Study On Human Detection In Video

Detection and tracking of the body targets in video is one of the most importantresearch orientation of computer vision. The task of this research is to capture thetargets’ trajectories through body detection and tracking. It has many importantapplications, especially in video surveillance, smart vehicles and human-robotinteraction. We did some research on this field and presented our methods in thispaper.Based on human detection using histograms of oriented gradients, this papercame up with an improved method. Using INRIA data library, we trained our ownLinear-SVM to detect bodies in videos. After computed gradient for each pixel in theOPP color space, we calculated a histogram for every detecting window, which wasinput to SVM for training and detecting. In order to detect all possible persons ofdifferent scales in the image, we resized the original image at a fixed rate and didintensive scanning at every level. Compared with that in RGB color space, ourmethod made better performances, e.g. locating the body more exactly, less falsedetection and so on.The algorithm of detecting single person first and counting then would causepossible deviations when people were crowded in the scene. To count people in crowdbased on independent motion is a good way to solve this problem. We utilized KLT tocapture corners’ trajectories, then counted independent motion through spatialclustering and coherent motion clustering, and estimated the number of people in thescene. Compared with algorithms based on objects’ appearance, our method worksmore effective when independent motion was distinct.In some situations it’s no need to get the exact number of people, but if we onlyneed to judge persons in the scene was crowded or not, we could detect the crowdbased on Hough Transform. The procedures indicated as below: Given imagesequences, we constructed spatio-temporal images with pixels at fixed positions inoriginal images. The moving points in time domain will generate into a line inspatio-temporal images. We used Hough Transform to reflect these lines to the Houghparameter field, and these statistics could help us to judge if crowd existed or not. Theexperiments demonstrated that the method above could get the result effectively andspeedily, which can fully satisfy the needs in real-time application.