Pedestrian Detection Under The Vehicle Environment

Driver Assistance Systems has an important role in ensuring traffic safety, pedestrian safety, and reduce traffic accidents. The important part of intelligent vehicle driver assistance research is how to use computer vision technology for pedestrian detection and gives distance warning for driver. From the image based on visual pedestrian detection and auxiliary security application problems is how to find typical characteristics of pedestrian more accurate, faster, and robust and how to use this algorithm in application. In this context, this thesis focuses on monocular vision pedestrian detection algorithm, distance measurement in front of vehicle and safety warning methods. The main work and research results are as follows:1.The edge information is often the key to detect pedestrians, while the divergent and saliency characteristics of human eye are not considered during selecting edge’s gradient direction, which need calculate directions respectively. In this paper, we propose a novel descriptor, called the Saliency Texture Structure Descriptor (STS), which is inspired by Weber Local Descriptor. First, compute the ratio between two terms:one is the relative intensity differences of a center pixel against its neighbors, the other is the intensity of the center pixel as local saliency factor. Second, extract texture though divergent gray level co-occurrence matrix. At last, combine both using two dimensional histograms as the feature vectors. The descriptor is computed simply and has strong properties of describing not only the saliency texture’s distribution but also pixels’divergence. The algorithm is compared with related local descriptors such as CENTRIST and HOG, and state-of-the-art is achieved on pedestrian detection. At the same time, it has high application value for vehicle driving safety.2. In order to measure the distance between pedestrian and vehicle in the road, presents a camera calibration method that does not require specific flags on the ground and fixed calibration object. First, designs a move mode for calibration. Secondly, the end-points of moving target are linearly fitted to obtain two parallel lines. Thirdly, calculates the camera external parameters based on three-line method. Finally, uses affine transformation to calculate the actual distance between pedestrian and vehicle itself. Experimental results show that the proposed camera calibration method operate simply and the error less then2.5%within50meters. So it fully meet the requirements of driving distance measurement accuracy. 3. Combined with the calibration parameters of the camera, the relative positional relationship between pedestrian and vehicle, presents a method to warn the pedestrian. Experimental results show that the algorithm can accurately detect pedestrians in front of the vehicle within25meters, and can warn pedestrian within10meters. The warning distance fully meets the emergency braking distance under the speed of60km/h.