Moving Object Detection And Segmentation Based On Background Modeling

Moving object detection and segmentation is challenging in the field of computer vision and it is also an important part for intelligent monitoring system. The result of which directly influences the effect of target classification, tracking identification and behavior analysis. In this thesis, moving object detection and segmentation technology is studied by regarding the requirement of intelligent monitoring as starting point and the video sequences shot by fixed camera as study object.In this thesis, we concentrate on the moving target detection algorithms and compare different methods such as Gaussian mixture model, Codebook and Vi Be using videos with different scenarios. The research is mainly focused on the advantages and disadvantages of Vi Be, and also its ability to adapt in different scenarios. Experimental results show that Vi Be performs well in simple scenarios with less background interferences. It not only has a high detecting speed, but also be able to detect moving targets from foreground image completely. However, Vi Be has a weak adaptability in complex scenarios. This algorithm is not robust enough under the condition of background disturbance or illumination changes, so its precision needs to be improved.The detection result of Vi Be is not ideal under complex scenarios, in this case, a moving target detecting algorithm combining Vi Be with spatial texture information is proposed. To describe background features accurately, the CSLBP texture descriptor is improved. Adaptive threshold is imported, and thresholding on the absolute difference is applied to generate binary representations using both intra-region and inter-region operations. After that, a background model based on color and texture feature could be obtained by adding spatial features in Vi Be. Experimental results show that the proposed method makes up the deficiency of the Vi Be in feature representation,and improve algorithm's adaptability in complex scenarios with shadow, illumination changing and background interference. As a result, the precision of detection can be improved.After moving target detection, it is need to segment moving objects from foreground image by using segmentation algorithm with threshold. In this thesis, we research on the problem of threshold selection. For the poor adaptability of fixed threshold in different video scenes, an adaptive threshold algorithm for moving target segmentation is proposed. The method is mainly divided into three steps: initial threshold selection, foreground areas classification and threshold updating. In initial time period, a strategy based on the combination of background modeling and Grabcut is presented to extract foreground objects and set an initial threshold. On the base of this, we can choose some foreground as samples and classify them by applying K-means clustering method. Finally, an appropriate threshold could be selected for moving object segmentation according to the classification result. Experimental results show that the proposed method can overcome the disadvantages of fixed threshold. It has strong adaptability to various scenes and can effectively segment moving targets from foreground image.The algorithms proposed in this thesis have certain theoretical value to moving target detection and segmentation in the field of intelligent monitoring. However, their speed is limited by the complexity of the algorithms themselves, and it is difficult to meet the requirement of real-time in practical application. How to improve the efficiency of detection and segmentation algorithms is a problem need to be solved in the future work.