Research On Moving Target Detection And Tracking Based On Image Processing

With the development of theory and technology,the needs of science and technology to serve human production and life have become increasingly diverse and complex.Video target detection and tracking,which is the core task of computer vision,has changed people's lifestyles to varying degrees,and a large number of successful practices have been applied to many fields.The detection and tracking of targets has become a research direction that has attracted much attention.In recent years,breakthrough developments and excellent results have also appeared,especially based on deep learning-based convolutional neural network algorithms and correlation filtering-based tracking algorithms Both target detection and tracking tasks have shown good performance.However,due to the complexity and diversity of the use scene,a variety of uncertain factors such as fast motion,scale transformation,motion blur,and target occlusion have seriously affected the performance of visual target detection and tracking algorithms.In this paper,the following researches are performed on the challenges of omission of small targets in the target detection process and occlusion and scale transformation in the target tracking process.(1)This paper summarizes and classifies classic target detection methods,analyzes the basic principles of convolutional neural networks,and analyzes classic detection algorithms based on regional suggestions and deep regression networks.Afterwards,for the low number of low-level feature convolution layers of the SSD algorithm,the existence of insufficient feature extraction led to the improvement of the poor detection of small targets.Inspired by human visual mechanisms,the network was enhanced by adding artificially designed receptive field modules Larger area feature extraction capabilities.Multi-scale dilated convolutions are used in the new module to increase the range of receptive fields.Borrowed from the idea of Inception on the structure,adding and replacing part of the original network's feature convolutional layer on the original lightweight backbone network.Finally,experiments on the VOC data set proved the average detection accuracy ratio of the improved algorithm The original algorithm improved by 2.6%,and basically maintained real-time speed.(2)A deterministic visual target tracking algorithm based on the mean drift algorithm is studied.Aiming at the loss or drift of the target when it is blocked or moving fast during the tracking process,a Kalman filter method is introduced to predict the target at subsequent frames.The possible positions in the algorithm reduce the number of algorithm iterations.By comprehensively considering the two algorithms,the accurate position of the target is obtained.The results show that compared with the single algorithm,the improved algorithm makes the tracking success rate effective and accurate.(3)Based on a deep understanding of the kernel correlation filtering algorithm,the KCF algorithm uses a single HOG feature as the target feature.Although it can capture the local shape information of the target area better,when there is more noise in the target background,HOG The ability to describe features is weakened,and its representation of target features is no longer obvious,while color features are better for tracking target deformation and motion blur.It is proposed to combine HOG features and CN features in KCF to enhance target local features and color information.Expressive ability.Then,the estimated position of the target is obtained using KCF,and the maximum response value of the scale filter output in the DSST algorithm is integrated to achieve the target's scale adaptation.Finally,experiments are performed to compare the algorithms of KCF,SAMF and DSST,and it is proved that the improved algorithm is more robust than several other comparison algorithms.