Study And Realization Of Pedestrian Detection Algorithm Based On Multi-Features Fusion

Pedestrian detection is a hot topic in computer vision, which has a broad application prospect in robotics, unmanned, virtual reality technology and some of the military fields. At the same time, China is a new rising vehicle-power, where many traffic accidents happen every year, and causing heavy casualties and property losses. Pedestrians are the most vulnerable group in the traffic system, so to detect them is of great significance in the realization of intelligent transportation.Under reality conditions, due to some factors such as the non-rigidity, diversity of huamans and the complexity of the background, pedestrians detection research is relatively difficult. Based on the the previous algorithms, this paper puts forward the pedestrian detection algorithm based on multi-features fusion. First of all, algorithms based on color segmentation and edge symmetry segmentation can be used to exact Regions of Interest(ROI). Color segmentation can quickly idetify the figure as humans's skin area, but it will not be reliable if there's no pedestrian bare skin or the background color is too close to the color of the humans'skin. Then segmentation algorithm based on edge symmetry can be compensated. And then to detect pedestrians on the ROI, this paper adopts a method based on sliding window with HOG-PCA+CENTRIST overall pedestrian detection.Haar+integal image head detection. HOG is good however its dimension is high and is sensitive to bending, so our algorithm adopts PCA to reduce its dimensions and CENTRIST to descript the overall outline of the target better. In the sliding windows with detected humans, detect human's head with Haar, and integral image is used to accelerate computation. In order to detect the pedestrians in different sizes, this paper has trained three types of classifiers in four scales based on SVM and AdaBoost. Though increasing the workload in classifiers-training phase, it simplifies the calculation of detection.Finally, fuse all the detected windows at all Levels by proposed window-fusion algorithm.The experimental results based on INRIA database show that the proposed algorithm which is based on multi-features fusion can achieve 91.5% detection rate when FPPI is 1, close to the best algorithm ever been published.DET figure indicates that proposed algorithm is significantly better than other traditional algorithms, like HOG et. al, and performs better than most of algorithms. This paper achieves good results even in the PASCAL challenge.