Counting Heads From Bus Video Streams Under Uncontrolled Conditions

This work mainly considers the problem of human head detection and count based on themonocular camera. This problem becomes challenge because the appearance of the object tends tochange intensely due to the impact of external factors such as posture, illumination, occlusion andpseudo-target. And in order to meet the practical demands, head detection and counting should be fastand accurate enough, and has a low false detection rate and undetected rate. In this work, we try toseek methods to deal with these complex changes, and we get the following idea: Considering thehead object’s appearance and texture information, we choose the CENTRIST combined with LBPfeatures as head descriptor. And we use a linear support vector machine to train the classificationmodel respectively. We tandem the models into a simple cascade classifier and use it for headdetection. Then we track and analysis the objects detected and establish their trajectories. At last, wecluster the trajectories to get the automatic head count. Our main works are given below:1. For the real-time requirements of the video-based head count, we use the easily computedCENTRIST features as the head descriptor. But the CENTRIST features mainly focus on the overallprofile of target, while ignoring the texture information. So we proposed a CENTRIST-LBP descriptorwhich combined the CENTRIST with LBP features.2. For the video stream environment is complex and changeable, there may be multiple headsappeared in the video screen at the same time, even pseudo-objective, so we proposed moving targettracking algorithm based on a rectangular area, which will output the head motion trajectory of theobject in each frame. Then, we used a DFT coefficient to represent trajectory feature rather than asimple feature point-based representation, because the parameters DFT coefficients of the trajectorycharacteristic form of representation can be expressed more fully and closely track, while the clusteranalysis is also very effective. Finally, this paper adopt DPMM-based bayesian maximum a posterioriprobability estimation method on target motion trajectory (head) to cluster analysis, the final numberaccording to the result of clustering analysis in statistics.