Research On The Method Of Obstacle Detection Of Off-Road Robot Running Environment

Recently,USA had held competitions about ALV navigation in complex environment many times. The purposes of these were to check the ability of ALV's environment perception. Cross-country environment has high complexity,many kinds of obstacles and coarseness and bumpiness of the earth surface. All of these cause great difficulties for ALV's safe running. So it is a very challenging research problem for rough terrain perception. For this purpose, ALV always installs different sensors such as CCD,LADAR,sonar, etc. Vision sensors are important"apparatus"for environment perception and obstacle detection, which have the same function as our eyes. This thesis studies techniques of obstacle detection based on vision sensors. The obstacles which were researched include heave and slope. The main contents of this paper are introduced below in detail:1. Two kinds of methods about the receiving of obstacle's 3-D information based on vision sensors are analyzed from arithmetic theory, including stereo vision and monocular vision. In the stereo vision part, we analyzed and compared different kinds of stereo matching algorithms. And in the monocular vision part, applications,merits and shortcomings of monocular vision methods are summed up. From the analysis, we know that the feature matching has good anti-jamming performance. But it can only get sparse disparity and the feature distilling and orientation are instability in complex environment. The area matching can get compact disparity and has high orientation precision. It is fit for off-road. In the methods of computing slants of surfaces based on monocular vision, shading and photometric stereo suffer the effect of lamp intensities and are fit for lamps that can be controlled. The texture analysis computes slant by changes of surface texture structure in the perspective process. It needs not estimate directions of the illumination and apples to natural lamp condition.2. Binocular stereo vision is the primary method for heave obstacle detection. Stereo matching is the key tache of this method. So this paper stress researches on stereo matching. Single grey correlative matching algorithm can not satisfy requires of robustness .So the method of adding color in similarity matching feature is proposed. This paper had studied RGB,HSI and L*a*b* color model. Also we improve the L*a*b* color model combining the opponent characteristic of color cells in the retina. In this paper, we compare and analyze the performances of improved L*a*b* color model,RGB color model,HSI color model and grey correlative algorithms from many aspects, including environment adaptability,stability,real-time and the influence of matching window size. The experimental results show that performances of algorithms including grey and color(improved L*a*b* color model,HSI color model and RGB color model) exceed single grey correlative or color matching algorithms. Grey and color supplement each other and can't lack either. Improved L*a*b* algorithm has better matching performanc- es than other algorithms in smooth areas of the image because of its good distinguishing ability in the tiny color difference area.3. Slope is also a kind of obstacle on cross-country environment. The surfaces of them often cover random textures that are dense and littery, which make against stereo matching. Reference to the form of human visual understanding of the natural environment, we propose computing the slant of slope using textures analysis of the monocular vision. Aiming at difficult of random textures analysis, radial competition method is presented for extracting texture elements which can reflect the size of texture elements and barycenter distributions of the same size texture are received by the statistic method. Based on this, we fit the tilt using Agglomerative Nesting Clustering and Linear method and compute the gradients using continental distance with varying weight. In the last, tilt computing experiment,slant and gradient calibration and algorithm applicability analysis are carried on. Experiment results show that the method can compute the tilt and slant of the surface effectively and the computing error is close to the estimate error of human vision. Furthermore, the method is little affected by gray scale.4. Although the matching performance of color model algorithms exceed grey algorithm, the computing complexity of color model algorithms are higher than grey algorithm because of more similarity measures. For this problem, hardware accelerating system is designed using FPGA which has high ability of data parallel processing. In this paper, CMOS color image sensors HV7131B are used for image collection and Stratix FPGA EP1S30 is used for core processing module of the whole matching. A ping-pong 2-cache system is designed according to the transmission of HV7131B and the matching efficiency is improved in this way. In the phase of algorithm implementation, we adopt the system that matching in row unit and design 3-class gliding structure which can mitigate the time stress effectively. In this paper, we mostly design the accelerating system for the tache of the original matching cost computing using Series and Parallel turn System and Box filter technology. Also the real-time performance of improved L*a*b* and RGB algorithms are analyzed and compared according to the logical function and time simulation in Quartus. The simulation results show that the time complexity of improved L*a*b* is 1.1times higher than RGB algorithm.