Research On Pedestrian Detection And Tracking Based On SSD

Pedestrian detection and tracking,as an important technology of artificial intelligence,has a wide range of applications,especially in the fields of automatic driving,service robot,video monitoring and so on.In recent years,with the wave of deep learning sweeping the field of computer vision,the research of pedestrian detection and tracking has made rapid progress.In pedestrian detection,there are many detection algorithms based on convolution networks.In the aspect of pedestrian tracking,the main research direction is the combination of depth feature and correlation filtering.However,it is still a difficult and challenging task for pedestrians to detect and track pedestrians accurately and quickly because of the factors such as complex and changeable environment,light change,size change,and similar background.In this paper,pedestrian detection and tracking are the main research contents.A pedestrian detection and tracking algorithm based on SSD(Single Shot Multi-box Detector)is proposed.At present,pedestrian detection algorithm based on convolution neural network can not give consideration to both detection quality and detection speed.To solve this problem,a pedestrian detection algorithm based on SSD is proposed in this paper.On the basis of the SSD algorithm,the prior information of pedestrians in traffic scene is introduced and the connection mode of neural network is adjusted,which can effectively alleviate the missing detection problem when the SSD algorithm is detected for small targets.The algorithm first uses the adjusted network to get the initial pedestrian location information and pedestrian feature information.Then,the AdaBoost(Adaptive Boosting)decision forest algorithm is used to further classify the pedestrian frame,and the algorithm can improve the discriminant ability of the difficult to distinguish example.The proposed algorithm has higher detection accuracy and has certain advantages in detecting speed.The main research idea of pedestrian tracking is the combination of depth feature and correlation filtering.The problem of this kind of algorithm is that the extraction of pedestrian characteristics and the training of the filter are fragmented,and the overall advantages of the deep learning framework can not be fully played.We design a new convolution network substitution correlation filter,which integrates the extraction of the target pedestrian feature and the training of the correlation filtering into the depth learning framework,so that the tracking algorithm can be updated with the constant change of the target appearance.Pedestrians encounter and occlusion often occur in pedestrian tracking,which leads to the tracking of erroneous objects.In this paper,we use the main sidelobe ratio to judge whether the target is occluded,and enhance the robustness of the tracking algorithm combined with the motion information of the target pedestrians.In addition,on the basis of the pedestrian detection algorithm proposed in this paper,we use the residual network of the pedestrian re recognition field data set to extract the characteristics of different pedestrians,and calculate the similarity of different pedestrian frames with the location and appearance information of pedestrians,and then realize the tracking of pedestrians.The algorithm shows better results than other algorithms when tracking pedestrians in occlusion.