Research On Vision-based Pedestrian Detection Algorithm In DAS

Along with the improvement of people’s living standards, motor vehicles areincreasing, and traffic accidents happen frequently. Since entering the new century,traffic accident has become a serious social problem. As pedestrians are importantparticipants in traffic system, and are also the main victims in traffic accidents, theimportance of protecting pedestrians is so obvious. Pedestrian detection of this paperrefers to using a vehicle-mounted camera to capture on-board video and judgingwhether pedestrians exist or not in the video. If any, the pedestrians’ positions aremarked, then danger is estimated and some measures are taken to protect thepedestrians.Although pedestrian detection has been studied for decades, there is still no onealgorithm that is universal, accurate and robust. Causes include the diversity ofpedestrian posture and dress, the complexity of light and background change, and theinteraction between people and environment. Therefore, pedestrian detection isalways a research emphasis and difficulty of machine vision.This paper focuses on monocular vision based pedestrian detection in DriverAssistance System (DAS). Firstly, take account of the specificity of vehicle-mountedcamera, a simple ROI segmentation is implemented on the input images, whichreduces detection region and improves real-time performance. Secondly,“Haarfeature+Adaboost classifier” is used to extract the pedestrian candidates, this stepruns very fast but results in some false positives, which can be filtered by verticaledge symmetry verification. Thirdly, HOG feature and SVM classifier are adopted toverify all the candidates. At last, Kalman filter and template matching are used forpedestrian tracking.Multi features and classifiers are utilized in this paper, which takes the advantageand avoids the disadvantage, taking both accuracy and detection speed intoconsideration. Experiment results show that this method can detect pedestrians withdifferent postures, even the small pedestrian objects. Besides, real-time requirement issatisfied. All these have certain reference significance for pedestrian detection in DAS.