Research On Motion Object Detection Based On Video Context

While the popularity of video surveillance soars,massive video data have become an important source of unstructured data for big data application.Automatically detecting motion object is an important research topic of large scale video surveillance,which is of great significance to the subsequent pattern recognition and behavior analysis.Motion object detection has been widely used in human-computer interaction,pedestrian and transportation and other applications.However,the inconsistent object scale,interference from complex background and illumination variation and other factors often enlarge miss-detection rate and error-detection rate,which hinder the promotion of application.Targeting the above problems,some robust algorithms are proposed,e.g.SVM,Faster RCNN and SSD,which,however,still face performance bottleneck due to the complexity of application scenes.Detecting timber transportation vehicle is a typical case of motion object detection with complex scenes,which shows object diversity,scale inconsistent,complex background and illumination variation.Therefore,in addition to evaluate algorithms on the public dataset,the proposed algorithms are tested and verified for detecting timber transportation vehicle on more rigorous and complicated forest scenes.The proposed research is inspired by the fact that features of motion object have strong correlation in the context of video,which is then used to provides adaptation for algorithms for the complex background and illumination variation.And the scale adaptability of algorithms are improved with an adaptable network model to suit scale change of the motion object in the shot.The specific research contents and innovations of paper are as follows:(1)Background modeling of moving vehicles.Vi Be algorithm has the advantages of high computing efficiency and easily implementing.But it is susceptible to ghosts and stationary object and so on.The paper combines the Vi Be algorithm with a MSRCR(Vi Be-M)to enhance processing power of ghosts and stationary object.The experimental results are compared with Vi Be algorithm and Vi Be-otsu algorithm,which show that the proposed method better improve accuracy under complex background and illumination variation.(2)SSD object detection based on context.The paper specifically studies SSD algorithm and the relationship between the m AP of object detection and the feature map numbers.DRF(Detection with Refined Feature)detector based on the SSD algorithm is proposed which adds center map and scale map.DRF fusions the features of each layer network and enrichs object feature through contextual information.Compared with existing algorithms,the proposed method significantly improves classification confidence and accurate location of the object on the Pascal VOC dataset.In addition,The paper combines the SSD algorithm based on the DRF detector with the generalized intersection over union(GIo U)to detect the timber transportation vehicle and optimize the m AP when prediction boxes and ground truth boxes are disjoint.(3)Detection of timber transportation vehicles based on adaptive context.The paper mainly discusses the advantages and disadvantages of Faster RCNN,YOLOv3 and YOLOv3-tiny algorithms and improves the tiny network to adapt to the scales of timber transportation vehicles.The proposal areas of the motion object are determined by obtaining the contextual feature information.Feding the proposal area into the network decrease the miss-detection and error-detection rate of timber transportation vehicles under illumination variation and complex scenes.Therefore,the accuracy of object detection is further improved.