| Object detection is one of the basic tasks of machine vision,and its main purpose is to identify specific object classes in a scene and determine the location of object bounding boxes.Nowadays,due to the continuous upgrading of hardware devices,a large amount of video data is generated,and video object detection technology has great potential for application in areas such as automatic vehicle driving,Internet video auditing,and security surveillance data auditing and retrieval,so video object detection is of great significance for practical engineering applications.In recent years,with the success of deep neural networks in various fields,the accuracy and efficiency of video object detection technology based on deep neural networks has made breakthroughs and has been used successfully in some fields such as abnormal event monitoring,industrial detection,and intelligent security.Video object detection is a key technology in video processing,which is very important for video information decoding,so video object detection has become a hot issue for research in recent years.Compared with image object detection,the difficulty of video object detection lies in how to effectively extract and use spatio-temporal information and solve the degradation of detection performance caused by blurring,deformation,and occlusion due to motion,all of which put forward higher requirements for video object detection algorithms.This thesis addresses the characteristics of video data with spatial consistency and temporal continuity,combines the current popular deep learning networks,while using sparse coding to extract spatio-temporal information of video,and conducts researches and discussions on feature fusion mechanism and region proposal network in video object detection.The main contents of this paper are as follows:(1)A video object detection based on low-rank and sparse decomposition is proposed.To address the problem that the video object detection based on frame-level feature aggregation network lacks attention to local regions,which makes it difficult to detect blurred,large deformed or occluded objects in the video accurately,we combine low-rank and sparse decomposition with deep learning by taking the advantage that low-rank sparse decomposition can effectively separate the foreground objects of video motion.The foreground moving prior of the video is extracted,and then the features learned through the deep network are enhanced using the foreground sparse prior to improve the detection performance of moving objects in the video object detection.The validation analysis on the Image Net VID dataset shows that the accuracy of this method outperforms some existing methods and demonstrates the effectiveness of the sparse foreground prior for the video object detection.(2)A video object detection based on sparse prior guidance is presented.Most existing anchor-based object detection methods exploit the dense anchor mechanism to preset a large number of anchors on the feature map to ensure a high recall of detection results,which not only generates many negative samples but also increases the computational effort.Considering the problems caused by this dense anchor mechanism,we design a region proposal network based on sparse anchor,which uses the sparse information obtained from the non-negative low-rank sparse decomposition model to guide the generation of fewer and more precise anchor during the proposal extraction process,and thus obtain high-quality proposals focusing on the object of interest.The feature fusion module in(1)is also improved to make better use of the spatio-temporal information of the video through the spatial attention.It is experimentally verified that the method can not only enhance the representation of indistinguishable objects in the video,but also reduce the computational cost in the anchor generation with good detection accuracy and efficiency.(3)A video object detection based on deep sparse learning is proposed.Although the methods in(1)and(2)can effectively use sparse representation to improve video object detection performance,they cannot achieve overall end-to-end training and have more limited generalization capability.Therefore,we attempt to adopt deep networks to implement traditional sparse representation methods and deep sparse learning of features in video frames to achieve an end-to-end deep sparse learning based video object detection.This method utilizes deep sparse learning branches to obtain sparse encoding,and then fuses the global features of video frames with the sparse encoding,so as to enhance the representation capability for objects in the global features and improve the generalization capability of the deep sparse learning based video object detection method. |
