Improving Visual Path Odometry for Smart Autonomous E-Bikes

Accurate global localization is very crucial in Autonomous Road vehicles. Visual Odometry (VO) uses data obtained from the movement of the vehicle and camera images to apply image processing and computer vision techniques to estimate the trajectory of the vehicle. VO can also be fused with other odometry systems such as LiDAR-based odometry to try to enhance the accuracy of the algorithm. This paper introduces innovative techniques and proposes a system to improve path odometry for smart autonomous E-bikes. Firstly, we present an image pre-processing technique using VGG16, a deep convolutional neural network, which enhances feature extraction and tracking. Secondly, we propose a method for generating optical flow reference points using the Shi-Tomasi corner detection algorithm, improving motion estimation in challenging scenarios. Lastly, we implement eight monocular visual odometry (MVO) algorithms on a smart autonomous E-bike prototype, comprising four feature-based and four appearance-based MVOs.