Real-time Object Track And Application Based On Template Updating And Feature Enhancement

Video object tracking is a task of recording the location and movement of specific objects in video.As an effective means of maintaining social security and assisting public security in investigating cases,it is receiving increasing attention.Object tracking in complex scenes is susceptible to environmental interference factors such as occlusion,light changes,angle of view changes,and rotation.The introduction of deep convolutional neural networks to obtain more robust features can alleviate these problems to some extent,but it is difficult to meet the application requirements of object tracking in real scenes characterized by significant changes in the appearance of objects and frequent occurrence and disappearance of objects,making object tracking face bottlenecks.The template matching tracking model based on siamese networks has the advantage of strong real-time performance,but it often uses the first frame as the matching template,which is difficult to adapt to the drastic changes in the appearance of the object,the disappearance of the object,and the interference of similar objects.In order to maintain the real-time nature of the tracking model,this article uses a template matching network as the main network,supplemented by a re detection network to respond to drastic changes in the appearance of the object and the disappearance of the object.At the same time,reinforcement learning is used to optimize template selection,and Transformer is used to better utilize the context sensitive information of the video.The specific work of this article is as follows:(1)Aiming at the problems of sharp changes in the shape of tracking objects and easy loss of objects,this paper first proposes a object tracking model combining template matching network and re detection network.The problem of template selection in template matching networks is modeled as a Markov decision process,and the optimal strategy for template selection is optimized through proximal strategy optimization algorithms in reinforcement learning to better respond to changes in the appearance of the object.In order to avoid tracking failure caused by object disappearance in the tracking model,a re detection network for object detection is introduced,and a binary restart judgment model is constructed.When the restart judgment model determines that the object has been lost,immediately start the restart detection network to retrieve the object and update the template in the template pool;Otherwise,a template matching network is still used for tracking to ensure strong real-time performance of the tracking model.Combining the two mechanisms mentioned above,the proposed tracking model improves the OTB-2015 dataset by 2% compared to Siam FC++based on siamese networks,and is also competitive compared to other mainstream algorithms.(2)To solve the problems of tracking objects with sharp changes in shape and being easily interfered by similar objects,this paper first introduces a template pool memory storage mechanism.Combining multiple appearance features into memory features as a matching basis improves the adaptability of the algorithm to appearance changes.Although the template selection mechanism of innovation point 1 is faster,its accuracy is not as good as the template pool memory storage mechanism.The two mechanisms are suitable for different scenarios.Secondly,in order to avoid tracking tasks being interfered by similar objects,this paper proposes a Transformer framework to enhance appearance features,and establish a temporal relationship between template features and search domain features to reduce interference from similar objects.The object tracking algorithm combining the two mechanisms has achieved excellent results in data sets such as GOT-10 k,La SOT,UAV123,and Tracking Net,and is also competitive compared to mainstream algorithms.(3)Based on the proposed two algorithms,this article combines the current demand for traffic video analysis to develop a traffic video monitoring and analysis system,and deploy it on the desktop.Users can input traffic video through the interface,and the system can monitor fixed pedestrians or vehicles,as well as real-time monitoring of vehicle and pedestrian flow,recording and analyzing traffic conditions under fixed cameras.The algorithm proposed in this paper is based on the siamese network tracking algorithm,ensuring real-time performance while possessing strong robustness,and is suitable for complex scenarios.The proposed system reduces manual workload and can be applied to different traffic scenarios using adaptive handover algorithms.Therefore,the research content of this article has certain theoretical value and good application prospects.