A Study On Key Techniques Of Multi-Vision-Feature Based Rear-Vehicle Detection

Aiming at improving safety, high efficiency and comfort of road transportation system, the Intelligent Transportation System (ITS) takes full advantage of available road elementary facilities to solve problems such as traffic jam, frequent occurrence of traffic accidents, and environment pollution, etc., by adopting multiple technologies including sensor technology, image processing, pattern recognition, and artificial intelligence. In ITS, Driver Assistance System (DAS) is an important part that is urgent needed with wide scale applications. Rear-vehicle real time detection technology is the key function of DAS, which plays considerable role for parking assistance, changing-lane assistance, avoiding collision accidents etc. As a result, Rear-vehicle real time detection technology has become a hot issue for research in ITS. In particular, with the development of computer vision technology , the vision-feature based rear-vehicle detection has attracted much more attentions domestically and abroad for its easy installation and maintenance, real time, and easy of promotion..However, most current approaches of vision-feature based rear-vehicle detection are prone to being affected by factors of circumstance such as complex light, weather, and background, hence, the robustness of these approaches are not good. For that reason, to enhance the robustness of rear-vehicle detection, it is necessary to synthetically leverage all kinds of features and methods based on the analysis of vehicle and background vision features. Contemporarily, most current approaches of rear-vehicle detection are in still in experimental research phase, that is to say, the practicability of these approaches are not very satisfactory. For this reason, to answer the requirements of industry application, it is necessary to strike a balance for recognition precision and real time of detection approaches.Given on the problem of bad robustness and real time faced by most existing approaches , driven by putting the rear-vehicle detection into industry application, this dissertation proposes a set of rear-vehicle detection techniques based on interfusion of muti-vision-features, and apply these proposed techniques into the implementation of vehicle recognition engine of Advanced Automotive Technology Research Center, which exhibits good performance. The main research works implemented by the dissertation include:●To ensure the accuracy of segmentation of region of interest in the daytime, the dissertation proposes an improved approach for accurate segmentation of region of interest based on the shadow features underneath vehicle. Not only, the dissertation presents the process of segmentation, but also establishes the model for calculating the threshold of grey of the shadow as well as the rules for clustering the candidate shadow and determining the road casting shadow. Based on these rules and the threshold, this dissertation comes up with the algorithm for calculating the threshold of the shadow based on the road edge features, the algorithm for clustering the candidate shadows based on the 3D geometrical; and the algorithm for dealing with the road casting shadows based on color features. In brief, the proposed approach can help avoid the inaccuracy of vehicle detection caused by the road casting shadow.●To ensure the accuracy of vehicle locating in the daytime, the dissertation proposes a vehicle detection technology based on interfusion of multi-vision features. Not only the dissertation presents the process of vehicle locatiing in the daytime, but also comes up with approaches for vehical symmetry axis locating, the vehicle left and right edges locating, and vehicle upper edge locating, by making good use of a variety of vision features, such as the binary symmetrical profile, symmetrical grey, S symmetric in HSV, vertical edge, the left and right border of the shadow, the color of the vehicle body, the vertical and horizontal feature of the background, and the length feature. In brief, he proposed technique can avoid the inaccuracy of vehicle detection caused by some factors in the background.●To ensure the accuracy of vehicle locating at night, the dissertation proposes an improved vehicle detection technology based on vehicle light feature in the night time. Not only the dissertation presents the process of vehicle locating at night, but only establishes the rules for recognizing the single light, deciding the collocability of a pair of lights and the headlights of the vehicle. Based on these rules, approaches for recognizing the single light and localizing the pair of highlights are proposed. The proposed technique can avoid the inaccuracy of recognition of single light caused by the reflective light sources, and ensure the accuracy of the collocability of the pair of the light as well as the vehicle locating. ●To ensure the accuracy of verification of the existence of vehicle in the vehicle detection region, the dissertation proposes a vehicle existence verification mechanism based on knowledge and feature statistics.. This mechanism filters out the background with approaches based on knowledge at first, and then, does the verification for vehicle exhistence on the output from the previous step using the statistics based approach. For implementing the mechanism, this dissertation presents the approaches for feature extraction based on Gabor, classifier training based on radial Gaussian kernel function and background filtering based on multi-vision-knowledge. In brief, this mechanism can avoid the inaccuracy of verification caused by using a single approach.Aiming at the application of the proposed technology, approaches and mechanism, on the basis of the analysis of the architecture of vehicle recognition engine of Advanced Automotive Electronic Research Center (AAC), the dissertation presents the process of multi-feature based rear-vehicle detection in the vehicle recognition engine. For evaluating the performance of the vehicle detection, the dissertation designs and establishes the experiment approaches including the method of image acquisition, the evaluation criterion, and the steps of the experiment. The experiments and applications show a better robustness and practicability of the proposed rear-vehicle detection in the dissertation.