Vision-based obstacle detection using optical flow

• Main Authors: Go, Candice Jean V., Medina, Sherry Maine E., Navoa, Ronel P., Tan, Daniel A.
• Format: text
• Language: English
• Published: Animo Repository 2006
• Subjects: Visual perception; Optical films; Computer Sciences
• Online Access: https://animorepository.dlsu.edu.ph/etd_bachelors/14181
• Institution: De La Salle University

Numerous robots use vision-based techniques in detecting obstacles and tracking object. As vision becomes more prominent in the field of robotics, there is a call for equally intensive research on other vision-based techniques. Optical flow is a monocular vision-based technique that is gaining attention for the past three decades. Optical flow is a technique that derives motion from a sequence of images. This motion is embodied in the optical flow field, illustrated by a vector diagram of the movement of all pixels in the image sequence. Vision-Based Obstacle Detection Using Optical Flow adapts an existing optical flow algorithm and analyzes its importance on both real and synthethic images. Four implementations of optical flow were tested for it's accuracy with respect to the time complexity, both the OpenCV and Barron implementation of Lucas Kanade and Horn Schunck algorithms. Among these implementations, the Lucas Kanade algorithms has the best tradeoff between accuracy and time. Time-to-collision information is computed using gradients of the velocity components of the optical flow. Using time-to-time collision information, the image was segmented into different regions, each region corresponding to an obstacle. Tests were conducted to measure the accuracy of the time-to-collision computation. Results show a large amount of errror due to the sparsity of the flow field as well as the varying values of the time-to-collision in each pixel, suggesting the use of refinement methods.