Pedestrian Detection Based On The Implicit Shape Model

The understanding and description of human sensory information is a hot and difficult issue of current artificial intelligence. 80 percent of the information obtained from outside is percept through vision. Although object detection has been studied for more than ten years in computer vision community, it is still an active research area. At present, there is not an object detection method which is general, robust, accurate, efficient and real-time. Pedestrian detection is an important branch of object detection. Nowadays, it's intensively investigated and becoming a hot topic in the field of computer vision. It could be widely used in smart surveillance, driver assistant systems, advanced human-machine interfaces and so on.The main work of the paper include:1. Object detection: in order to solve the problem of the slow learning rate of traditional Mixture Gaussian model, we propose a moving object detection algorithm with a novel background updating method. Firstly, the mixture Gaussian model was constructed, and a new way is adopted to update background, which utilizes different equations at different phases. Then, a coarse foreground image could be obtained by background subtraction.2. Shadow removal: When detecting the moving object, moving shadow projected by the object may be detected as the foreground, which would cause the combination, geometric deformation of the target, or even the loss of the target. Moving shadow could be eliminated by the shadow detection algorithm based on light, color and gradient information. Finally, we perform post processing. The experimental results show that this method could remove shadow well and extract moving objects accurately, wherever in indoor or outdoor environment.3. The construction of the Implicit Shape Model( ISM). Difference-of-Gaussian (DoG) detector, simple Greyvalue Patches descriptor and RNN clusterig method are chosen to construct the ISM. 4. Pedestrian detection with the ISM: at the detection stage, we apply an interest point detector and extract features around the selected locations. The extracted features are then matched with the codebook to activate codebook entries. From the set of those matches, we collect consistent configurations by performing a Generalized Hough Transform. Each activated entry casts votes for possible positions of the object center according to the learned spatial distribution. Consistent hypotheses are then searched as local maxima in the voting space. Finally, pedestrians are detected by top-down segmentation. This method can detect pedestrians from static images, and solve the partial occlusion problem.