Research On Pedestrain Detection And Tracking Method Based On CenterNet And DeepSORT

Pedestrian detection and tracking technology plays an important role in intelligent monitoring,virtual reality,unmanned driving and other fields.Detection-based multi-target tracking technology shows great advantages in pedestrian tracking tasks,and its tracking effect is closely related to the performance of the detector used.In pedestrian detection tasks,pedestrian targets are often missed due to environmental interference.The situation leads to low detection accuracy and inaccurate positioning,which in turn affects tracking performance.In pedestrian tracking tasks,due to unpredictable pedestrian trajectories,occlusions between pedestrians or between pedestrians and obstacles are prone to occur,resulting in wrong identity ID switching.Therefore,it is necessary to improve the performance of the pedestrian detection algorithm,optimize the pedestrian tracking model,reduce the number of wrong identity ID switching,and then improve the pedestrian tracking effect.To this end,this paper proposes a pedestrian detection and tracking algorithm based on CenterNet and DeepSORT.The specific research contents are as follows:(1)In view of the missing detection of pedestrian targets in the process of pedestrian target detection due to environmental interference,this paper uses Res Net18 as the backbone network of the CenterNet detection algorithm,and adds CBAM(Convolutional Block Attention Module)hybrid attention mechanism in Res Net18 network to improve the detection algorithm’s ability to locate the detected target.Use the Smooth L1 loss function to improve the loss function in the CenterNet detection algorithm to improve the detection accuracy and recall rate of pedestrian targets.Compared with the original CenterNet detection algorithm,the improved model has a 1.4%increase in precision.(2)Aiming at the wrong ID switching problem in the process of pedestrian tracking,the ordinary Kalman filter is easily affected by low-quality detection,and ignores the information on the detection noise scale.In this paper,the NSA Kalman filter algorithm is used to replace the ordinary Kalman filter.Mann filtering algorithm to enhance the tracking performance of DeepSORT algorithm.The DeepSORT pedestrian re-identification network is trained on the Market-1501 dataset,and the improved CenterNet model is used as the detector of DeepSORT to improve the pedestrian tracking performance.Compared with the original DeepSORT detection algorithm,the improved target tracking algorithm improves MOTA by 1.8%,MOTP by 1.2%,reduces false identity switching by 155 times,and runs at 22 Hz.The experimental results show that the improved algorithm in this paper can effectively reduce the number of wrong identity switching,meet the real-time requirements of pedestrian target detection and tracking tasks,and has certain application value.(3)In order to solve the problem of statistical analysis of pedestrian flow based on target tracking,this paper builds a pedestrian crossing counting system based on the improved CenterNet and DeepSORT pedestrian detection and tracking algorithms,which can accurately track and count pedestrian targets in video sequences.