Combination of MLF-VO-F and loss functions for VOE from RGB image sequence using deep learning

Abstract

Visual odometry estimation (VOE) is important in building navigation and pathfinding systems. It helps entities find their way and estimate paths in the environment. Most of the computer vision (CV)-based VOE models are usually evaluated and compared on the KITTI dataset. Multi-layer fusion framework (MLF-VO-F) has had good VOE results from red, green, and blue (RGB) image sequence in Jiang et al. study, using the DeepNet to extract the low-level textures, edges, and deeper high-level semantic features for estimating motion between consecutive frames. This paper proposed a combined model of MLFVO-F as a backbone and loss functions (LFs) (LMSE, LMSE−L2, LCE, and Lcombi) to optimize and supervise the training process of the VOE model. We evaluated and compared the effectiveness of LFs for VOE based on the KITTI and TQU-SLAM datasets with the original MLF-VO-F. From there, choose the appropriate LF combined with the backbone for VOE. The evaluation results on the KITTI dataset show that LCE(RT E is 0.075m, 0.06m on the Seq. #9, Seq. #10, respectively), and Lcombi (trel is 2.21%, 2.67%, 3.59%, 1.01%, and 4.62% on the Seq. #4, Seq. #5, Seq. #6, Seq. #7, Seq. #10, respectively) have the lowest errors and LMSE has the highest errors (AT E is 133.36m on the Seq. #9).