Research On Pedestrian Target Tracking Method Based On Siamese Network

Pedestrian target tracking technology is one of the important research directions of computer vision,and has important applications in intelligent video surveillance,automatic driving and transportation.However,in the actual tracking scenario,due to a variety of complex distractor,it is difficult for the existing tracking methods to achieve accurate tracking.In recent years,the pedestrian target tracking method based on twin network under deep learning algorithm has attracted attention because of its high accuracy and computing speed,but the traditional siamese network tracking method,due to the imperfection of pedestrian training dataset,lack of redetection mechanism,lack of effective feature selection,etc.,its tracking robustness and accuracy under complex distractor will be significantly reduced.In order to solve the above problems,the main research content of this dissertation is as follows:Firstly,six complex disturbances,such as pedestrian occlusion,similar background interference,pedestrian attitude change,lighting change,scale change,fast motion and motion blur,which often occur in the current pedestrian target tracking process,are analyzed in depth,and they are classified into two categories:complex background information interference and semantic information complex interference.The existing problems of the current pedestrian target tracking method are analyzed,and the improvement strategy is proposed.Secondly,a siamese network pedestrian target tracking method combining distractor-aware training and re-detection is constructed,which can solve the complex distractor problem of background information such as pedestrian occlusion and similar background distractor.In the offline training stage,considering that the training sample determines the feature learning direction,this method adopts a distractor-aware training mechanism to train the tracking model from three aspects:location distractor-aware,occlusion distractor-aware and semantic distractor-aware.In addition,in the online tracking stage,a re-detection mechanism is introduced for the problem of similar background distractor,which effectively realizes the screening and elimination of similar background.Experiments show that this method can better cope with the problems of pedestrian occlusion and similar background distractor,and effectively improve the tracking accuracy.Finally,a siamese network pedestrian target tracking method based on selfattention mechanism is constructed,which can solve the complex distractor problem of semantic information such as pedestrian attitude change,illumination change,scale change,fast motion and motion blur.Aiming at the problem that the tracking effect of shallow convolutional neural networks decreases significantly under complex distractor,this method adopts an improved deep residual convolutional neural network to replace the traditional shallow network to obtain more image semantic features.In addition,a feature selection module is designed using the channel self-attention mechanism and the spatial self-attention mechanism,which can adaptively screen the features of the target.The effectiveness of the proposed method is proved by experimental results,which can achieve accurate pedestrian target tracking.