Forward Collision Warning Technology Based On Monocular Vision

As the increasing demand in vehicle market, industries of car manufacturing and retail has gained great development in recent years. Car parc are growing continuously in big cities, allowing for more and more people to enjoy the convenience brought by individual cars. However, this also results in the increasing occurrence of traffic accidents, posing a threat to people's life and property safety. Consequently, governments, organizations and car manufacturers has put great efforts in improving road safety from different aspect, looking for possible solutions in policies and technologies. Among them, advanced driving assistance systems(ADAS) has drawn much attention, not only because of its satisfying performance, but also because it meets the future trends of road traffic, namely the autonomous driving.This paper research the most important part of ADAS, namely the forward collision warning system(FCW). This system can detection all cars in front using a monocular camera, and measuring the distance to them, thus providing warning message when the distance is too near, and avoiding the happening of traffic accidents effectively. The system consists of two subsystems: vehicle detection and vehicle tracking.In vehicle detection step, the horizontal and vertical edges as well as the shadow underneath the vehicle are combine as clues to determine the vehicles' positions roughly. Sub-images output from this step often contain non-vehicles, because the clues used are not unique to vehicles, so the step is called hypotheses generation. In order to pick out real vehicles, the histogram of oriented gradient feature of features is extracted, and classified by AdaBoost classifier to choose the class labels of images, thus eliminating all non-vehicles. This step is call hypotheses verification.In the vehicle tracking step, Harris feature points are extracted, and matched with each other between two frames using normalized cross correlation to track the vehicles in the first stage. The feature points are put in to P-N dictionary and new feature points are measured by Mahalanobis distance between points in the dictionary to attain stable tracking. Finally, the relation between the number of pixels which is contained between the image bottom line and the vehicle's bottom line and vehicles' distances is determined so that the distance can be measured, and the purpose of forward collision warning can be achieved.In the end, experiments about the detection rate, classification correctness in hypotheses generation step, as well as the tracking successful rate and the accuracy of distance measurement is carried out to evaluate the overall performance of the whole system. The results have shown good performance, indicating the system is promising to apply to various ADAS.