Pedestrian Detection And Counting Based On Single Camera

How to detect and count pedestrians quickly and accurately in video is an importantresearch field in computer vision and image processing. This paper dedicated to detect andcount pedestrians based on a single camera. The main research contributions are listed asfollow:Firstly, in order to handle partial occlusions in videos shot by single camera, a pedestriandetection method based on motion information and part detectors is proposed in this paper.Motion areas are extracted in the video image by fast convergence Gaussian mixture modelas the candidate region of the pedestrian. In each candidate region the part detectors whichhave trained by the liner SVM combined with the HOG were used to scanning detectindividually. The Max Margin Hough Transform is used to verify the detection result.Experiments show that our method has high performance in detecting pedestrians with partialocclusion.Secondly, for the traditional pedestrian detection algorithm based on histogram oforiented gradient (HOG) features cannot achieve the accurate positioning of the pedestrianwhen the pedestrians are partially obstructed in the actual scene. A pedestrian detectionalgorithm based on HOG feature pyramid and mixed star-structure model combined withweak label structure SVM(WL-SSVM) is proposed. The HOG feature pyramid ofstar-structure model include overall models and part models of human are extracted fromtrain samples. And then multiple star-structure target models will be trained by the classifierweak label structure SVM. The model is not only able to obtain coarse resolution edges of thetarget, but also to capture details of human. So the robustness of the algorithm can beimproved in some extent when the environments are complex. Last, the mixed star-structuremodels cascading detection is used to get the accurate positioning of the pedestrian.Experimental results show that the proposed algorithm is not only able to detect partialoccluded pedestrians accurately but also enhance the detection performance under noocclusion.Finally, considering on the lack of two mainstream statistical methods of peoplecounting-based on feature-regression, and this paper presents an approach to combine thetwo methods legitimately. For away from the lens region of the video frame, using thebackground segmentation method to extract the foreground blocks firstly, the features of theforeground blocks are extracted to combine Bayesian multiple kernel support vectorregression to estimate the number of people. For close shot area, the accurate positioning ofpedestrians and statistics of the number of human are achieved by cascade pedestriandetection. Experimental results show that this method is not only able to achieve moreaccurate statistics of people, to some extent, reduce the statistical time, but also can determine the locations of pedestrians in some range accurately.