Pedestrian Tracking Algorithm Based On Convolutional Neural Network

Target tracking is an important research direction in the field of computer vision,which is widely concerned by domestic and foreign researchers.Pedestrian tracking,as a very important branch of target tracking,is one of the research hotspots in the field of pattern recognition and machine learning.It is widely used in intelligent monitoring and other fields,and has very important application value.With the advent of the era of big data and artificial intelligence,deep learning has achieved great success in the field of image classification and recognition.Correspondingly,the pedestrian tracking algorithm based on deep learning also has developed rapidly.In recent years,trackers based on the siamese neural network have attracted much attention,because they perform well in tracking accuracy and speed.But there are still some problems to be solved.For example,most Siamese trackers' and improved Siamese trackers' backbone networks directly adopt the neural network designed for target classification task and lack the ability to distinguish similar objects.When the pedestrian target is disturbed or occluded by the adjacent pedestrians,the algorithm has a high probability of losing the correct target and tracking other disturbed pedestrians.Secondly,the Siamese tracking method transforms the single target tracking problem into a one-shot target detection task,failing to make full use of the time-domain correlation of visual tracking.Thirdly,as the fixed frame of video sequence is adopted as the tracking template in the Siamese tracking method,no template update mechanism is provided during the tracking process,which can not deal with the drastic changes in the appearance of the pedestrian tracking scene.This thesis mainly studies the pedestrian tracking algorithm based on deep learning Siamese networks.First,Siamese tracking method takes classification neural network as the model subject,which leads to lack the ability to distinguish similar objects.This thesis introduces the seemingly non-essential features to identify different pedestrians.SiamRPN tracker is used as the baseline,which introduces the region proposal network after the Siamese network,and the apparent information,color information,is added to solve the problem of other pedestrian interference.Experimental results show that the proposed algorithm is comparable to the existing improved Siamese tracker,who improves tracking accuracy mainly by using a deeper and more complex network.And this provides a lightweight and effective new idea for improving pedestrian target tracker performance.In addition,for problem that Siamese network only makes use of the spatial information within the frame,Kalman algorithm is applied to make effective use of inter-frame information,which effectively solve the problem of pedestrian target moving fast,blocking and semi-blocking.Finally,aiming at the template fixation problem in the tracking process,a template updating mechanism based on similarity score of target position predicted by SiamRPN algorithm is designed.When the similarity score is lower than the set threshold value,the matching features between the pedestrian in the template frame and in the detected region is considered to be insufficient,and the template is replaced with the recently successful tracked target.Experiments show that this strategy can effectively solve the tracking problem caused by the drastic change of the object's appearance in the pedestrian tracking scene.