Real-time Detection of Wearable Camera Motion Using Optical Flow

The efficient use of image sensors has been one of the top challenges for computer vision researchers for several years. Detecting and tracking objects, video surveillance, navigation, and many other real-time applications depend on motion estimation for moving camera. In this paper, a real-time method for the detection and classification of the motion of a wearable, moving monocular camera is proposed. This approach was adopted to be used with smart glasses to assist people with visual field defects. Five main motion classes (corresponding to the five primary degrees of freedom) were detected using optical flow and motion velocity vectors calculation. These classes cover different degrees of freedom including rotation and translation. The proposed method classifies the type of camera motion as static, translation/rotation left, right, up or down. This classification is important for object detection and tracking that can alert the user to the potential hazards outside their field of view. The proposed approach has been tested on a real first-person perspective video captured by a wearable camera. The experimental results demonstrate that the proposed method classifies the type of motion successfully in real-time and can be used as part of low-cost wearable solutions for various forms of vision loss assistive technologies. Promising performance results of 84% correct states for camera motion detection were obtained.