Research And Implementation Of Algorithm For Forward Vehicle Detection And Location

The increment of cars leads to the fact that traffic accidents occur more frequently,therefore automobile auxiliary driving technique has been becoming more important. Compared with microwave, radar, infrared technology, ordinary video technology have high imaging resolution, strong anti-interference ability, low price, low cost, high penetration rate of the comprehensive advantages, without radiation hazards on the human body and so on.It has various aspects and it is irreplaceable. Now, it has beening widely used in Video Surveillance, Human – Machine Interaction, Visual Navigation of Robots and other civil as well as military fields.Moving target detection technology has been a research hotspot for many years.But the change of the dynamic background, such as changes in weather effects, and light moving shadow, disrupting and camera motion,lead to moving object detection becomes quite difficult in practical application.The vehicle detection technology in this paper is based on the video moving target detection technology and it is a spacific application for car driver assistance.Relying on the Department of scientific research project, we performed a research work on video vehicle detection and location. The contents and results are shown as follows:1.Multiscale detection on the frame directly is really inefficient and prone to false alarm. Two algorithm introduced by this paper can extract vehicle potential region.Extracting the potential vehicles regions reduces the number of candidate detection positions significantily, improves the validity of detection and the detection accuracy.2.As the huge chagnge of background,a large number of positive samples are introduced with different models, different point of view and different contrast.The same as a large number of negative samples in different scales with different illumination, different scenes and different distractors.This improves the robustness of the trained classifier to different scene, ensure the detection rate.3.According to the difference of the visual appearance of the different vehicle, a algorithm,which we could call Multi-svm here, is designed to train several support vector machine classifiers by the way of clustering, comparison. Non-supervision mode is adoped here,as well as the num of classfier is not fixed in advance to avoidunreasonable division.4.The kalman filtering method is adopted for vehicle target tracking, and the simulation experiment is carried out.The above algorithms have been programmed to verify the validity and enforceability. Results show that algorithms which this paper studies can detect and locate vehicles quickly and accurately.