Adaptive Multi-Dimensional Coordinated Comprehensive Routing Scheme for IoV

1 Adapti v e Multi-Dimensional Coordinated Comprehensi v e Routing Scheme for IoV Ruixing Ren İD , Minqi T ao İD , Junhui Zhao İD , Senior Member , IEEE , Qiuping Li İD , and Xiaoke Sun İD Abstract —The characteristics of high-speed node mov ement and dynamic topology changes pose great challenges to the design of internet of vehicles (IoV) r outing protocols. Existing schemes suffer from common problems such as insufficient adaptability and lack of global consideration, making it difficult to achieve a globally optimal balance between routing reliability , real-time performance and transmission efficiency . This paper proposes an adaptive multi-dimensional coordinated comprehensiv e routing scheme for IoV envir onments. A complete IoV system model including network topology , communication links, hierarchical congestion and transmission delay is first constructed, the routing problem is abstracted into a single-objecti ve optimization model with multiple constraints, and a single-hop link comprehensiv e routing metric integrating link reliability , node local load, net- work global congestion and link stability is defined. Second, an intelligent transmission switching mechanism is designed: candidate nodes are screened through dual criteria of connectivity and progressiv eness, a dual decision-making of primary and backup paths and a threshold switching strategy are introduced to av oid link interruption and congestion, and an adaptive update function is constructed to dynamically adjust weight coefficients and switching thresholds to adapt to changes in network status. Simulation r esults show that the proposed scheme can effectively adapt to the high dynamic topology and network congestion characteristics of IoV , perform excellently in key indicators such as routing interruption times, packet delivery rate and end-to- end delay , and its comprehensive performance is significantly superior to traditional routing schemes. Index T erms —Internet of vehicles, routing optimization, net- work congestion, reliable communication, dynamic topology , link stability I . I N T R OD U C T I O N W ith the rapid dev elopment of intelligent transportation systems, the internet of vehicles (IoV), as a core enabling tech- nology , has become a research hotspot in both academia and industry [1]. By enabling lo w-latency and high-reliability com- munications between v ehicle-to-vehicle (V2V) and vehicle-to- infrastructure (V2I), IoV provides a critical data interaction platform for applications such as autonomous driving, intelli- gent route planning, and traf fic safety early warning [2], [3]. Howe ver , IoV networks exhibit distinct characteristics, including high-speed node mobility , dynamically changing Corresponding author: Junhui Zhao. Ruixing Ren, Junhui Zhao are with the School of Electronic and Information Engineering, Beijing Jiaotong University , Beijing 100044, China. (e-mail: renruixing0604@163.com; junhuizhao@hotmail.com) Minqi T ao is with the School of Information and Software Engineering, East China Jiaotong University , Nanchang 330013, Chiina. Qiuping Li and Xiaoke Sun are with the National Computer Net- work Emergency Response T echnical T eam/Coordination Center of China (CNCER T/CC), Beijing 100029, China. topology , drastic channel quality variation, and frequent net- work congestion. These pose sev ere challenges to design- ing efficient and robust routing protocols [4], [5]. How to guarantee high packet deliv ery ratio (PDR), low end-to-end delay , and low bit error rate (BER) in such complex and dynamic environments remains a core ur gent issue in current IoV routing research. Extensiv e research has been conducted on routing optimiza- tion in IoV en vironments. Early studies mainly focused on topology-based routing protocols, but they incur enormous routing overhead and are prone to link breakage in highly dynamic IoV scenarios [6]. Recently , W ang et al. [7] inte grated a dynamic topology evolution model with routing optimization to provide a basis for topology prediction in decision-making. Howe ver , the coordination mechanism between local link changes and global topology optimization was not clearly reflected, leaving room for improvement in multi-objective performance balance. T o improve protocol scalability , geographic location-based routing protocols have emerged [8]. Howe ver , it is prone to local optimality in the case of network holes or uneven node distrib ution. Y e et al. [9] used satellite maps and SD maps to predict lane maps, alleviating the local optimality problem caused by ov er-reliance on local perception. Suo et al. [10] utilized node location information to construct proacti ve multipath routes, and adopted adaptive path adjustment as well as a static-node-based local routing maintenance scheme. Nev ertheless, end-to-end connected paths may be absent for a long time in scenarios such as sparse vehicle density or suburban areas. T o address this, opportunistic routing or delay- tolerant network routing ideas have been introduced [11], abandoning the pre-setting of complete paths and adopting a store-carry-forward mode [12]. This mode uses vehicle mobility as a transmission medium, temporarily storing data when no suitable next hop is av ailable until a forwarding opportunity arises. W ith the enrichment and differentiation of IoV application scenarios, a single routing strategy can hardly meet all re- quirements, making hybrid and clustering routing protocols important research directions. Hybrid routing addresses com- plex and dynamic network en vironments by integrating the advantages of different routing strategies [13]. For example, geography-based unicast routing is adopted in high-vehicle- density areas, switching to opportunistic routing in sparse areas [14]; alternativ ely , cluster-structure routing is used at the backbone network level, while simple flooding or geo- graphic forwarding is employed within clusters. Clustering routing forms a virtual backbone network through dynamic 2 vehicle clustering and cluster head election, reducing nodes in volved in routing calculation, lowering control o verhead, and improving routing scalability and stability [15]. In recent years, the de velopment of artificial intelligence has profoundly reshaped the research paradigm of IoV [16], [17]. In particular , intelligent routing algorithms based on deep reinforcement learning (DRL) hav e emerged as a cutting- edge research topic. Moon et al. [18] took the dynamic traffic en vironment as the state input and achiev ed dynamic routing decision-making through agent training. By enabling continu- ous interaction between routing entities and the en vironment, such methods learn optimal forwarding policies under various network states from historical or real-time data, without ex- plicitly predefined complex routing rules. They thus exhibit great potential in handling high-dimensional state spaces and complex optimization objectives, and can adaptively learn and approximate optimal routing strategies in dynamically uncertain environments [19]. F or instance, a DRL-based effi- cient routing algorithm for intelligent QoS optimization selects optimal data transmission paths by dynamically adapting to vehicular network variations [20]. On this basis, more studies hav e proposed targeted DRL routing schemes to address the unique characteristics of IoV , such as highly dynamic topology , unpredictable contact patterns, and resource constraints. As an e xample, the CR-DRL algorithm, which integrates an actor-critic framework and heuristic functions, realizes real- time optimal relay selection and dynamic overlapping cluster adjustment [11]. Although e xisting IoV routing protocols hav e achiev ed performance optimization in different scenarios, they still hav e common shortcomings. T opology-based protocols fail to adapt to highly dynamic topologies. Geography-based protocols tend to fall into local optima. Opportunistic routing is only suitable for sparse scenarios and cannot balance transmission ef ficiency in dense ones. Cluster management and strategy switching mechanisms of hybrid and clustering routing remain to be improv ed. While DRL-based routing algorithms exhibit cer- tain adaptability , they mostly focus on local network states, neglect global congestion and link stability , and rely on fixed parameters that are difficult to adapt to dynamic IoV scenarios, thus failing to globally optimize routing reliability , latency and transmission efficienc y . T o address the abov e issues, this paper proposes an adapti ve multi-dimensional coordinated comprehensive routing scheme for IoV . The main contributions are summarized as follows: • W e model the IoV system and establish a multi- dimensional comprehensive ev aluation model that incor- porates link reliability , local node load, global network congestion, and link stability , providing a comprehensiv e and objectiv e quantitativ e basis for routing decisions. • An intelligent V2I/V2V handov er mechanism is de- signed. Candidate nodes are selected via connectivity and progressiveness, while primary-backup dual-path and threshold-based switching strategies are adopted to av oid link disruptions and congestion. An adaptiv e update func- tion is dev eloped for diverse network densities and loads, enabling real-time adaptation to network dynamics. V2 V l in k V2 I l in k Fig. 1: IoV System Network T opology Model. • Simulation results demonstrate that the proposed scheme outperforms conv entional approaches in key metrics in- cluding routing disruption frequency , packet delivery ra- tio, and end-to-end delay , verifying its effecti veness and superiority in highly dynamic IoV environments. The rest of this paper is organized as follows. Section II constructs the IoV system model and formulates the problem. Section III presents the proposed scheme. Section IV validates the proposed scheme through simulation experiments. Section V concludes the paper . I I . S Y S T E M M O D E L A N D P R O B LE M F O R M U L A T I O N A. Network T opology Model The IoV system constructed in this paper consists of two communication entities: vehicle nodes and road side units (RSUs). As shown in Fig. 1, the road region is defined as a 2D rectangular space R = [0 , L road ] × [0 , W road ] . The set of vehicle nodes is denoted by V = { v 1 , v 2 , ..., v n , ..., v N } , where N is the total number of vehicles. Each vehicle v n has a communication radius R v , and its state is jointly described by the position coordinate P n ( t ) = [ x n ( t ) , y n ( t )] T , velocity vector v n ( t ) = [ v nx ( t ) , v ny ( t )] T , and queue load q n ( t ) . The vehicle mobility model satisfies the following:  x n ( t + ∆ t ) = x n ( t ) + v nx ( t ) · ∆ t y n ( t + ∆ t ) = y n ( t ) + v ny ( t ) · ∆ t (1) where ∆ t is the time slot interval. The set of RSU nodes is denoted as R = { r 1 , r 2 , ..., r m , ..., r M } , where M is the total number of deployed RSUs. Each RSU has a much larger communication cov erage than vehicle nodes, i.e., R r > R v . Each r m can establish direct communication links with all ve- hicle nodes within its communication coverage, which assists in forwarding data packets and alleviates the link instability and network congestion of V2V communications. B. Communication Link Model The communication link set of the IoV can be expressed as the union of two core modes: L = L V2V ∪ L V2I . The V2V direct link set is defined as L V2V = { l ( v i , v j ) : v i , v j ∈ V , d ( v i , v j ) ≤ R v } . (2) 3 The condition is that the Euclidean distance between two vehi- cle nodes v i and v j does not exceed the vehicle communication radius R v . The V2I communication link set is defined as L V2I = { l ( v n , r m ) : v n ∈ V , r m ∈ R , d ( v n , r m ) ≤ R r } . (3) This holds if the Euclidean distance between vehicle node v n and RSU node r m does not exceed the communication radius of the RSU. In urban IoV scenarios, signal propagation is affected by path loss, shadow fading, multipath fading, adjacent-channel interference, and other factors. The channel model established in this paper comprehensiv ely considers the abov e factors. Based on the recei ved power P r , single-hop link quality is characterized from three dimensions: signal strength, trans- mission error , and data reception success rate, using three core metrics: signal-to-noise ratio (SNR), BER, and packet reception rate (PRR). For any single-hop communication link l ( n i , n j ) ∈ L in the IoV , where n i , n j ∈ V ∪ R , the recei ved power P r at the receiv er [21] adopts the unified form as follo ws P r ( d ( n i , n j )) = P t · G t · G r · ( λ 4 π d 0 ) 2 · ( d 0 d ( n i , n j ) ) ζ · X σ · | h | 2 (4) where P t is the transmit power of the transmitting node, G t and G r are the transmit and recei ve antenna gains, respecti vely , λ is the carrier wav elength, d 0 is the reference distance, ζ is the path loss exponent, and X σ is the shadow fading factor following a log-normal distribution 10 log 10 X σ ∼ N (0 , σ 2 ) , which characterizes the slo w fading caused by fixed obstacles. | h | 2 is the amplitude of multipath fading, characterizing the fast fading after signal reflection and scattering. The SNR is defined as the ratio of the useful signal po wer to the noise power at the receiver: SNR( n i , n j ) = P r ( d ( n i , n j )) N 0 (5) A higher SNR indicates stronger anti-interference capability of the link. The BER is approximated by the complementary error function: BER( n i , n j ) ≈ 1 2 erfc( r SNR( n i , n j ) 2 ) (6) It directly reflects the accurac y of data transmission. PRR is a core metric characterizing the link data reception success rate, defined as the ratio of the number of data packets successfully receiv ed by the receiv er to the total number transmitted by the sender . It comprehensi vely reflects the ov erall link transmission quality and serves as a key basis for link selection in routing algorithms. Under ideal assumptions (independent bit errors, no channel coding or interleaving), PRR and BER satisfy the exponential relationship: PRR( n i , n j ) = (1 − BER( n i , n j )) L p (7) where L P is the packet length. Howe ver , in urban vehicular communication scenarios, the application of channel coding and interleaving alters the mapping between BER and PRR, such that they no longer follo w the simple exponential re- lationship under ideal assumptions. Meanwhile, as BER → 0, PRR → 1; as BER → 1, PRR → 0. Such extreme values cause singularity in routing metrics, impairing the numerical sta- bility and decision rationality of the algorithm. Therefore, to simplify computation, avoid e xtreme-v alue interference, and ensure accuracy suitable for short-range urban vehicular communication, this paper adopts the following clipped ap- proximation: PRR( n i , n j ) = clip(1 . 0 − BER( n i , n j ) , 0 . 01 , 0 . 999) (8) where clip( · ) denotes the clipping function, which strictly confines PRR to the interval [0 . 01 , 0 . 999] .This av oids extreme- value interference while maintaining sufficient accuracy in short-range vehicular communication, providing stable quan- titativ e support for subsequent routing decisions. C. Congestion Model T o accurately capture congestion dynamics, this paper es- tablishes a hierarchical congestion model from node-level and global-lev el dimensions, providing a quantitati ve basis for routing decisions. Node-lev el congestion focuses on the local load state of a single v ehicle node. This paper adopts normalized node load as the core metric. The node load is dynamically updated with packet forwarding during operation and decays naturally over time. T o eliminate the de viation caused by dif ferences in b uffer capacity among nodes, the normalized queue load is defined as q ( v n ) = min(1 , L n ( t ) L max ) ∈ [0 , 1] (9) where L n ( t ) is the current actual queue length, and L max is the maximum buffer capacity of the node. The value of q ( v n ) is positiv ely correlated with the node congestion level, which can be broadcast to one-hop neighbors via beacons to provide a basis for routing selection. Node-lev el congestion metric only reflects the local load of a single node, and cannot characterize global congestion caused by excessi ve vehicle density in the region and over - occupied channel resources. This paper selects channel oc- cupancy as the global congestion metric. Assuming vehicles are uniformly distributed, the v ehicle density in the region is N / ( W road · L road ) . The communication coverage area of each vehicle is π R 2 v , but the cov erage areas of different vehicles ov erlap. The channel occupancy level can be approximated as the ratio of the total coverage area of all vehicles to the total area of the region [22]. The formula for the global congestion lev el C global is giv en as follows: C global = min(1 , N · π R 2 v W road · L road ) (10) Its value is positively correlated with the degree of global congestion: a higher overlap indicates greater vehicle density in the re gion, more intense channel contention, and more sev ere global congestion. 4 D. Delay Model The end-to-end transmission delay in IoV consists of three core components, corresponding to the three physical processes: signal propagation, data transmission, and queue waiting. In the short-range communication scenario of IoV , the propagation delay is usually small and can be neglected. T ransmission delay is the time required to send a fixed-size data packet from node n i to node n j , which is related to packet length and link transmission rate: τ t ( n i , n j ) = L p rate( n i , n j ) , n i , n j ∈ V ∪ R (11) where rate( n i , n j ) is the physical-layer transmission rate of link ( n i , n j ) , determined by link distance and channel conditions. V ehicle nodes use normalized load to approximate queuing delay , adapting to distributed, low-ov erhead routing decisions: τ q V ( v n ) = τ 0 · q ( v n ) , v n ∈ V (12) where τ 0 is the maximum queuing delay reference, used to map the load to a reasonable delay magnitude. As a centralized infrastructure, the RSU uses an in-degree-processing time model to characterize queuing delay , reflecting its scheduling capability and processing pressure: τ q R ( r m ) = k R · deg in ( r m , t ) · τ R , r m ∈ R (13) where k R is the RSU scheduling coefficient, used to adjust the impact of scheduling efficiency on delay [23]. deg in ( r m , t ) is the number of acti ve incoming links to RSU r m at time t . τ R is the average processing time of a single data packet for the RSU. The av ailable path P from source node n S to destination node n D is defined as an ordered sequence consisting of k + 1 nodes: P = { n S = n 0 , n 1 , n 2 , ..., n D = n k } , ∀ n i ∈ V ∪ R (14) where any adjacent node pair ( n i , n i +1 ) must satisfy d ( n i , n i +1 ) ≤ R v or R r , i.e., a valid direct communication link exists. The path must also satisfy the acyclic constraint, i.e., contain no repeated nodes, to avoid routing loops and the resulting resource waste and delay surge. The total end-to-end delay τ total of the transmission path P from the source node to the destination node can be e xpressed as: τ total = X n i ,n j ∈P τ t ( n i , n j ) + X v n ∈P ,v n  = D τ q V ( v n ) + X r m ∈R ,r m  = D τ q R ( r m ) (15) E. Routing Pr oblem F ormulation The total end-to-end routing cost of path P is defined as the sum of the comprehensiv e routing metrics of each single-hop link: Cost( P ) = k − 1 X i =1 M ( n i , n i +1 ) (16) where M ( n i , n i +1 ) is the comprehensiv e routing metric for a single-hop link, which jointly reflects link quality , node load, network congestion, and path stability . The core objective of routing optimization is to select the optimal path P ∗ with the minimum total cost from all a vailable paths Ψ , while ensuring routing v alidity and quality of service: P 1 : P ∗ = arg min P ∈ Ψ Cost( P ) s.t. C1 : d ( n i , n j ) ≤ R v , C2 : d ( n i +1 , n D ) < d ( n i , n D ) , n i +1  = n D , C3 : PDR( P ) = Y PRR( n i , n j ) ≥ PDR min , C4 : q n ≤ q max , v n  = n D , C5 : τ total ≤ T max , C1 ensures physical connectivity of the path, i.e., each hop must be within communication range. C2 ensures each hop is closer to the destination node n D than the previous hop. C3 guarantees path reliability and avoids severe packet loss caused by poor link quality , where PDR is the end-to-end packet deliv ery ratio. C4 prev ents overloaded nodes, since packet dropping may occur if the queue of an intermediate node is full. C5 is the end-to-end total delay constraint. I I I . P R O PO S E D S C H E M E The previous section has abstracted the IoV routing problem as a single-objectiv e optimization model with multiple con- straints, clarifying the core goal of finding the path with mini- mum total routing cost while ensuring constraints such as path connectivity , reliability , congestion control, and end-to-end de- lay . According to this objective, the single-hop comprehensiv e routing metric M ( n i , n i +1 ) should adopt multi-dimensional criteria, including link reliability , node load, global congestion, and path stability . Link stability quantifies the resilience of a link to topology changes by combining relati ve speed and communication distance: s ( n i , n i +1 ) = 1 − 0 . 5 × min(1 , v rel ( n i , n i +1 ) v ref ) − 0 . 5 × min(1 , d ( n i , n i +1 ) R v ) (17) where v rel ( n i , n i +1 ) is the relative speed between nodes n i and n i +1 , and v ref is the reference speed threshold. A larger value of s ( n i , n i +1 ) ∈ (0 , 1] indicates a more stable link. Then M ( n i , n i +1 ) can be obtained as: M ( n i , n i +1 ) = α · 1 PRR( n i , n i +1 ) + β · q n i +1 + γ · C global + δ · (1 − s ( n i , n i +1 )) (18) where α, β , γ , δ are weighting coefficients satisfying α + β + γ + δ = 1 . Based on the above multi-dimensional metric, to further address the core challenges of IoV—high dynamics, frequent congestion, and time-varying topology—this paper proposes three ke y technical strate gies. These strategies form a complete optimal routing selection framework from the perspectives of robustness enhancement, resource optimization, and scenario adaptation. The routing decision procedure of the proposed scheme is shown in Algorithm 1. 5 Algorithm 1 The designed routing selection algorithm Require: n S n D , V , R , C th , etc. 1: Calculate the global congestion degree C global by Eq. (10). 2: It is defaulted to the V2V mode. 3: for Each r m ∈ R do 4: if The source node can communicate with this RSU. then 5: Switch to V2I mode; break 6: end if 7: end f or 8: if The mode is V2I. then 9: for Each reachable RSU and each vehicle node do 10: if The vehicle satisfies both V2I connectivity and the destination-oriented advancement constraint then 11: Add the vehicle to the candidate set Ψ 12: end if 13: end for 14: end if 15: if Ψ = ∅ then 16: for Each vehicle node do 17: if The vehicle satisfies both V2V connecti vity and the destination-oriented advancement constraint. then 18: Add the vehicle to the candidate set Ψ . 19: end if 20: end for 21: end if 22: for Each vehicle node ∈ Ψ do 23: Calculate the PRR (Eq. (8)), stability (Eq. (17)), load (Eq. (9)), and M ( n i , n i +1 ) (Eq. (18)). 24: if The current value of M ( n i , n i +1 ) is smaller then 25: Update the primary path to this node. 26: end if 27: if The current stability is higher . then 28: Update the backup path to this node. 29: end if 30: end f or 31: next hop = primary path 32: if The primary path is v alid and its metric exceeds the threshold, and the backup path is v alid and different. then 33: next hop = backup path 34: end if 35: return next hop (1) Global congestion state ev aluation. Node n i first obtains the current global congestion lev el C global . If it is within the cov erage of an RSU, the v alue can be obtained from the periodic broadcast of the global congestion statistics by the RSU. Otherwise, the node can estimate it by counting the local av erage number of neighbors. (2) Adaptive V2I/V2V transmission mode selection. By continuously monitoring beacon messages broadcast by the RSU, the node determines whether it is within RSU cov erage. If an RSU with signal strength abov e a predefined threshold is detected, the V2I mode is preferred. If no av ailable RSU is detected, the node switches to V2I mode. (3) V alid candidate node screening. According to the current transmission mode, neighbor nodes are subjected to dual screening (satisfying constraints C1: connectivity , C2: ad- vancement): • V2I mode: Select vehicle nodes that communicate with both the reachable RSU and the source node n i , and are closer to the destination node n D . • V2V mode: Select vehicle nodes within the communica- tion range of the source node and closer to the destination node n D . • If the candidate set is empty after screening, enter the recovery mode, temporarily relax the geographic advancement constraint, or adopt the carry-and-forward strategy . (4) Primary and backup path decision-making. For each valid candidate node n j , the algorithm calculates the link stability s ( n i , n j ) and comprehensiv e metric M ( n i , n j ) by Eq. (17) and Eq. (18), and determines the next hops of the primary and backup paths based on these values. The next hop of the primary path is selected as the node with the minimum M ( n i , n j ) , representing the optimal comprehensive routing quality . The next hop of the backup path is chosen as the node with the maximum s ( n i , n j ) , representing the most stable link. When the comprehensi ve metric M of the primary path exceeds the predefined threshold C th , the algorithm automatically switches to the backup path to av oid potential link breakage or se vere congestion, thus improving routing robustness. Finally , the selected next-hop forwarding node is output. Notably , the performance of the algorithm highly depends on the selection of the weighting coef ficients α, β , γ , δ and the switching threshold C th . Different network scenarios exhibit different sensiti vities to each metric, and static parameters cannot achie ve optimality across all scenarios. Therefore, we design the following adapti ve adjustment strate gy . When the network density is high (i.e., C global is large), global congestion becomes the dominant issue. It is neces- sary to increase γ to enhance congestion av oidance, while appropriately reducing α and δ . When the network density is low , the risk of link disruption increases, and δ should be enlarged to emphasize stability . The dynamic weighting function is designed as follo ws: γ = γ 0 · (1 + C global ) , (19) δ = δ 0 · (1 − C global ) (20) where γ 0 and δ 0 are the initial baseline v alues. When the queue length of node v n is high, it indicates local congestion. Thus, β should be increased to steer traffic away from other con- gested nodes and av oid aggrav ating congestion. The formula is adopted as follows: β = β 0 · (1 + q ( v n )) (21) These weights are then normalized together with the other weights, so that heavily loaded nodes prefer to select next-hop nodes with shorter queues. When frequent link disruptions are detected, C th should be appropriately decreased to improv e switching sensitivity and reduce disruptions. Conv ersely , if ex- cessiv e switching causes routing oscillations, C th is increased to stabilize the routing. 6 T ABLE I: Hyperparameter settings Parameter V alue Road length L road 20,000 m Road width W road 200 m Channel bandwidth 10 MHz T ransmit po wer P t 20 dBm Noise po wer spectral density 10 − 13 W/Hz RSU communication range R r 500 m V2V maximum communication distance R v 300 m Reference speed threshold v ref 30 m/s Minimum PDR threshold PDR min 0.7 Maximum end-to-end delay T max 10 s Packet length L p 10000 bit σ , ζ , d 0 8, 3.5, 1 I V . S I M U L AT I ON R E S U LT S A N D A N ALY S I S In this section, simulation experiments are conducted to verify the ef fecti veness of the proposed scheme. The initial x and y coordinates of vehicles follow a uniform distribution. The vehicle speed follows a uniform distribution over 15 ∼ 30 m/s. The vehicle heading angle follows a uniform distribution ov er -0.05 ∼ 0.05 rad. The number of RSUs M is giv en by max(5 , N // 200) , and RSUs are uniformly and equidistantly deployed along the road centerline. The maximum hop count is set to 15. The mainstream 2.4 GHz wireless communication band for IoV is adopted, so the carrier wav elength λ is 0.125 m. V ehicles generate tasks, so source nodes and destination nodes are randomly selected from all vehicle nodes to av oid result bias caused by fixed paths. Through multiple simulations and parameter tuning, C th is set to 1.5, and α, β , γ , δ are set to 0.4, 0.2, 0.2, and 0.2, respecti vely . Other hyperparameter settings are shown in T able I. T o verify the performance superiority of the proposed scheme, four typical comparati ve algorithms in the field of IoV routing are selected, and the core design of each algorithm is as follows: • RSU-Prioritized V2V (RSU-V2V): Firstly , vehicles that are within the communication range of the same RSU as the source node and closer to the destination node are selected as candidates. When no av ailable RSU exists, the algorithm degrades to traditional V2V direct commu- nication. The next hop is determined by a static scoring function that combines transmission rate, link BER, and distance to the destination node. • Load-A ware V2V (LA-V2V): Adopts full V2V direct communication. On the basis of the scoring dimensions of RSU-V2V , it adds the instantaneous load of nodes and avoids local congestion by penalizing highly loaded nodes. Ho we ver , it only perceives node-lev el local load, without global congestion and link stability metrics. • Model-based Reinforcement Learning (MRL): Constructs a Q-table with source-neighbor node pairs as state ke ys, calculates the immediate rew ard based on transmission 200 300 400 500 600 700 800 900 1000 Vehicle density 0 10 20 30 40 50 Number of routing interruptions Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 2: The trend of routing interruption count v ersus vehicle density . 200 300 400 500 600 700 800 900 1000 Vehicle density 0 20 40 60 80 100 PDR(%) Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 3: The trend of PDR v ersus vehicle density . rate, BER, and distance, and updates the Q-values ac- cordingly . It adopts the ε -greedy strategy , fusing the immediate reward and historical Q-v alues to select the next hop, thus achieving adaptiv e routing optimization. • DRL for QoS Optimization (DRL-QoS): Constructs a four-dimensional state vector consisting of link interrup- tion probability , node load, transmission rate, and BER, and designs an exponential reward function to minimize the deviation between QoS metrics and target v alues. It selects the next hop by combining rew ard values and distance through the ε -greedy strate gy . Fig. 2 illustrates the variation of routing disruption times with vehicle density . The proposed scheme exhibits a continu- ous decrease in disruption times from lo w to high v ehicle den- sity , outperforming the comparative algorithms significantly . This superiority verifies that the global congestion aware- ness, link stability metric, and primary-backup path switching mechanism of the proposed scheme can effecti vely adapt to the highly dynamic topology of IoV . Even under ultra-high- density congestion scenarios, the routing disruption rate can be maintained at an extremely low level. In contrast, LA-V2V only perceives local load, RSU-V2V lacks intelligent learning capabilities, and both DRL-QoS and MRL fail to incorporate global dimensions, rendering them unable to achiev e compa- rable disruption control performance. As shown in Fig. 3, the PDR of the proposed scheme is significantly improved with the increase of vehicle density , 7 200 300 400 500 600 700 800 900 1000 Vehicle density 0 1 2 3 4 5 6 7 8 BER(%) Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 4: The trend of BER v ersus vehicle density . 200 300 400 500 600 700 800 900 1000 Vehicle density 60 80 100 120 140 160 180 200 220 240 Throughput (Mbps) Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 5: The trend of throughput v ersus vehicle density . achieving fa vorable performance under various density scenar - ios and consistently outperforming the comparative algorithms. DRL-QoS and RSU-V2V also exhibit strong QoS guarantee capability in terms of PDR under high-density conditions. Limited by its single-node load perception only , LA-V2V achiev es slightly inferior PDR performance at high density . Meanwhile, MRL suffers from the lack of congestion and stability awareness, resulting in a PDR close to 0, which cannot meet the communication requirements of IoV . This indicates that the proposed scheme effecti vely enhances the packet delivery ratio via multi-metric ev aluation and primary- backup path switching, and demonstrates outstanding rob ust- ness especially under heavy congestion scenarios. Link BER and throughput are key metrics for ev aluating link quality and transmission efficienc y of routing algorithms, which jointly determine the reliability and ef fecti veness of IoV communications. As shown in Figs. 4 and 5, the proposed scheme exhibits significant adv antages in both indicators. Its BER remains stably at an extremely low level, ef fecti vely av oiding high-error links; meanwhile, the throughput increases steadily with vehicle density , reaching approximately 226 Mbps in high-density scenarios, which is notably superior to the comparison algorithms. The BER and throughput of DRL- QoS, RSU-V2V , and LA-V2V are slightly inferior to those of the proposed scheme. In contrast, due to the lack of awareness of link quality , congestion, and stability , the MRL algorithm experiences a sharp rise in BER to around 8% and a drastic drop in throughput to approximately 70 Mbps when vehicle 200 300 400 500 600 700 800 900 1000 Vehicle density 0.00 0.02 0.04 0.06 0.08 0.10 Delay (s) Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 6: The trend of delay v ersus vehicle density . 200 300 400 500 600 700 800 900 1000 Vehicle density 0 2 4 6 8 10 12 14 16 Average path length Proposed DRL-QoS RSU-V2V LA-V2V MRL Fig. 7: The trend of average path length versus vehicle density . density exceeds 300, failing to meet the core IoV requirements of low BER and high throughput. As shown in Fig. 6, the end-to-end delay of the proposed scheme remains consistently the lowest and increases most gently with rising vehicle density , significantly outperforming the comparison algorithms. The delays of DRL-QoS, RSU- V2V , and LA-V2V rise slo wly in the ranges of 0.020 ∼ 0.036 s, 0.025 ∼ 0.041 s, and 0.030 ∼ 0.048 s, respecti vely , reflecting the impact of different sensing capabilities on latency control. In contrast, due to the lack of congestion and link stability awareness, the MRL algorithm exhibits generally high latency across all vehicle densities, failing entirely to meet the real- time requirements for IoV safety information transmission. As sho wn in Fig. 7, the a verage path length of the pro- posed scheme increases steadily with vehicle density , reaching approximately 14.4 hops in high-density scenarios, which is significantly longer than that of the comparison algorithms. This is because the proposed scheme prioritizes paths with good link quality and low congestion rather than the short- est path in routing decisions, thereby ef fecti vely improving communication reliability . The average path lengths of RSU- V2V , LA-V2V , and DRL-QoS fluctuate between 11.7 and 13.9 hops, reflecting their trade-off between QoS guarantee and path length. In contrast, due to the lack of congestion and link quality awareness, the MRL algorithm always selects the shortest path, with an av erage path length stable at 5 hops, resulting in extremely poor communication reliability in high- density scenarios. T o comprehensiv ely e v aluate the ov erall performance of 8 DRL-QoS RSU-V2V LA-V2V MRL Proposed 0 20 40 60 80 100 Overall Score 89.0 87.2 85.4 29.9 92.3 Fig. 8: Comprehensi ve Performance Scores of Dif ferent Algo- rithms. various algorithms in IoV scenarios, we weighted and normal- ized the number of interruptions, link BER, end-to-end delay , throughput, and PDR to obtain the o verall scores sho wn in Fig. 8. The proposed scheme significantly outperforms others with a score of 92.3, while the scores of DRL-QoS, RSU-V2V , and LA-V2V are 89.0, 87.2, and 85.4, respectively , and the MRL algorithm achieves only 29.9. This result demonstrates that the proposed scheme achiev es the optimal overall balance of com- munication reliability , real-time performance, and transmission efficienc y in highly dynamic and congested IoV scenarios via its inno vati ve multi-dimensional metric, primary-backup path switching, and V2I/V2V adaptive mechanism. V . C O N C L U S I O N T o address issues such as high dynamic topology , channel fluctuation, and network congestion in IoV , this paper proposed an adapti ve multi-dimensional coordinated comprehensiv e routing scheme for IoV . The scheme constructed a complete system model including network topology , communication links, hierarchical congestion, and transmission delay , defined a multi-dimensional routing metric fusing link reliability , node load, global congestion, and link stability , designed an intelligent V2I/V2V switching mechanism, combined primary- backup path dual decision-making and threshold switching strategy to a void link interruption, and realized dynamic parameter adjustment through an adapti ve function to adapt to network state changes. Simulation experiments comparing with four typical algorithms showed that the proposed scheme effecti vely reduced the number of routing interruptions and bit error rate, improv ed packet delivery rate and throughput, maintained lo w end-to-end delay with gentle growth, and achiev ed the global optimal balance of routing reliability , real- time performance, and transmission ef ficiency . R E F E R E N C E S [1] J. Zhao, R. Ren, D. Zou, Q. Zhang, and W . Xu, “IoV-oriented integrated sensing, computation, and communication: System design and resource allocation, ” IEEE T ransactions on V ehicular T echnology , vol. 73, no. 11, pp. 16283–16294, Nov . 2024. [2] J. Contreras-Castillo, S. Zeadally , and J. A. Guerrero-Ibañez, “Internet of vehicles: Architecture, protocols, and security , ” IEEE Internet of Things Journal , vol. 5, no. 5, pp. 3701–3709, Oct. 2018. [3] R. Ren, J. Zhao, D. Zou, Q. Zhang, D. W ang, and W . Xu, “Collaborativ e computation in integrated sensing, communication, and computation system for autonomous driving, ” IEEE Tr ansactions on Intelligent T ransportation Systems , vol. 27, no. 1, pp. 883–894, 2026. [4] F . Cunha, L. V illas, A. Boukerche, G. Maia, A. V iana, R. A. Mini, and A. A. Loureiro, “Data communication in V ANETs: Protocols, applications and challenges, ” Ad Hoc Networks , vol. 44, pp. 90–103, Jul. 2016. [5] R. Ren, J. Zhao, Q. Zhang, D. W ang, and J. Li, “DRL beamforming in RIS-aided IoV for integrated-sensing-communication-computation, ” IEEE Internet of Things J ournal , vol. 12, no. 14, pp. 28201–28213, May 2025. [6] J. Chen and et al., “Enhancing routing performance through trajectory planning with DRL in U A V-aided V ANETs, ” IEEE T ransactions on Machine Learning in Communications and Networking , vol. 3, pp. 517– 533, Apr . 2025. [7] H. W ang, “Dynamic topology evolution and multi-objective routing op- timization for efficient V ANET communication, ” IEEE Access , vol. 13, pp. 36124–36134, Feb. 2025. [8] Y . He, Y . W ang, Y . Guo, and Y . Y ang, “Spectrum sensing based geo- graphic opportunity routing protocol in cogniti ve radio network, ” in 2025 28th International Confer ence on Computer Supported Cooperative W ork in Design (CSCWD) , pp. 2502–2507, 2025. [9] J. Y e, D. Paz, and et al., “SMAR T: Adv ancing scalable map priors for driving topology reasoning, ” in 2025 IEEE International Conference on Robotics and Automation (ICRA) , pp. 3298–3304, May 2025. [10] L. Suo, L. Liu, and et al., “ A reliable low-latency multipath routing algorithm for urban rail transit Ad Hoc networks, ” Sensors , vol. 23, no. 12, 2023. [11] M. S. Sani, S. Iranmanesh, R. Raad, and F . Tubbal, “Energy-efficient routing protocol in vehicular opportunistic networks: A dynamic cluster- based routing using deep reinforcement learning, ” IEEE T ransactions on Intelligent Tr ansportation Systems , pp. 1–16, 2026. Early Access. [12] E. Puka, P . Herrmann, and A. T aherkordi, “ An efficient and robust protocol to accelerate message deliv ery in cellular dead spots, ” IEEE T ransactions on V ehicular T echnology , pp. 1–16, 2025. Early Access. [13] P . P . Y adav and T . B. Reddy , “Smart city-oriented routing protocols in internet of vehicles: A comprehensiv e survey , ” in Proceedings of the International Conference on Advanced Materials, Manufacturing and Sustainable Development (ICAMMSD 2024) , pp. 115–128, Atlantis Press, Mar . 2025. [14] A. O’Driscoll and D. Pesch, “Hybrid geo-routing in urban vehicular networks, ” in 2013 IEEE V ehicular Networking Confer ence , pp. 63–70, Dec. 2013. [15] X. Bi, H. Huang, B. Zhang, and X. W ei, “ A hybrid routing algorithm for V2V communication in V ANETs based on blocked Q-learning, ” IEICE T ransactions on Communications , vol. E106.B, no. 1, pp. 1–17, 2023. [16] Y . Y ao, J. Zhao, Z. Li, X. Cheng, and L. W u, “Jamming and eav esdropping defense scheme based on deep reinforcement learning in autonomous v ehicle networks, ” IEEE T ransactions on Information F or ensics and Security , vol. 18, pp. 1211–1224, Jan. 2023. [17] R. Ren, J. Zhao, and Q. Zhang, “U A V-assisted collaborati ve sensing task offloading and resource allocation in IoV, ” IEEE T ransactions on V ehicular T echnology , pp. 1–10, 2025. [18] S. Moon, S. Koo, Y . Lim, and H. Joo, “Routing control optimization for autonomous vehicles in mixed traffic flow based on deep reinforcement learning, ” Applied Sciences , vol. 14, no. 5, no. 5, 2024. [19] J. Zhao, R. Ren, Y . Nie, and D. W ang, “Decision intelligence empower- ing resource management for 6G-V2X communications, ” IEEE Internet of Things Magazine , vol. 8, no .6, pp. 88–96, Nov . 2025. [20] S. Y e, L. Xu, Z. Xu, and F . W ang, “ A deep reinforcement learning-based intelligent QoS optimization algorithm for ef ficient routing in vehicular networks, ” Alexandria Engineering J ournal , vol. 107, pp. 317–331, Nov . 2024. [21] J. Zhao, R. Ren, Y . Wu, Q. Zhang, W . Xu, D. W ang, and L. Fan, “SEAttention-residual based channel estimation for mmW ave massive MIMO systems in IoV scenarios, ” Digital Communications and Net- works , vol. 11, no .3, pp. 778–786, Jun. 2025. [22] F . García V idal, E. Egea López, and J. Santa Lozano, “Evaluation of multichannel operation mechanisms for vehicular networks, ” V ehicular Communications , vol. 56, p. 100978, Dec. 2025. [23] B. Su, Y . Ju, and L. Dai, “Deadline-aware scheduling for transmitted RSU packets in cooperative vehicle-infrastructure systems, ” Applied Sciences , vol. 13, no. 7, Mar . 2023. 9