Research On Video Object Detection In Complicated Scenes

As a kind of science that uses computer to process visual data,computer vision occupies an increasingly important position in human production and life with the continuous improvement and development of human science and technology.The goal of computer vision is to make computers have the same ability of processing visual information as human beings.It is one of the important research directions of artificial intelligence.As a basic task in computer vision,video object detection has been used widely in video surveillance,manless driving,human-computer interaction and so on.This made video object detection as the basis of many intelligent video analysis tasks.Therefore,the study of video object detection in complex scenes can improve the overall performance of computer vision related tasks.For decades,video object detection technology continues to develop,gave birth to some excellent video object detection algorithm.To some extent,these algorithms solve the problem of video object detection in some scenes,but most of them can only get good detection results under certain conditions and require a lot of tuning skills.For complex backgrounds and challenging challenges The complex scene lacks a sufficiently efficient and robust algorithm.Therefore,the study of video target detection algorithm in complex scenes is a challenging task.The main work is as follows:(1)A moving object detection approach based on recurring pattern prior voting is proposed to handle the problems of traditional moving object detection methods,which are susceptible to background interference and misclassification in complex scenes.The recurring pattern prior voting is based on the observation that the same patterns tend to recur frequently in same semantic regions(background or foreground).Through introducing this prior,we can detection moving objects robustly to the background interference in complex scenes.First,a Gaussian mixture model(GMM)is created for each pixel in the video frame.After the model parameters are initialized,the GMM is used for each subsequent frames to generate a rough background probability map,and then we segment video frames into superpixels.The superpixel-level foreground probability map is calculated in a semi-supervized way given the above pixel-level ones.In particular,we construct a spatially constrained graph to take both neighborhood and long-range relationships among superpixels into account,and then employ the manifold ranking algorithm to generate the superpixel-level foreground probability map.Finally,the probability map is processed by a threshold to generate a detection results.By comparing with other moving object detection methods on the standard video sequences,we conclude that our approach has good performance in dealing with background interference.(2)Aiming at detecting specific objects in a video,we propose a novel approach based on the 3D convolutional neural network(CNN)and the YOLO2 framework.The proposed method extracts the feature by separating the front and back frame data of the frame to be detected and then combines the timing information through 3D convolution and finally performs the object detection through the YOLO2 object detection network.Compared with CNN-based image object detection methods,the proposed method,when detecting each frame,incorporates the information of the frames before and after the detection frame and models the timing information of the video.The validity of our approach is verified by comparing with the original YOLO2 object detection method on the vehicle detection task on the public dataset DETRAC.