Non-GPS navigation system for criminalistic investigation on mobile robot

Abstract

In the paper we describe visual odometry algorithm for monocular camera. It is developed for use on a mobile robot involved in criminal investigation conducted by Polish police forces. FAST keypoints with ORB descriptors are used as input to the motion estimation. Concept of pose graph is implemented in order to improve accuracy of the results obtained. The algorithm was evaluated on a 400 m long outdoor dataset included in KITTI Benchmark (18.41 % position error) as well as on a short dataset collected indoors in one of the offices at our University (5.34 % position error). Detailed discussion of the results is presented together with the guidance for further development of the method.