Reseach On Biomimetic Optical Flow Based Mobile Robot Obstacle Avoidance Method

Mobile robots autonomous navigation in unknown complex environment hasbeen a hot research topic, and robot obstacle avoidance is the most basic function ofthe autonomous navigation system. Although the visual sensor has the advantages thatother sensors can not match, the study of visual sensors used for obstacle avoidancehas been less. Currently, visual sensor in mobile robot navigation and obstacledetection has drawn more and more attention. Especially in recent years, through theintense study of insects flight mechanisms, scientists found that due to the specialstructure of their compound eyes, flying insects don’t use binocular stereo vision toestimate depth, instead they use optical flow to control their various flight behavior.Inspired by insects biological visual mechanisms, the study of the monocular opticalflow based obstacle avoidance has become a hot issue recently. In this project, thetopic of optical flow based robot obstacle avoidance techniques was studied, the mainresearch work includes:(1) The traditional optical flow algorithm measure was expanded and used foralgorithm evaluation. As optical flow here is used for real-time obstacle avoidance, sothe calculation method must be both fast and accurate. Therefore a time measure wasadded to the traditional error measure when comparing the performance of opticalflow algorithms. As spatial-temporal gradients being very important in differentialmethods, experiments of five gradient operators were done using two synthetic imagesequences, and the expanded measure was used for quantitatively evaluate the fivegradient operators.(2) As for the flaws that traditional differential HS algorithm and LK algorithmhave, three improved differential methods were presented, including sparse LKalgorithm, Gaussian pyramidal LK algorithm and the combination algorithm of localand global method (CLG). Compared to the original method, experimental resultsshow that these three improved algorithms improved the performance in one aspect orso. But overall comparison of the three algorithms shows that the combinationalgorithm of local and global method outperforms the other two.(3) An algorithm for mobile robot obstacle avoidance was proposed based on theTTC obstacle map. First, we used the least square method to calculate FOE (focus ofexpansion), and then to calculate TTC (time to collision/contact) from optical flowfield. Then an obstacle map was constituted based on the TTC computed, and obstacleavoidance was achieved based on the obstacle map. Experimental results show thatthis algorithm can guarantee a robot wander without collision in an unknownenvironment.(4) Because original balance strategy only fits to corridor-like environment, and it doesn’t apply to complex environment. So a modified balance strategy wasproposed and used for robot obstacle avoidance. The modified balance strategy firstdetermines if there will occur imminent front collision based on sum of the magnitudeof optical flow in the central part of the image. Then original balance strategy isapplied when avoidance behavior is needed, and if isn’t, the robot just don’t take anyavoidance action and move on. Experiments found that the modified balance strategycan make better decisions for the robot.