This paper proposes an image-processing-based method for personalization of calorie consumption assessment during exercising. An experiment is carried out where several actions are required in an exercise called broadcast gymnastics, especially popular in Japan and China. We use Kinect, which captures body actions by separating the body into joints and segments that contain them, to monitor body movements to test the velocity of each body joint and capture the subject's image for calculating the mass of each body joint that differs for each subject. By a kinetic energy formula, we obtain the kinetic energy of each body joint, and calories consumed during exercise are calculated in this process. We evaluate the performance of our method by benchmarking it to Fitbit, a smart watch well-known for health monitoring during exercise. The experimental results in this paper show that our method outperforms a state-of-the-art calorie assessment method, which we base on and improve, in terms of the error rate from Fitbit's ground-truth values.
It is suggested by several health experts that people should be concerned of their calorie intake and consumption (Hill et al. 2003). Nowadays, the assessment of calorie consumption remains challenging. There exists a gas analysis system for calorie consumption assessment (B Böhm, Hartmann, and H Böhm 2016), which seems highly accurate, but it needs large space and expensive devices. In addition, users of their system also lose freedom to move. Another method (Tsou and Wu 2015) was developed by Tsou and Wu where Kinect, a line of motion sensing input device that can detect the gesture of a whole body, is used for calorie assessment. This kind of device is expected to be extensively used in constructing rehabilitation applications in calorie assessment that are related to health promotion (Da Gama et al. 2015). In Tsou and Wu's method, the coordinates of body joints in 3D space are captured by Kinect and used to calculate the velocity of each joint movement, and then a kinetic energy for estimating calorie consumption. The method yields promising performance; however, there are still issues that can be improved, in particular, the issue that assessment does not take the body size of individual users into account.
In this paper, we propose an improved version of the method by Tsou and Wu. Note that in their method, kinetic energies are computed by using the velocities of body joints and the standard mass of each joint (a mass represents the portion of a joint of interest to the whole body, including muscles and bones attached to that joint). On the contrary, in our work, the mass of each body joint is derived by processing an image of the subject’s body. In other words, calorie consumption assessment by our method takes the body size of each user into account. Following an existing protocol for system evaluation (Ryu, Kawahawa and Asami 2008), we use a reference device, Fitbit, to evaluate the assessment accuracy of our system.
Nowadays, we could know how many calories a human consumes during walking by some smartphone applications. But accuracy is still in question. Most applications do not consider mass, which means they do not weight the importance of each body segment. Therefore, a method is required that adapts to the body size and weight of each individual user.
For the aforementioned exiting work on calorie consumption assessment based on gas analyzing, we stated that, based on their result, it is an accurate system. However, considering a high cost, largely needed space, it is impractical to adopt their approach to applications for promoting users’ physical health through daily exercise or motion gaming.
Our work is mainly based on the aforementioned existing method by Tsou and Wu, in which Kinect is used to monitor users’ activities and assess their calorie consumption. They showed error rates to a ground truth that is calorie consumption assessed by a reliable assessment tool, i.e., a heart rate monitor. In addition, the longer the training time, the less the error rate. They used kinetic energies of the body joints to build a regression function for estimating calorie consumption. The kinetic energy of each body joint is calculated as a multiplication of the joint’s standard scale with the body weight. We conjecture that assessment can be improved if the body scale is measured specifically for each individual user.
According to Tsou and Wu’s method, kinetic energy parameters are used to assess calorie consumption. This shows that such energies are related to the calorie consumption amount. Following their recipe, we also use kinetic energy parameters to assess calorie consumption.
The kinetic energy needs mass and velocity to calculate. In Tsou and Wu’s method, the kinetic energy in each joint is used in multiple linear regressions for predicting calorie consumption. The assessment of mass, velocity, and calorie consumption are described in the subsections below, respectively.
Tsou and Wu’s method assumes that the shape of body is universal to all people while in our method, the system obtains mass by analyzing the body shape of each user specifically. Image processing is done on a depth image (An example is shown in Fig. 1), where the ratio of each body segment to the whole body is computed and used to represent the mass percentage of each joint. By multiplying the mass percentage with the weight of the user, we obtain the mass of each part for calculation of the energy. To obtain the mass for each of Kinect’s 20 joints, we used software called Im-ageJ to measure the ratio of the number of pixels in each joint’s area to that in the whole body.
While a user is exercising, the system obtains his/her streaming skeleton data from Kinect (see Figure 1). The skeleton data represent 3D coordinates of all body joints in each row. We set the data frame rate to 25 fps. We derive the velocity of a given joint over a period of time by using the differentiation method.
The differentiation method is wi
This content is AI-processed based on open access ArXiv data.
