Fast Lane Detection Based On Computer Vision

Lane detection technology is a very fundamental and importantcomponent of intelligent transportation systems. Computer vision based lanedetection algorithms use cameras mounted on the board of the car to collectroad images. Lane detection algorithms are performed to detect lane markersand warn drivers to avoid lane departure. The thesis focuses on the researchof vision based lane detection technology. On the basis of previous lanedetection technology, the thesis proposes a novel fast lane detectionalgorithm for embedded systems in order to meet the real time speed needsof practical intelligent transportation systems. The algorithm can detect lanemarkers with high speed and detection rate. The detection rate is88.7%. Onthe ATOM based Asus embedded platform, the algorithm can process23.2frames per second.The project uses a visual sensor equipped in the car to collect real roadimage. Image pre-processing methods are introduced, including RGB to grayscale converting. Pre-processing methods create a favorable way for otherimage processing.Lane marker edge detection algorithms are illustrated. The thesis makescomparisons among common edged detection algorithms. Image processingtechnology uses first degree differentiator to approach the gradient of grayscale, like Roberts and Sobel. It also uses the maximization of signal noiseratio to do lane detection, such like Canny detector. On the basis of color andgeometric feature of lane marker, the thesis uses local extreme value to dolane edge detection. The proposed lane edge detection algorithm iscompared with other edged detection algorithms in terms of edge detectionrate, computational cost and non-lane-marker edge pixel ratio. Theexperiments show that the proposed algorithm has less computational costand contains less non-lane edge pixels.The thesis illustrates lane marker modeling and fitting algorithms. Previous lane marker detection algorithms include straight line based Houghtransform and third degree B-Spline based random sample and consensusalgorithms. The thesis analyzes the geometric feature of lane detection anddevelops a novel lane marker geometry matching algorithms. The algorithmuses the overlap feature of the path lane segments do the fitting andmatching process. Discussion is held to compare the proposed fittingalgorithm with previous algorithms by similarity of the model andcomputational cost. The proposed geometric lane fitting model is fast andcapable of detection all kinds of lane markers.The thesis collects multiple data sets to conduct comparisonexperiments of the proposed algorithm and other previous lane detectionalgorithms. Lane marker positions are manually marked to get numericaldetection rate and speed. The experiments prove the algorithm to be betterthan previous algorithms in terms of detection rate, robustness, applicabilityand speed.The algorithm is transplant to two embedded platforms. One is ARMbased and the other is ATOM based. Optimization is conducted to make useof the hardware architecture. After optimization, it can operate at12.9frames per second on the OMAP4430platform and23.2frames per secondon the Asus MiniMax platform. After discussion and analysis, it is proved tomeet the speed needs of intelligent transportation on the ARM and ATOMplatforms. By comparison with other algorithms, the proposed algorithmshows speed advantage in the embedded platforms.