Research On Target Tracking Algorithm Based On Deep Learning

Object tracking technology has always been one of the important branches in the field of computer vision.Tracking a specific target in a continuous motion video sequence is the main problem of target tracking research.Many researchers have done a lot of work.This paper focuses on the work of deep learning on target tracking,and studies a large number of excellent algorithms in recent years.Based on some shortcomings of current research,this paper carries out the following work.Aiming at the problem that the traditional deep learning object tracking algorithm uses multi-sample tracking to cause large computation,a deep direction network target tracking algorithm is proposed.We first construct a deep direction network.The deep direction network output consists of two branches,one branch outputs eleven class direction categories,and the other outputs positive and negative sample scores.We use the output direction category to judge the next operation,then the sliding window operation slowly approaches the target position.Until the sample output is true,it is considered that the depth direction network has found the target.The deep direction network also takes the target scale change as an output category and does not require separate training of the scale filter.The positive and negative sample branches exist for the re-detection mechanism when the target is lost.When the positive sample score is lower than the threshold,the target tracking is considered to be lost,and the re-detection mechanism is used to re-detect the new position of the target.The use of the deep direction network ensures that only one sample is calculated at a time,which reduces the amount of computation for deep learning,thereby improving the tracking speed.The re-detection mechanism can retrieve the target when the target is lost,making the algorithm more robust.The algorithm results are verified on the OTB dataset,and the algorithm achieves certain expected results.In view of the network fine-tuning problem and the advantages of using traditional artificial features,a time domain information object tracking algorithm based on global artificial features is proposed.In order to reduce the redundancy calculation,the global CN feature is extracted from the original image,then CN feature are reduced in dimension by the PCA algorithm.The candidate samples are extracted on the dimension reduction feature map,and the candidate samples with the highest score in the depth network are used as the final tracking result.During the tracking phase,a time domain information sample set is initialized and the time domain information sample set is updated over time.The time domain information sample set is used to fine tune the network when the target tracking score is below a certain threshold or continuously tracking a certain number of frames.Experiments show that this algorithm has its unique advantages in dealing with video sequences with some challenge factors,and it has good robustness against some classical algorithms.