A novel visual odometry correction using sensor fusion technique on multiple cameras

Abstract

Visual odometry estimates the position and orientation of the mobile robot using vision information. A good estimate is always the key for a robotic task implementation. The proposed technique is based on finding the translation component of the robot's movement with low level feature extraction using multiple cameras. The change in successive images from a camera gives an estimate of displacement of the robot while it is moving. This change is found by extracting features using the edge detection and correlation on small portion of two successive images. The proposed faster and computationally less expensive method of feature extraction is not very accurate as compared to existing feature extraction. But, proposed system with multiple cameras does not alter the speed of operation and improves accuracy using simple sensor fusion technique. Proposed method is also merged with odometry using encoder technique and results show further improvement in accuracy. Experiments reported in this paper validate that odometry calculation with multiple cameras using simple direct sensor fusion technique reduces error by atleast 60%. The results demonstrate that the visual odometry is possible with low level feature extraction.