Research On Detection And Tracking Methods For Dense Targets In Complex Environments

The detection and tracking of dense small targets in complex environments is a challenging problem.It has a wide application background in the fields of UAV aerial target detection and tracking,surface vegetation evaluation,square and other public places security monitoring.At present,researchers at domestic and overseas have carried out a series of researches in the field of target detection and tracking.However,when the video sequence image information of the input detection and tracking network is unstable,the background is complex,and the interested targets are small and dense,the accuracy of the target detection and tracking algorithm will decrease.Therefore,how to effectively compensate the image background jitter caused by shooting equipment and improve the detection and tracking accuracy of dense small targets in complex environments become the core of the problem research.Therefore,the following research work is carried out in this thesis:(1)Research on video global motion compensation based on affine inverse transform model.To solve the problem of blurred video sequence images and jittery background caused by the motion of shooting equipment,firstly,SURF algorithm is used to extract the feature points of adjacent frame images,and then MSAC algorithm is used to obtain the global motion parameters of video sequence.Based on the obtained motion parameters,an affine inverse transform model is constructed to compensate the global motion caused by the motion of shooting equipment in the video sequence.(2)An attention-lightweight YOLOv5 dense small object detection method.Aiming at the low detection accuracy of the classical target detection model for dense small targets in complex environments,a YOLOv5-CBG detection model is proposed based on the image data of dense small targets taken from aerial photographs.Firstly,the rando stochastic pooling layer and residual structure are added to the CBAM attention mechanism to obtain SX-CBAM,which is embedded into the YOLOv5 detection network to improve the network’s feature extraction capability for dense small targets;Then,the Ghost bottleneck module is used as a lightweight module to replace the cumbersome CSP1＿X structure in YOLOv5 detection network,reduce the amount of parameters and calculations in the detection network,and ensure the real-time detection.(3)An improved small target tracking model based on Deep SORT.In order to solve the problem that the existing Deep SORT tracking algorithm faces high track loss rate and large error ID conversion times in the process of dense small target tracking,the feature extraction module and re matching module are improved.Firstly,YOLOv5-CBG model is used as the detector,and the feature extraction and fusion part is reused as the appearance feature extraction network of the improved tracking algorithm to obtain a more comprehensive appearance descriptor of dense small targets;Secondly,CIo U is used to replace Io U as the measurement between the unmatched detection box and the unmatched track,so as to better avoid the error deletion of the track and target by the algorithm and obtain better tracking effect;Finally,the performance of the improved algorithm is verified on the MOT-16 dataset.The experimental results show that the MOTA index of the improved algorithm is improved by 3.6% and the error ID conversion is reduced by 269 times in the process of tracking dense small targets,which proves the effectiveness of the improved algorithm.