Pedestrian Detection Based On Visual Attention Model

Pedestrian detection under video surveillance system has always been a hot issue in computer vision research, which is widely used in train station, airport, large commercial plaza and other public places.In recent years, with the development of visual attention mechanism, the visual attention mechanism is paid more and more attention in object detection and tracking, and makes the substantial progress and breakthrough. Therefore, it is of great significance and application value to apply visual attention mechanism to pedestrian detection.In this paper, the pedestrian detection method based on the visual attention mechanism is mainly studied. The main work includes three aspects as follows.Static visual attention modelin the spatial domain is constructed by combining bottom-up and top-down attention guidance. According to the characteristics of pedestrian, Itti's bottom-up visual attention model is firstly improved via intensifying the orientation vectorsof elementary visual features so that the visual saliency map will be suitable for detecting pedestrian. With the attributes of pedestrian, the regional model and Gaussian model are adopted to construct skin color model. For the requirement of pedestrian detection, visual attention guidance based on skin features is then put forward to complete the top-down process. Finally, bottom-up and top-down visual attentions are linear combined via the proper weights obtained from amount experiments so as to construct the static visual attention model in spatial domain.The experimental analysis is carried out on MIT, CAT2000, and EyeCrowd data set. The experimental results have verified the effectiveness of the proposed attention model.The spatial-temporal visual attention model is constructed via motion features in temporal domain. On the foundation of static visual attention model in spatial domain, frame difference method and optical flowing are combined to detect motion vectors. The filtering is applied to process the field of motion vectors. Moreover, the saliency of motion vectors can be evaluated by motion entropy so that the motion feature selected is better suitable for spatial-temporal visual attention model. i LIDS of AVSS 2007 conference, the PETS2006 data set and practical videos are chosen for experiments. The experimental results have verified that the adaptation of the proposed attention model for detecting pedestrian.A pedestriandetection and tracking is proposed based on visual saliency. The video saliency maps are firstly obtained by the proposed spatial-temporal visual attention model. The Histogram of Gradient(HOG) and SVM classifier are then combined to detect pedestrian in the salient regions. Finally, mean shift algorithm is applied for tracking pedestrian under the guidance of the salient map. The experimental results have shown that the proposed algorithm can improve the performance of pedestrian detection and tracking.