A Real-time Embedded Target Tracking System Based on Deep Learning Model

Target tracking has been arousing increasing attention in current production and life. Usually, target tracking algorithms are deployed on high-performance servers. There are significant challenges in deploying real-time target tracking in embedded systems. Therefore, a real-time embedded target tracking system is studied based on a deep learning model in this paper. Firstly, the hardware implementation of the proposed system is presented in detail. Then, to resolve the problems of high latency and low performance in embedded platforms, we offer an adaptive detection strategy that combines the advantages of deep network model inference and prediction. This strategy makes full use of the target relevance between the front and rear frames of video, reduces the frequency of calling the depth model, and improves the system's speed. Finally, Experiments have proved that the system can effectively reduce the time delays by more than 50% while ensuring the same detection effect. Meanwhile, the system's power consumption and resource occupancy rate are also significantly improved.