Research And Implementation Of Lightweight Target Tracking Algorithm For IoT Terminals

Nowadays,intelligent transportation based on object tracking plays an increasingly important role in people’s lives,which makes a large number of target tracking systems gradually deployed into our daily traffic.Considering factors such as the cost and the tracking ability,the existing target tracking system mostly adopts the client-server-client deployment method.In this way,the IoT terminal is only a platform for collecting data and presenting the results,and the cloud server performs the calculation of the target tracking results.Such systems with a high degree of coupling cannot meet the needs of time-sensitive services.We study and design a lightweight target tracking algorithm for IoT terminals based on Deep Sort,we also implement a real-time target tracking system.The system is free from the constraints of cloud servers and networks so that all processes are carried out on the IoT side,which improves the real-time efficiency of the target tracking system.Directly deploying the target tracking algorithm is a huge challenge for IoT terminals with poor performance.Therefore,the core goal of this thesis is to use the Deep SORT algorithm as the main tracking method to lighten the algorithm.Firstly,the basic detector of the Deep SORT algorithm is optimized,and an improved target detection algorithm based on YOLO v4 is proposed.Referring to the structure of the basic module of MobileNet,a lightweight network is designed to replace the backbone network of the original YOLO v4.At the same time,the idea of depthwise separable convolution is used to optimize the ordinary convolution in the feature fusion network,and the channel branch reduction technology is used to further slim down the overall network.Finally,the improved algorithm based on YOLO v4 has a considerable or even a slight improvement in detection accuracy,while the detection speed has been greatly improved.Then,based on the lightweight of the basic detector,we also improve the Deep SORT algorithm process.Firstly,the extended Kalman filter algorithm is used to improve the prediction of the detection frame,which makes the prediction of the detection frame position in the nonlinear scene more accurate.Secondly,the generalized intersection ratio is used to replace the ordinary intersection ratio measurement method used in the algorithm process,so that the relative position measurement of the two targets is more accurate.Finally,considering the motion state of pedestrians and vehicles in real traffic,we propose an adaptive tracking algorithm based on motion judgment.Combined with the real motion state of the target,the Deep SORT process is improved and the call of the target detection algorithm is reduced.frequency,which improves the efficiency and speed of the tracking algorithm.Finally,based on the above improved Deep SORT target tracking algorithm,we design and implement a real-time target tracking system,which consists of only one embedded terminal and does not require the participation of servers and networks.In our embedded system,Data processing,tracking,and post-processing all form a closed loop.Ultimately,field tests show that the system’s detection capability and detection speed are usable in real traffic.