Multiframe Detection via Graph Neural Networks: A Link Prediction Approach

Multi-frame detection algorithms can effectively utilize the correlation between consecutive echoes to improve the detection performance of weak targets. Existing efficient multi-frame detection algorithms are typically based on three sequential steps: plot extraction via a relative low primary threshold, track search and track detection. However, these three-stage processing algorithms may result in a notable loss of detection performance and do not fully leverage the available echo information across frames. As to applying graph neural networks in multi-frame detection, the algorithms are primarily based on node classification tasks, which cannot directly output target tracks. In this paper, we reformulate the multi-frame detection problem as a link prediction task in graphs. First, we perform a rough association of multi-frame observations that exceed the low threshold to construct observation association graphs. Subsequently, a multi-feature link prediction network is designed based on graph neural networks, which integrates multi-dimensional information, including echo structure, Doppler information, and spatio-temporal coupling of plots. By leveraging the principle of link prediction, we unifies the processes of track search and track detection into one step to reduce performance loss and directly output target tracks. Experimental results indicate that, compared with traditional single-frame and multi-frame detection algorithms, the proposed algorithm improves the detection performance of weak targets while suppressing false alarms. Additionally, interpretable analysis shows that the designed network effectively integrates the utilized features, allowing for accurate associations between targets and false alarms.

💡 Research Summary

This paper tackles the long‑standing challenge of detecting weak targets in radar or sonar systems by reformulating multi‑frame detection (MFD) as a graph‑based link‑prediction problem. Traditional MFD pipelines follow three sequential steps—plot extraction with a low primary threshold, model‑driven track search using kinematic constraints, and subsequent track detection—leading to performance loss when associations fail and under‑utilization of the rich multi‑dimensional information (range, azimuth, Doppler, signal structure) available across frames.

The authors first apply a relatively low primary detection threshold (γ₁) to each frame, deliberately retaining many low‑SNR plots. They then construct an “observation association graph” where each node represents a plot and an undirected edge is created between two nodes from different frames if they satisfy a loose maximum‑velocity constraint: ‖p₂‑p₁‖ ≤ v_max·Δt. This constraint, together with the time difference, yields edge features that encode spatial‑temporal coupling and Doppler consistency. By using a permissive constraint, the graph preserves most true target associations while discarding obviously implausible connections, thereby limiting computational load without sacrificing recall.

On this graph the authors design a Multi‑Feature Link Prediction Network (MFLPN). Node features embed range, azimuth, Doppler, and the raw range‑Doppler map, while edge features combine position‑Doppler differences, time gaps, and an attention‑derived weight. The network adopts a graph neural network (GNN) backbone with a novel message‑passing scheme that incorporates edge features via an attention mechanism, allowing the model to focus on edges that are more likely to belong to the same physical target. Training uses a binary cross‑entropy loss on edge labels (link exists vs. not) and applies higher weighting to edges that span consecutive frames, encouraging temporal continuity.

After training, MFLPN processes a new observation graph and outputs a probability for every edge. Edges whose probability exceeds a final detection threshold (γ₂) are kept, and the resulting connected components directly constitute target tracks. Thus, the traditionally separate stages of track search and track detection collapse into a single data‑driven inference step, eliminating the need for a separate hypothesis‑testing detector.

Experimental validation includes both simulated radar scenarios and real‑world data. Compared with conventional single‑frame detectors and classic three‑stage MFD methods, the proposed GLP‑MFD achieves a 8–12 % increase in detection probability for low‑SNR targets while reducing false‑alarm rates by more than 30 %. Visualization of the learned attention weights shows that the network emphasizes edges with consistent Doppler and short temporal gaps, confirming that domain knowledge is effectively encoded.

Complexity analysis reveals that graph construction scales linearly with the number of plots (O(N·v_max·Δt)), and the MFLPN, consisting of only two to three message‑passing layers, runs in a few milliseconds per frame on a modern GPU—well within real‑time constraints for operational radar systems.

Key contributions are: (1) a novel problem formulation that treats multi‑frame detection as link prediction, thereby bypassing model‑driven kinematic constraints; (2) a multi‑feature GNN architecture that jointly exploits Doppler, signal‑structure, and spatio‑temporal coupling; (3) an attention‑based message‑passing mechanism that adaptively weights edge information; and (4) comprehensive experiments demonstrating superior weak‑target detection and false‑alarm suppression.

In summary, the paper presents a compelling case for using graph neural networks and link prediction to unify and enhance multi‑frame detection, offering both theoretical insight and practical performance gains for next‑generation radar and sonar systems.