Research On Small Target Tracking Method Based On YOLOv5+DeepSort

With the rapid development of deep learning technology,based on the detection of multiple target tracking algorithm on the large and medium-sized target tracking accuracy and precision,shows great advantage,but because of various objects in the image size,posture,and appearance is not the same,plus shade effect,background interference,and so on the reason,give the small target tracking algorithm research has brought the huge challenge.In addition,detection-based small target tracking algorithm is extremely dependent on the performance of detector,while small targets are easy to miss and misdetect due to their too small proportion in the image,resulting in low detection accuracy and inaccurate positioning,which greatly reduces the tracking performance.Therefore,it is necessary to further improve the tracking effect on the basis of improving the performance of small target detection.In this paper,small target tracking for pedestrians and vehicles is mainly studied as follows:(1)Aiming at the problem that small targets occupy a small area in the image and are easily disturbed by background factors,attention mechanism is introduced on the basis of YOLOv5 to enhance the ability of small target feature expression;To solve the problem that the detailed features of small targets are lost due to multiple down-sampling operations,a P2 small target detection layer is proposed to strengthen the feature information fusion between high and low levels.Through comparative analysis of experimental data in Vis Drone2019 data set,it can be seen that YOLOv5＿CA＿P2 model,which integrates attention mechanism and small target detection layer,has improved accuracy and recall rate in small target detection and has better practicability.(2)The improved YOLOv5＿CA＿P2 model is used as the detector of the small target tracking method.Aiming at the problem that the kalman filter in the original DeepSort has poor prediction effect on the non-uniform motion model,the acceleration parameter is introduced to rebuild the uniformly accelerated motion prediction model,and the prediction accuracy rate of the trend target with non-uniform motion is improved.In order to solve the problem of target loss or identity switching caused by occlusion and other factors,the full-scale feature extraction model OSNet was used to train the re-recognition model of specific scene to improve the expression of apparent feature,and the data association matching was combined with the motion feature to improve the overall accuracy of the tracking algorithm.Through comparative analysis of experimental data on Vis Drone2019 tracking data set and self-made human-vehicle data set,it can be seen that the small target tracking algorithm based on YOLOv5+DeepSort improved in this paper has improved tracking accuracy and accuracy rate,and is more advantageous in practical tasks.(3)Based on the improved algorithm of YOLOv5+DeepSort,a small target tracking system based on uav perspective is built,which can track small targets of human and vehicle in video sequence,and the ID switching rate of the same target is low.