| Summary: | 碩士 === 國立高雄第一科技大學 === 電腦與通訊工程研究所 === 102 === Pedestrian detection and counting is an important topic in developing an intelligent surveillance system. In this work, we propose a vision-based system for detecting pedestrians in an image. Be robust to crowded scenes and adapt to incomplete foreground from background subtraction algorithm, expectation maximization (EM) algorithm is applied to impose the constraint of body part for achieving successful detection. First, the corner points at body part are all detected and described using histogram of oriented gradients (HOGs). In addition, one of three body part labels (head, torso, and leg), a kind of locality property, is encoded in corner points for overcoming the mutual occlusion situation. Then, we apply a grouping algorithm in HOGs feature space to form a set of clusters. Each cluster center is considered as a code word and the probabilities of this cluster belonging to head, torso, or leg are also computed, respectively. During detecting phase, all detected corner points are matched to the construct code words and are assigned to three body part probabilities. After that, an EM algorithm is applied to iteratively estimate the likelihood probability of all corner points to the pedestrian candidates (E-Step) and update the parameters of the pedestrian models (M-Step). In the experiment, three videos are used to validate the performance of the proposed method.
|
|---|
