Self-Driving Car Control Using Visual Ego-Motion Estimation

Unsupervised depth learning and ego-motion estimation of individual objects present in the view of a driverless car is a challenging task. Unsupervised learning removes the need of separate ego-motion using ground truths images and multi view cameras. For ego-motion estimation previous work mostly utilized the local pixel neighborhood and its position in the next frame. In this paper a control algorithm is proposed which utilized visual information for detecting individual objects, calculating estimated depth, and ego-motion estimation of each object detected. The Tensorflow object detection API is used to detect the objects in the view and then this result is used to calculate depth of each object and for ego-motion estimation of with respect to detected object in the view. A prototype car equipped with four cameras is used for acquisition of live video data, and subsequently each camera video is then used for fusion of extracted useful information for object detection and ego-motion estimation.