Research On Pedestrian Detection Based On Several Features Fusion

Pedestrian detection is a hot research direction which has a wide application prospect in many areas such as safety driving, video surveillance, robot applications and so on. The content of this paper is the pedestrian detection algorithms of static high-resolution images under complex background, which is a detection of pedestrians who usually remain upright. The research direction is the basic research subject which is rather applicable and universal. Since there are some challenges for the pedestrian detection in the images such as partial occlusion, illumination changes,differences of size, color, posture, location and so on, thus the performance of the pedestrian detection needs to be improved.Pedestrian detection algorithm develops from the shape template matching to the feature extraction based on statistical learning, from single feature and single sensor to multi-feature fusion and multi-sensor combination. However, feature extraction based on the statistical learning is the mainstream of the pedestrian detection at present. This paper do researches on commonly used features like HOG feature, Haar feature and so on, so as to improve the HOG feature effectively which is the mainstream of those features. Aiming at the partial features of pedestrians, this paper do researches on common features such as skin color, hair color and curvature,and establishes feature models.Firstly, based on the traditional HOG algorithm structure, this paper proposes a HOG algorithm, improves the HOG feature, trains the entire template by utilizing the SVM classification algorithm, and builds image pyramid to detect complex images by sliding window scanning. Secondly, it makes further efforts to add the human body partial templates and use the template elastic model to propose the partial and entire template elastic model algorithm. Realizing the elastic combination of human body and parts by setting the offset value, it effectively improves the robustness of the detection for the pose-varied pedestrian and partial occlusion. Meanwhile, it can not only raise recognition rate, but also increase the amount of calculation. Finally this paper proposes the pedestrian detection algorithm based on part quadratic weighting and several features fusion. It adds HOG detection for the first time to eliminate a lot of interference region, fuses the feature of skin color, hair color and curvature,overcomes the limitations of the single feature detection, detects the partial feature inarea of partial template effectively. Therefore, features fusion has high efficiency, and the algorithm complexity of pedestrian detection is better controlled.The experimental results show that the partial and entire template elastic model algorithm has greatly raised the recognition rate compared with the HOG algorithm,but the amount of calculation has also increased significantly. However,the pedestrian detection algorithm based on part quadratic weighting and several features fusion makes up for the inadequacy of single feature. Meanwhile, it detects the whole and part key features of pedestrian more effectively, raises the recognition rate, and effectively controls the algorithm complexity.