A Study On Hierarchical Recognition Of Pedestrians Based On Multiple Features

Pedestrian recognition is the foundation of the analyzing of human behavior action. It has great important significance and practical prospects in the field of intelligent surveillance, intelligent transportation and human-computer interaction. However, because of the variety of different appearance, movements and the complex and changeable environment it is surrounded, the present pedestrian recognition method is still limited and cannot satisfy the demand of accuracy and time. The main characteristics of recognition include two parts:feature extraction and the designing classifier. This paper does some relevant research work about the difficulties in the two aspects of the recognition process.According to the problem that it is difficult to describe the pedestrian contour, for example, the robustness of using single feature is weak, pedestrians can be divided into three parts:head-shoulder, torso and legs, and then extracting the three parts’feature individually. To solve the problem that numerous background interferences exist when extracting features of head and shoulder, the background area above the left and right shoulder should be removed according to the HSCENTRIST (Head Shoulder CENsus TRansform hISTogram) features of head-shoulder shape, and the integral images are adopted to speed up the extracting process.The feature vector dimension of describing pedestrian in traditional object recognition is too high, and the speed of classifier recognition is slow. Thrilling to this problem, the essay comes up with the method of hierarchical feature extraction and hierarchical classifier combination. The method firstly extracts a HSCENTRIST feature, and trains a double-level classifier according to the feature.The first level classifier takes initial recognition by HSCENTRIST feature, removing much non-pedestrian area, so the recognition speed is guaranteed. The second level classifier recognizes more accurately in the candidate area which is ensured by the first level classifier. If the reliability is very high according to the HSCENTRIST feature, then the torso part is extracted to recognize; if it is comparatively reliable, the leg part is extracted. If not, it is defined as non-pedestrian area, a kind of hierarchical recognition process which extracts features step by step is realized. The experiment shows that the method can improve the recognition speed as well as guarantee the accuracy.Trilling to the problem that traditional tracking methods can easily lose target when it encounters with blocks, the essay comes up with the hierarchical recognition method that can used in the video sequences. In this method, the current frame identifies according to the information of the last frame. When the pedestrian is recognized, the hierarchical evaluation method of the object similarity between two frames can be used to judge whether the object is the same or not, and then track the object. Experiments demonstrate that the presented method has strong robustness and can realize the dynamic object tracking under occlusions, what’s more, it can detect the new object that enters into the scene and track it.