Moving Object Detection Based On Superpixel Image Segmentation In Complex Environment

In recent years,with the rapid development of communication,electronics and computer technology,detection and identification technology are becoming familiar to people.The emergence of ADAS and Cameras-Surveillance need higher standards for visual detection.Video can clearly describe objects and excavate the detection target.In particular,movement targets are identified in complex situations,such as occlusion of targets,target deformation and background changes.Traditional object detection algorithms have poor effect in accuracy and robustness,which cannot adapt to complex scene because of its variation and deformation.Therefore,the algorithm of motion target detection based on super-pixel segmentation and learning objective characteristics is proposed in this paper.Firstly,based on the AdaBoost model and the improved SLIC super-pixel segmentation method,we construct the moving target appearance model.Secondly,the maximal posterior estimation is calculated by using particle filter and Bayesian framework.Finally,we use the algorithm to detect targets in video images.The main contribution of this thesis includes the following aspects:1)We calculate the luminance component to constraint the SLIC super-pixel segmentation and to improve the semantic boundary the accuracy of the segmentation.The target search area is segmented by using improved SLIC super-pixel.The algorithm uses a Mean-Shift to cluster the characteristics of the pool and then calculates the confidence diagram of moving target to construct target appearance model.2)The influence of noise change is researched on target detection in complex dynamic background video.Using HOG and LBP features of the training images get weak classifier is proposed by means of AdaBoost algorithm.Then,we construct cascaded reinforced classifier to confirm search area and classify the target.3)Based on Bayesian theory,the particle filter is utilized to calculate the histogram similarity and reconstruction error of the target to construct motion model and observation model.With the suitable reinforced classifier and model occlusion and update strategy,the state of the target can be estimated in the tracking process.The simulation results show that the AdaBoost and Super-pixel(ABSP)algorithms can simplify the complex background information,reduce the dynamic noise effects and effectively detect the moving and achieve high accuracy and robustness in complex environment.The target appearance model can be effectively extracted to improve the influence of target deformation and occlusion in the complex background.