Trajectory Visualization of Ego-Motion Videos with Pedestrian Based on Monocular Visual Odometry and Machine Learning

摘要

In this paper, we proposed MVO system combined withmachine learning. Both methods had accumulated errorproblem. However compared with geometric MVO, theperformance of MVO combined with machine learningshowed higher accuracy and precision. It indicates that themoving objects effect poses estimation. The pipeline inthis paper provides a baseline for future work to detectmore moving objects. From this experiment, it alsoindicates that the result of machine learning also dependson good labelled training and test data, and computationalburden is another issue with the increase of errorcorrelations.