Research On Fast Detection Algorithm Of Forward Vehicle Based On On-road Camera

With the rapid development of computer vision, vehicle detection based on mo-nocular camera is widely used in Advanced Driving Auxiliary System (ADAS). Through real-time monitoring of the front road, ADAS helps driver in braking and changing the lane when facing a danger situation, aiming at avoiding or reducing the damage of the rear-end accident. On one hand, vehicle detection faces big challenge in terms of detection accuracy due to the various changes of the vehicle itself and the external environment. On the other hand, vehicle detection in ADAS requires a very strict real-time speed to handle the danger in time. In this thesis, several aspects are studied for forward vehicle detection in ADAS including:(1) The representation of the rear-end of the vehicle based on LBP and Co-HOG features along with the Boosting algorithm and cascaded classifiers for fast and robust vehicle detection.(2) The divide-and-conquer strategy for different types and views of vehicles. Taking different types and views of vehicles as a whole and training one single clas-sifier result in poor detection performance. So divide-and-conquer strategy is adopted in our algorithm. A unique vehicle division scheme according to different types of vehicle and different views of vehicle is designed. Then for each divided type, a classifier is learned respectively. Finally, the classifiers are applied to specially de-signed image regions to ensure a high computation efficiency. Experimental results demonstrate that the proposed strategy can largely improve the detection accuracy.(3) Region-Of-Interest (ROI) extraction. In forward vehicle detection for ADAS, to run classifiers on each of the image positions is not necessary. It results in higher computation cost and unexpected false positives. The vehicle detection onlyneeds to deal with the image area (ROI) that a vehicle might appear and has a risk to cause a rear-end collision. We propose a method to extract the ROIs based on a calibrated camera and the prior knowledge about the road and vehicle. By running the classifier only on the extracted ROI areas, the computational cost is reduced largely. Besides, the unexpected false alarms falling outside the ROIs are removed.Extended experimental results show that the proposed vehicle detection method can achieve a high detection accuracy and a real-time processing speed as well.