Study On Road And Vehicle Detection Based On Machine Vision

With the development of transportation system and the progress of science and technology, Intelligent Transportation Systems(ITS) has received widespread attention and develops rapidly. As an important part of ITS and the future directions of vehicle, Intelligent vehicle(IV) which integrates the functions of environmental perception, route planning, and safe driving, is developing rapid.Environmental perception system based on machine vision is a very important part of the intelligent vehicle system. This dissertation has researches on environment perception system of intelligent vehicle, taking the BJUT-IV, which is an electric smart car of Beijing University of Technology as a platform, and based on monocular vision. The researches have main contents of following four aspects: image pre-processing, lane detection, vehicle identification, coordinate transformation and conditions warning.Firstly, the original image taking by IV camera is preprocessed according to characteristics of road. Aiming at road environment, image is grayed using weighted average method. And the noise is reduced using bilateral filtering, which could save the edge information when denoising. Meanwhile the different lighting conditions are classified and using different gray stretch methods according to those conditions.Secondly, a three-lane detection method is proposed, aiming at multi-lane road environment. A lane marking segmentation method based on kernel varying Top Hat is proposed, reducing the affection when environment changes. A lane marking locating method is proposed using three-lane straight line model, which firstly searching straight lines using Hough transforms based on density function of double parabola on ?-axis, and then choosing straight lines using vanishing point which established using weighted least squared(WLS) method. After that the lines are clustered and are matched to a three lanes model under polar coordinate. Besides, a curve lane marking searching method is proposed, which firstly searches the region of interest(ROI) and lane markings iteratively in close and far region respectively, then matching the lane marking as cubic curve using least squared method, after comparison of two matching methods. Furthermore, to solve the problem of lane markings flashes, the probability of lane markings are filtered, in processing of Kalman filtering, which are used to keep lane markings between continuous images. Besides, a side-lane driving judgment method using random seeds are proposed, which solves the problem of error detection of side-lanes. Experiment shows that the algorithm has a high detection rate and has strong environment adaptability.Thirdly, a vehicle recognition method using multi-feature is proposed. In which, vehicles are searched using image entropic units, according to entropic threshold which calculated by OTSU method. And the combination rules of the entropic units are proposed to eliminate the disturbance that comes from regions which are small and have high complexity. In order to optimize the results of vehicle detection based on image entropy, all the results are filtered using the shadow of vehicles, which based on the remarkable gray-level difference between lane markings and the shadow of vehicles. After that, the position of vehicles that in next frame are calculate using Kalman filtering method, which reduces the size of ROI and improves the efficiency of the algorithm. Experiment verifies that the vehicle recognition method has a high recognition rate and could adapt to complex road environments and varying illuminations, and also has a good real-time performance.Finally, according to imaging model, the transformation relations between image coordinate system and vehicle coordinate system are calculated. Furthermore, position the intelligent vehicle using the transformation relations and the lane detection results, and then warning information are calculated according to positions of the intelligent vehicle and other vehicles on road, to ensure driving safety.In order to verify the performance of the algorithm mentioned in this dissertation, the software of driving assistant system is designed and programed, and are tested together with BJUT-IV. The results show that the system has reached the expected target and also has a good environment adaptability and real-time performance.