Studies And Applications On Moving Object Detection Algorithm In Complex Scenes

Video surveillance has been widely used.Nanjing alone has already held over 500 thousand surveillance cameras.The need of analyzing and recognizing computer images is prominent,especially in aspects of extracting moving object's size,trajectory,speed,and flow statistics.As a result,moving object detection in video sequences is a fundamental step of extracting information in many visual surveillance applications.However,the environment is always complex and out of control,which obstructes accurate moving object detection.The detection result often has a lot of false positives and negatives.Our lab has carried on several research projects at national or ministerial level such as"Research and Application of Key Technologies in Real-Time Video Monitoring and Recognizing System for Ships"," Security Incident Detection System Based on Intelligent Visual Perception "and "Multi-sensor,Convergence Mining,Warning Decision,Integrated Delivery Network System for Express Ways ".However,scenes cannot be assumed to be single due to camera jitter,light changes,bad weather,shadow or other unpredictable and complex events,which make moving object detection a challenging task.In the case of moving ship detection,the difficulty is extremely big owing to the shimmering water,the weak color contrast between ship and water,and the ship ripple's sailing wave.Related research at home and abroad is far less enough.So we do need to do more research on the complex environment conditions and on the negative impact they bring to the moving object detection.Based on the research bottleneck encountered in practical application scene,composition of video monitoring modules and the significance of moving objects detection as well as its relevance to the complex environments are described.To find out the key points which cause error detection and failed detection,the rationality and limitations of existing related algorithms at home and abroad are analyzed.Difficulties and encountered problems in key algorithms are figured out.Then corresponding algorithms or optimized solutions are proposed and applicated in practice.Algorithm research mainly includes pavement ROI area extraction based on vanishing point estimation,moving object detection based on visual background extractor(VIBE)in dynamic background scenes,moving object detection based on motion information distribution in camera jitter scenes and shadow detection and eliminate based on color ratio and texture consistency.Since traffic monitoring PTZ cameras is remote-controlled,a new pavement ROI area extraction method based on vanishing point estimation is proposed,which uses lane marks as reference information.Unlike usual ways which need manual settings,this method is self-adaptive to PTZ motion to automatically locate camera position and automatically extract ROI area.Besides,the watershed algorithm is introduced to segment the left and right lanes,which prepares for the after traffic flow count and anomalous event detection.Aimed at moving object detection in dynamic background scenes,the mainstream moving object detection algorithms in recent years are analyzed,and the visual background extraction(VIBE)is improved.The background model is initialized with several continuous frames,which can remove the bad influence on precision caused by the ghost that occurs when using single frame initialization.Decision threshold is self-adaptive to background dynamics so that background models can better adapt to dynamic background.To obtain real time background model and improve the robustness,selective update mechanism is applied based on space consistency and fuzzy criterion,which narrows down the bad influence on model caused by error detection.In order to solve the problem that unstable camera may deteriorate the performance of detection,a moving object detection algorithm based on motion information distribution is proposed.After analyzing the motion information of motion pixels in camera jitter scenes,it is can be found that the difference of motion information between foreground and background can provide effective basis for distinguishing foreground and background area.Based on this,the motion information is used to build a nonparametric background motion information distribution model,and the probability of likelihood is introduced to remove false moving object detection caused by camera jitter.And the adaptive decision threshold is automatically calculated based on mean-shift and information entropy method,which can overcome the defect that using a single global threshold is not capable to adapt to the change of environment.Considering that the influence of moving shadow on detection accuracy,the imaging model and color attribute of shadow are analyzed and a color ratio feature with illumination invariance is built.However,those color features with single type are insufficient.Combining the texture consistency of shadow with that of background,a new shadow detection method using both color ratio and texture is proposed.Differing from tradition method that calculates texture consistency of the whole region,a new index is designed to quantize texture similarity of small region.The proposed method combines with the both type of features,remove the unreliability of single feature,and is more rational.The paper does not stay in theory.From experiments to practical application systems such as "Nantong Maritime,"and"Shanghai-Nanjing Expressway",the design and construction of the Real-time Video Surveillance System for Ships is mainly introduced,which is used and verified in real world.To sum up,the research has both important theoretical significance and application value.The innovations are as follows:? A pavement ROI extracting method based on vanishing point estimate is proposed,which improves the automation and humanization compared with manual method.In addition,computation cost and processing difficulties generated by redundant pixel information are reduced.? A moving object detection method based on visual background extractor is optimized.Dynamic background's interference to the motion foreground detection is effectively reduced,which provides a good foundation for after image analysis.? Dynamic information distribution is adopted in moving object detection to reduce the interference of camera jitter.? Shadow detection method based on color ratio and texture consistency is proposed,which makes up for the method using single feature and removes the interference of motion shadows to the accuracy of surveillance system.