Pedestrian Detection In Dynamic Scenes Based On Approximate IKSVM Classifier

Nowadays, as the car ownership and the urban traffic increase, trafficaccidents occur frequently and the traffic safety has become an important so-cial problem. As a matter of fact, pedestrians are the main victims in a trafficaccident. Therefore, to ensure the safety of pedestrians, intelligent auxiliarydriving systems have attracted much attention of researchers in the world. Asan important part of the system, the pedestrian detection aims at detectingpedestrians appearing in front of the vehicle by using vehicle-mounted cam-eras, so as to assess the danger and assist the driver in taking countermeas-ures. Due to the different postures of pedestrians, rapidly changing back-grounds and real-time requirements, the pedestrian detection in complex dy-namic scene is a challenging task.To solve these problems, we present a novel pedestrian detection method,using a road detection technique together with the appropriate intersectionkernel SVM (IKSVM) classifier. The method includes three steps. First, theroad surface is detected by a certain method. Second, the regions of inter-est(ROIs) are extracted through the pinhole imaging principle and the scalesize of ROIs is normalized based on the depth information. Finally, the ap-proximate IKSVM classifier is used for the pedestrian recognition.The advantages of our method are:1) The scale normalization of ROIs has greatly reduced the number of search windows and the false positives,which decreases the computational complexity;2) Though the approximateIKSVM is a nonlinear kernel SVM, its computational complexity is the sameto the linear SVM. Using the IKSVM classifier instead of the linear SVMclassifier can improve the detection accuracy, and its computational complex-ity remains unchanged. Experimental results show that our method has im-proved the detection accuracy and the detection time does not increase.