Research On Obstacle Detection Based On Binocular Stereo Vision

Autonomous obstacle avoidance is one of the important research contents in the field of mobile robots,and it is also one of the basic indicators for measuring machine intelligence.The obstacle avoidance problem first needs to use sensors to obtain information about the surrounding environment,and then through intelligent analysis,to detect various major obstacles,and calculate the range of accessible areas.In terms of sensors,there are two main categories: visual sensors and non-visual sensors.Non-visual sensors generally include laser radar,ultrasonic sensors,etc.,and visual sensors are more representative of monocular cameras,binocular stereo cameras,and the like.Among them,the binocular stereo camera has the advantages of long detection distance,rich information and low cost.This paper focuses on the detection of obstacles based on binocular stereo camera.The main contents of the research are as follows:(1)Introduce the imaging principle of binocular stereo camera,and analyze the performance,advantages and disadvantages of the main binocular stereo camera on the market,and independently construct a binocular stereo vision camera,in camera selection,baseline adjustment,calibration imaging.Research on synchronous acquisition,3D shell printing,etc.,and performance analysis of camera stereo matching accuracy by Block Matching algorithm,Semi-Global Matching algorithm and Efficient Large-scale Stereo Matching algorithm.Finally,the performance of the above-mentioned three binocular stereo cameras and self-assembled binocular stereo cameras was tested by contrast experiments.The advantages and disadvantages of the self-organized binocular stereo cameras were analyzed.At the same time,the distance measurement experiments on fixed targets were verified.The relationship between the focal length and baseline length of the self-assembled binocular stereo camera lens and the measurement error.(2)Introduce the static obstacle detection algorithm based on binocular stereo vision.Firstly,the algorithm acquires the three-dimensional spatial structure of the forward field of view through the binocular stereo camera,and obtains the ground plane constraint by camera calibration.Then,according to the ground plane height parameter of each spatial point,it is judged whether it is passable.At the same time,it is judged whether it is a slope point by the trend of the adjacent space points of the non-passing point,and it is judged whether it is possible to pass according to the slope.Finally,the inaccessible points are clustered to complete the detection of obstacles.Experiments show that the algorithm is robust to static obstacles.(3)In this paper,introduce the dynamic obstacle detections segmentation and tracking algorithm based on deep learning detection and stereo vision.The algorithm consists of two parts: one is the target detection segmentation algorithm based on deep learning detection and target stereo information,and combines the depth learning target detection result and the scene three-dimensional information to obtain the target detection and accurate segmentation result;the second is based on stereo vision.The data association tracking algorithm solves the data association tracking problem of the partial occlusion target and the complete occlusion target by two different data association strategies,and obtains the real-time motion trajectory of the dynamic obstacle,and provides information for the subsequent obstacle avoidance.Through the experimental analysis under multiple scenarios,this paper builds some data sets and tests and analyzes the performance of the algorithm.The experimental results demonstrate the effectiveness of the algorithm.