Obstacle detection in air-to-air images

One means of locating and classifying obstacles in video images is through the analysis of optical flow, defined as the motion on the image surface that results from camera and/or object motion. We present a computer vision algorithm that analyzes two classes of optical flow to determine the direction of a moving object captured in image sequences. First, the optical flow is computed with the Lucas-Kanade gradient-based technique. Then, the optical flow at each point in the image is decomposed into a translation component and an expansion component. Translation optical flow can be caused by motion perpendicular to the line of sight, whereas expansion optical flow can be caused by motion along the line of sight. By analyzing both types of optical flow, our computer vision algorithm successfully classifies scenarios into four types: (1) a collision condition; (2) a crossing condition, in which the target crosses in front of the camera; (3) a passing condition, in which the target passes to the side of the camera; and (4) a safe condition, in which the target travels away from the camera.