Design And Implementation Of Video Object Long Sequence Tracking System

With the advent of the era of big data,computer hardware capabilities are constantly improving,and access to data information is becoming more and more convenient.Technologies such as computer vision,natural language processing,and speech recognition are constantly improving,and various applications in the field of artificial intelligence are gradually entering people's lives.With the continuous development of deep learning technology,the core technology of artificial intelligence,object detection and tracking technology based on deep learning has become a hot topic in the field of computer vision,and is widely used in intelligent monitoring systems and Internet mobile terminals.Among them,video object detection and tracking algorithms have always been difficult and hot spots of research,and also have extremely important scientific and practical significance.Due to the increase in video application scenarios and different shooting equipment,there will be problems in the video such as complex background,object occlusion,object ratio change,object and background interference,and real-time requirements.In response to the above problems,we design a object long sequence detection and tracking system,the main research contents are as follows:(1)In the object detection module,we propose a object detection module INYOLOv3 based on YOLOv3(NMS)by studying the principle and network structure optimization of YOLOv3 algorithm,that is,using darknet-53 network for feature extraction,using logistic regression for classification prediction.Then in order to improve the efficiency of processing preselected object frames,we use improved Non-Maximum Suppression algorithm to handle overlapping preselected object frames.It is verified through experiments that the designed detection module has good accuracy and real-time performance for object detection of pictures and videos,and can provide rapid and accurate object detection services for subsequent long-sequence tracking systems.(2)In the object tracking module,we first analyze the principles of the ECO algorithm,then combined with the scale optimization and feature extraction features of the ECO tracking algorithm,a new model update strategy is proposed to reduce the impact of error information on the model and ensure the accuracy of the object information.Experimental comparisons show that the improved algorithm can reduce tracking drift and improve tracking accuracy.Secondly,we propose an improved tracking algorithm based on Siam RPN algorithm.By optimizing the network structure,adding the anchor mechanism of the regional recommendation network and the interference perception model,the algorithm's tracking performance of the object is improved.And by expanding the training samples in the offline training phase,the algorithm's ability to distinguish the target ROI is enhanced,the tracking effect of the tracking algorithm on small objects is improved.The algorithm can adapt to the object's posture changes and the tracking accuracy is improved.(3)Finally,in order to solve the problem of tracking failure caused by the possible occlusion in the long sequence and continue to track the object,we propose a re-detection mechanism,and design a long sequence object tracking system combined with the object detection and tracking module.After the object is lost,the best object position is selected and the tracking is continued by comparing the correlation between the target ROI before the loss and the candidate object ROI.Experiments show that in the occlusion situation,the robustness and accuracy of the system tracking process have been improved.