Study On Obstacle Detection Algorithm In Electronic Travel Aids For Blinds

Obstacle detection is the key technology in the system of Electronic Travel Aids (ETA). The accuracy and instantaneity of the obstacle detection algorithms is an important symbol to weigh practicability of ETA. For a long time, the blind or visually impaired peoples walked basically by traditional ways such as white cane and dog guides, while those guide tools are limited severely. Therefore, it is helpful to get great research values and social meanings and broad application prospects by developing practical intelligent ETA system.In recent years, many researches about ETA system for the blind have been done by scholars at home and abroad and a lot of methods have been put forward, which are helpful for improving practicability of ETA system. However, existing ETA systems have generally detected obstacles with ultrasonic wave, laser or infrared ray. All of them have many shortcomings such as narrow beam and limited detection range, while detection distances of a few ETA systems based on monocular vision are still detected by additional diastimeter. In the way of information outputting, most of ETA systems have adopted electro-tactile displays or synthetic sonar, so the blind need long periods of training to learn and adapt them. Therefore, In this paper, the B-ETA system model and its key algorithm based on binocular stereo vision are the core research questions. The complete B-ETA system includes three important parts: the global navigation, local obstacle avoidance, and audio output, the thesis researches ETA system technology for the blind based on binocular vision, of which the core problem of obstacle detection is researched systematically and deeply. The main research achievements of this thesis include four parts as follows:1) We analyze the research status of the ETA technology in detail. These existing ETA systems are analyzed comparatively in various ways, and he defects of those methods are pointed out. Then a new electronic travel aid system model based on binocular vision for the blind is proposed (B-EAT, for short).2) For the image segmentation in the detection obstacles, the thesis presents two-dimensional Otsu thresholding algorithm based on inter-class with intra-class variance. And the thesis also presents minimum error thresholding segmentation algorithm based on 3D grayscale histogram (3D-MET, for short) by applying three dimensional normal distribution and minimum relative entropy of information theory to threshold image segmentation. In addition, the thesis presents two rapid implementation methods based on lookup table, or dimension reduction and rebuilding of the 3-Dimensional histogram to improve arithmetic speed of the new methods respectively. Then we will apply the new methods to obstacle segmentation in the follow chapters.3) Based on visual attention mechanism of human eyes, the thesis puts forward a new salient obstacle detection algorithm combining weighted frequency-tuned and 3D-MET to realize the automatic selection of salient obstacles in complex scenes. The results of experimental comparison show that new methods, compared with others, can highlight the most salient objects in images accurately and effectively, and it can produce full resolution saliency map.4) The nearest obstacle is the biggest security threat on the road for the blind or visually impaired peoples. So the accuracy and instantaneity of the obstacle distance measurement, is critical for obstacle avoidance. Based on binocular stereo vision, the thesis researches obstacle detection and distance measurement methods. For the determination problem of tolerance coefficient a in the process of designing the thresholding mechanism of the disparity error energy in Unreliable Disparity Estimation, the thesis presents the utility function about tolerance coefficient a and reali2es self-adaption determination of a, by combining satisfectory optimization with the utility theory, which solves well the weakness that only could be solved with experience before.