Research And Implementation Of Single Object Tracking Based On Deep Learning

With the rapid development of information technology,the problem of public security in our country has not been solved,and as people pay more attention to their own security and social security,more and more risks are waiting for the supervision department to solve them.Video monitoring system is the basic means used by the government or enterprise to control and discover the hidden dangers of security.It has become an important content of the new intelligent city construction,and the key technical problems involved in the video monitoring system have been widely studied.Target tracking is one of the basic technologies of the monitoring system,which will provide support for the upper technology of behavior analysis,anomaly detection and so on.In addition,deep learning is widely used in the field of computer vision and has obtained important research results.Therefore,the research and application of deep learning in target tracking is of great significance for the development of target tracking technology.The main research difficulty in target tracking is the impact of complex interference factors on target recognition in a video scene,and the changes of the target in the process of motion,such as occlusion and deformation.There are many problems that have not been solved for these complicated interference factors.The study mainly studies how to improve the representation ability and occlusion of targets,the main research contents are as follows:(1)An improved correlation filtering algorithm based on high confidence is studied and implemented.In the target feature processing,the background features are extracted from the upper and lower four directions to extract the related background features to enhance the ability to express the target.The color histogram model is used to extract the color feature of the target,and the feature value of the target is extracted in the other dimension.The two-channel is used to calculate the response map at the same time and predict the position of the target.The model is updated according to the oscillation degree of the response map.When the overall confidence degree is high,the model of the two channel is updated to reduce the interference caused by the error information.The optimization strategy improves the integrity of the feature,and reduces the updating of the model to a certain extent,and improves the tracking speed.Experiments prove that the improved strategy improves the accuracy of model tracking.(2)A deep residual learning tracking model based on feature fusion is studied and implemented.The model uses the depth learning network to extract the multidimensional features of the target,and uses the thought of the residual network to construct the network layer for fast learning,and calculates the target location according to the information of the next frame.The three-channel network is designed to extract the characteristics of different dimensions of the target,and train the corresponding network at the same time,to improve the accuracy of the tracking model.In addition,the model is supplemented by adding the feature of the extra network channel to extract the color histogram,and the multi dimension feature fusion strategy uses the linear fusion method to set different weights according to the importance of the feature.Experiments show that the model has been improved in various indicators,the model has better robustness,and it has the better adaptability to the video scene with the human target.By adjusting the feature fusion strategy,the ability of the model to deal with occlusion is improved.