Research On Online Tracking Algorithm Based On Deep Learning And Correlation Filtering

With the development of computer vision technology,video surveillance plays an increasingly important role in intelligent security,enterprise production,and intelligent transportation.To reduce the impact of multi-scale changes of pedestrians in the tracking process,this paper studies the online object tracking algorithms and uses multiple correlation filters to assist tracking,as well as develops auxiliary and constraint relationships among correlation filters to solve multi-scale tracking ploblem.The deep learning feature and the edge detection feature are combined to obtain more powerful features,which make the performance of the object tracking algorithm improved,and the research content of the subject is of positive significance.This paper proposes an improved object detection method based on Light-Head R-CNN algorithm.The detection algorithm requires high detection accuracy and speed.The detection accuracy and speed of the current one-stage detection algorithms and the two-stage detection algorithms are ideal,but under the premise of ensuring the detecting speed,the detection efficiency of smaller objects needs to be improved.This paper proposes a method of clustering the object size of the target scene to be detected and setting a suitable size candidate frame to enhance the detection ability of the detection algorithm for small objects.By experimenting with different training strategies,the detection accuracy of the detection algorithm is improved,and the detection of small pedestrians in complex scenes is better.As for the use of features,this paper studies the fusion of convolutional neural network features and edge detection features.On the basis of the analyzation of the components,working principle,forward propagation and backward propagation theory and network training strategy of the convolutional neural network,after visualising the network features,the method of fusing features of convolutional neural network and Canny features is proposed.Experiment shows that the fusion method works well on the twinning network tracking algorithm,thus this paper applies it to the tracking algorithm based on correlation filter and achieves good results.In this paper,an improved algorithm based on the KCF algorithm is proposed.Improvements mainly include three aspects.Firstly,auxiliary among correlation filters is utilized to solve the multi-scale target tracking problem.On the first frame,with the auxiliary constraints among correlation filters,the training of the filters of pedestrian head,hip and body are implemented at the same time to utilize as much object information as possible.And then the object scale can be obtained through the distance emerged from tracking results of head filter and hip filter.Secondly,if the range of the searching area is fixed,the object can hardly be located when the object scale changes during tracking.Then the variable padding range is proposed,which changes according to the distance between different filters and can provide a searching area with proper size for correlation filters.Thirdly,an effective two-stage searching method is proposed for tracking in certain area,and this method can reduce the time comsuption during tracking to some extent According to the experiment results on some pedestrian sequences of VOT2016 data set,the average overlap expectation(EAO)of the improved algorithm in this paper reaches 0.208.Finally,the research work carried out on this topic is summarized,and the future research direction is forecasted.