Research On Pedestrian Tracking Algorithm Based On The Fusion Of Deep Learning And Correlation Filtering

Pedestrian target detection and tracking technology has important research value and application value in security monitoring,intelligent transportation,human-computer interaction and other fields.In the practical application environment,due to the diversity of tracking scene and the complexity of pedestrian target change,the traditional target tracking algorithm with kernelized correlation filters has insufficient tracking performance in videos with fast motion,scale variation,occlusion and other attributes.When the target tracked is not accurate,updating filter model will reduce the quality of its detection quality.If this happens in several consecutive frames,it will cause model drift which leads to tracking target loss and error spread.To solve these problems,this paper proposes a grouped kernel correlation target tracking algorithm based on deep learning detection.The main contents and achievements are as follows:(1)Aiming at tracking target loss and error spread caused by the actual tracking environment,this paper proposes a grouped kernel correlation target tracking algorithm based on deep learning detection.Firstly,the algorithm divides the video sequence to be tracked into several groups.The Faster RCNN algorithm is used to detect the target in the first frame of each group.If the target to be tracked is detected,the filter is trained and updated with the detected feature information;otherwise,the filter model is trained and updated with feature information tracked by algorithm with kernelized correlation filters in the previous frame to complete the follow-up target tracking.(2)Due to the poor robustness of and-crafted features used in traditional target detection algorithms,this paper uses Faster RCNN target detection algorithm based on deep learning to complete pedestrian detection.In this paper,the Caltech data is used to complete the training of pedestrian detection network model under the Caffe deep learning framework and VGG network model.The experimental results show that the method has a good detection effect for pedestrians with different scales,different density and different attitude in the images.In order to verify the effectiveness of the improved algorithm,this paper selects the video sequences containing pedestrian targets in OTB100 database and UAV123 database for comparative experiment.The experimental results show that the performance of this algorithm is relatively optimal.The accuracy of the improved algorithm in OTB 100 database is 20.5%higher than that of the original algorithm with kernelized correlation filters,and the success rate is 17.8%higher;in UAV123 data set,the accuracy and the success rate of the improved algorithm are 26.1%and 18.5%higher,respectively.Both the tracking speed are over 100fps,which meets the real-time tracking requirements of the actual scene.