Collision detection based on scale change of image segments for the visually impaired

Abstract

A variety of portable or wearable navigation systems mounted on smart glasses and smartphones have been developed to assist visually impaired people over the last decade. In these systems, collision detection is one of the key components. Many conventional methods with the monocular vision estimate the collision risk based on the motion information of obstacles in the image by measuring the size change of objects using detected feature points and their corresponding motion vectors. However, the size change is sometimes incorrectly measured due to unreliable feature points and motion vectors. To overcome this problem, we present a motion clustering scheme to remove outliers among both feature points and motion vectors. Experimental results indicate that the proposed collision detection method outperforms the conventional one in terms of detection and false positive rates.

A Hardware-Friendly Optical Flow-Based Time-to-Collision Estimation Algorithm

“…The main advantage of using TTC is its relative simplicity. Unlike in collision warning systems based on range sensors (radar, laser, structured-light depth sensor), or stereo cameras that measure the distance to the obstacles by binocular matching, TTC computation, requiring only a monocular camera, can be done without knowing the physical distance to the objects [ 1 , 2 , 3 , 4 , 5 , 6 , 7 , 8 , 9 , 10 , 11 , 12 ]. Therefore, low cost smart video sensors for collision avoidance can be made feasible by employing TTC-based approaches.…”

“…Existing TTC estimation approaches can be classified into three broad categories: feature-based (inferring TTC from the change of obstacle scale with sparse feature tracking or matching) [ 1 , 3 , 4 , 5 , 6 , 7 , 8 ], gradient-based (direct estimation from image intensity or spatiotemporal gradients) [ 2 , 9 , 10 , 11 ], and dense optical flow-based [ 12 , 13 , 14 , 15 ]. From the perspective of deployment on smart video sensors for our application, feature-based algorithms are not very suitable for efficient hardware implementation.…”