| Ground intelligent vehicle is a kind of mobile robot equipped with sensors, which can perceive the surrounding environment, and run autonomously and continuously on roads and cross-countries. Its development has a significant influence on the defense, academy, society and economy, so it has become the strategic research objective of high technology of all countries. In the intelligent mobile robot community, autonomous navigation is an important research topic, and environment perception is a fundamental requirement for autonomous navigation. An autonomous navigation system is always aware of environment by3-D measurement techniques, which provide necessary terrain and obstacle data for safe driving and decision making of the mobile robot. Currently,3-D measurement techniques can be categorized into two major classes:active and passive measurement. The stereo vision technique belongs to the latter one which does not emit light and other radiation, and is a process of reconstructing3-D scene from2-D images. Compared with other techniques, it is rapid, relatively accurate, easy to be concealed and cheap. In addition, a stereo vision system is of light weight and low energy consumption. In the past decade, the stereo problem has gradually become an extensively topic of computer vision and autonomous navigation and made compelling progress. It has been widely implemented in the military and outer-space exploration, and plays a more and more important role.In this paper, binocular vision is chosen as the mode of autonomous navigation of a ground vehicle, and in-depth studies are carried on the solutions and difficulties of stereo vision. In order to improve applicability, accuracy, efficiency of a stereo vision system, several novel solutions and improved methods are presented. In addition, researches in hardware component, parameter configuration and error analysis of the system are also the main content of this dissertation. Details and results are as follows:(1) Zhang’s calibration method is a representative calibration solution based on planar pattern, and is widely applied to data. In order to make it more flexible and adaptable for calibrating extrinsic parameters, a four-control-point calibration algorithm based on collinear equation is presented. A linear expression of re-projection errors is derived on the basis of the collinear equation model, and the non-linear least square algorithm is introduced to iteratively optimize the extrinsic parameters. A four-point pyramidal method enforcing a rectangular geometric constraint is presented, which can produce more stable and accurate intial values and can converge faster. The presented algorithm is performed very well on accuracy and computational efficiency—its re-projection error is less than0.5pixels, and the number of iterations to convergence is less than10times, and the running time is less than100ms. Compared to Zhang’s method, its world coordinate can set unchanged so it produces greater practical value.(2) Dynamic programming (DP) and scanline optimization (SO) are two common methods to solve1-D energy optimization for stereo. Pix-to-Pix algorithm, being a representative matching method based on DP, can efficiently detect large occlusions by enforcing the ordering constraint. However, it doesn’t solve the difficulty of inter-row disparity discontinuity and huge matching errors in untextured regions. Therefore a matching algorithm based on bi-directional dynamic programming (BDDP) enhancing vertical consistency is presented. Firstly, raw matching cost is aggregated by a vertical window, and inter-row pixels with inconsistent disparities are punished in a new energy function.Then, after false matches are removed by the defined rules of disparity reliability, the disparity maps are filled by shiftable window method. The presented algorithm shows better performance than the other relevant works. Its average error rate is decreased by about40%compared to Pix-to-Pix algorithm, so its accuracy is significantly improved particularly in depth discontinuities and untextured regions.(3) Considering the typical problem of inter-row inconsistency operated by the SO-based matching algorithms, two pass scanline optimization (TPSO) is a feasible solution, however its computation costs are relatively large. A fast stereo matching algorithm based on TPSO is presented. By introducing disparity gradient theory, a weak consistency constraint is defined to restrict the searched range of disparity changes between adjacent pixels and incorporated in the TPSO technique along and across the scanlines respectively. The computation costs of the presented algorithm are inversely proportional to the number of smooth pixels, so the computational efficiency of the presented algorithm is improved with less risk of false match.(4) The hierarchical belief propagation (HBP) algorithm is a new milestone in the phylogeny of BP algorithms. It mainly applies a coarse-to-fine manner, so that the standard BP is capable of getting the optimal solution in polynomial time. On the basis of HBP algorithm, an efficient matching algorithm based on adaptive hierarchical BP (EBP) is presented. Firstly, adaptive data and smoothness terms are defined and messages are iteratively passed after the smoothness term is adjusted in per level to reduce the uncertainty in ambiguous regions. Secondly, messages’convergence is detected to reduce redundant computation introduced by unsynchronized convergence of pixels. Finally, the matching symmetry is employed to detect occlusions, and then the data term is reconstructed. The greedy method iteratively refines the results and propagates disparity information from stable pixels to unstable ones. The running time of EBP is independent on the number of iterations, and the accuracy and computational efficiency of EBP are both improved.(5) The graph cuts(GC) algorithms give the best stereo results among the state-of-the-arts, but their major difficulty lies in the enormous computational costs. For solving this problem, a stereo matching algorithm based on a-expansion, enforcing a segment constraint is presented. Firstly, an energy function is defined, which is robust against noises and preserves disparity discontinuity and occlusion asymmetry, so the anti-noise ability and applicability in ambiguous regions are improved. Secondly, graph-construction of the energy function is proved. Then, the segment constraint is defined, which can be combined with the distance transform to selecte the candidate correspondences of the disparity a as vertices in each expansion, thus the network size of original a-expansion algorithm is reduced greatly. Finally, the a-expansion process is operated over the decreasing order of disparity distribution to reduce the number of iterations. The presented algorithm reduces the converging computation of the other relevant works two fifth at least, moreover it obtains more accurate results and avoids noise interference.(6) The matching algorithms based on genetic algorithm (GA) are of huge space of solutions and easy to fall into a local optimal solution. For solving the problem, a global matching algorithm is presented based on twice segmentation and adaptive immune genetic algorithm (AIGA). Firstly, initial disparities in each segment are modeled by a liner planar equation and are fit with ground control points. Secondly, three techniques are presented to reduce the space of solutions and paths to be searched, which are merging color segments according to continuous disparity information, and searching solutions in neighborhood plane set instead of initial plane set, and estimating the optimal plane parameters on the segment level. The defined global energy function is optimized by AIGA. The optimizing process is guided by the immune operator to avoide the blind searching of GA. The probability of immune selection is chosen as a function of concentration and fitness to prevent the possible premature. The presented algorithm converges about9times faster than GA retaining the best individual. It obtains98%accuracy which is similarly as graph cut algorithms, but has less running time to convergence. |
PDF Full Text Request