The Realization of Moving Target Tracking in Monitoring Video based on Deep Leaming

Abstract

The purpose of this research is to expand the application field of deep learning, to apply deep learning to target tracking problems, to conduct an in-depth study on the monitoring video detection and recognition methods of suspect targets. In this research, from the perspective of feature extraction, the deep learning algorithm that can better reflect the essential features of objects was deeply understood. Then, the stack noise reduction self-encoder was trained with the offline training method without supervision. On this basis, support vector machine (SVM) classifier was added to improve the accuracy of target tracking, and the trained self-encoder was used to extract object features from the deep network. Finally, the deep learning algorithm of “offline training + online fine-tuning” was used to track the moving target in monitoring video, and the robustness performance of the algorithm in this experiment was compared with other kinds of the experimental algorithm through the visual tracker benchmark (VTB) data set. The results showed that the accuracy rate and success rate of all sequences of target tracking algorithm based on deep learning were more than 60%, and their accuracy rate and success rate were higher than other algorithms. And the accuracy and success rate of inter-robustness evaluation of target tracking algorithm based on deep learning were much higher than other algorithms. The tracking algorithm proposed in this research had a higher success rate and accuracy rate than many excellent algorithms such as direct linear transformation (DLT), circulant structure kernel (CSK) and tracking learning detection (TLD) in one-time pass evaluation (OPE) and time robustness evaluation (TRE), and the performance of video sequence attribute algorithm was obviously higher than other algorithms. Target tracking algorithm based on deep learning had good accuracy, robustness, and tracking effect, so it was worth further exploration and research.