Stereo visual odometry using a supervised detector

Abstract

Stereo visual odometry, which is widely used in agent navigation because of its fast execution time and relatively good accuracy, is a process that computes the position and orientation of an agent using a stereo camera. Additionally, the method can build up a map which can be used for path planning. However, traditional stereo visual odometry only provides a content-less point cloud, making it difficult for path planning under some circumstances such as crossroads. To give content to the map, a convolutional neural network (CNN) classifier for detecting specific objects is created, and has been applied on the keypoints detected by oriented FAST and rotated BRIEF (ORB) detector. Thereafter, the labeled keypoints are inputted to a simplified keypoint-based visual odometry to build up the map. Our CNN classifier is trained on data provided by 2018 Chalmers formula student driveless (CFSD18) and achieved 98.97% accuracy in test data. Our stereo visual odometry algorithm is tested on an elliptic track and the mapping is compared to a raw GPS sensor data as ground truth. The average processing time of the whole approach is 68.6ms per frame with a CPU-based multi-threaded algorithm. Since our keypoint detector is a combination of ORB detector and CNN classifier, which is based on supervised learning, we have called it a supervised detector.