Study On Methods Of Pedestrian Detection On Static Images

Computer vision is a discipline to study the simulation of biological vision by video cameras and computers. Object detection is one of the most basic tasks of computer vision. The detector extracts certain features to detect the specific category objects. It provides the necessary technical foundation for high-level tasks such as object recognition, behavior analysis, retrieval, etc. As a specific category object detection, there are many research difficulties in pedestrian detection which are non-rigid human body, pedestrian appearance diversity, complex background, illumination changes, scale changes and occlusions. Pedestrian detection has been applied widely in intelligent video surveillance, driver assistance pedestrian protection system, intelligent traffic control and many other fields and it has great commercial value. These factors caused that the pedestrian detection studies desorbed from object detection studies gradually. And it became an independent research field.We study the pedestrian detection methods on static images in this paper, and use INRIA person dataset as the experimental platform. The research work are as follows:Firstly, we studied simplified non-maximum suppression which is the typical representative of post processing methods. However, the traditional simplified non-maximum suppression only uses one constraint and cannot discard multiple nearby detections well. An improved simplified non-maximum suppression with two additional constraints was proposed. With respect to the traditional simplified non-maximum suppression only calculates the proportion of intersection area to the area of candidate detection bounding box, the two additional constraints named "completely covered detection suppression" and "PASCAL VOC overlap criterion" calculate the proportions of intersection area to the area of selected detection bounding box and to the union area respectively.Secondly, we supplemented the detection performance of the standard HOG human detection model under FPPI vs. miss rate evaluation method, and discussed the context of standard model. A view that 32×96 model which the context was removed is better than the standard model was put forward. At the same time, we proposed one novel feature selection method combined Fisher-like ratio to calculate blocks’ discrimination performance with NMS to select feature subset.24 blocks were selected from 144 designed blocks and composed a 1854-dimensional human detection model.Finally, we studied the attention mechanism of the human visual system and the computational models of visual attention mechanism in computer vision. A pedestrian detection method under visual attention mechanism was proposed. Firstly, we used the saliency detection to simulate the attention mechanism of the human visual system. Then we did the gray-level saliency map binaryzation and extracted the candidate detection area to simulate ROI under the visual attention mechanism. Finally we used the detector to detect pedestrians in the candidate area.Our research work include features extraction, classifier training, post processing in traditional pedestrian detection methods, and also try to build a pedestrian detection method under visual attention mechanism. The experimental results indicate that the false positives reduced a lot after our improved simplified non-maximum suppression adopted. It tapped the potential of the detection method further, and can be extended to all the object detection methods based on sliding window method. Our proposed feature selection method could obtain a sub-optimal and semantics reserved feature subset at a reasonable cost. The improved human detection model using feature selection achieves significant performance improvements. The pedestrian detection method under visual attention mechanism significantly reduced the search range of traditional pedestrian detection methods. However, subject to the imperfect visual attention models, the false negatives increased reluctantly while the false positives reduced.