Research On Target Detection And Tracking Algorithm Based On Video

Visual target tracking is an important research direction in computer vision and has a wide application prospect.Although visual target tracking technology has made great progress in the past 20 to 30 years,especially with the development of deep learning in the past two years,video target tracking method has made some progress.However,many of the twin network architectures adopted by the tracking network use the old network skeleton,which is not conducive to the continuous optimization of the network architecture.Therefore,it is of great significance to study how to complete higher level image comprehension tasks more efficiently based on deep learning algorithm.Based on the existing problems of video tracking,this paper explores and studies the target tracking framework based on deep learning.Based on the current situation of shallow skeleton layers in twin network with single target tracking,in this paper,the network skeleton is replaced by other networks with different depths(residual networks,etc.),the factors affecting the performance of twin networks are discussed from different angles,such as network depth,sensing field,filling or not.According to the experimental data,this paper summarizes the methods of optimizing the performance of twin networks as follows: The receiving field for the output feature should be set in proportion to the size of the sample image.When designing the network architecture,the network step size,the perception domain and the output characteristic quantity should be considered as a whole.For fully convoluted Siamese matching network,it is very important to deal with the problem of perceptual inconsistency between two network streams.In this paper,the residual element and the down sampling element of the original residual network are used.Based on the above optimization of twin network,the factors(filling,step size,sensing area,etc.)that are not conducive to the final performance of the twin network are improved.The filling effect of the primary cell is reduced,and the parameters are reduced.And by reducing the step size,the network can feel more features and make it more fit with Siamese architecture.A new video target tracking network,NewResNet,was built based on the optimized residual unit and the down sampling unit.This paper uses the vot-2016 benchmark video data set to set up a control test,and evaluates the impact of different factors in the optimization module on network performance.The experimental data prove that the optimized module has better performance than the original module,and verify the correctness of the above-mentioned twin network influencing factors.In this paper,the performance of the new network was tested under the vot-2016 benchmark video data set.The performance of NewResNet was compared with the original network SiamFc and other networks through the three indicators of accuracy,robustness and expected average overlap.The test results show that: compared with the original twin network architecture SiamFc,the new network has better accuracy and expectation average overlap index,lower robustness,but less real-time performance,and better comprehensive performance than SiamFc.In comparison with various networks,the performance of the new network ranks above the average and is better.