Learning-Based Heatmap-Guided Model for Monocular Visual Odometry

Visual odometry (VO) is a key part of autonomous navigation systems, particularly for robots and autonomous vehicles. Conventional feature-based or direct approaches for VO are powerful but have limitations in environments with dynamic lighting or few features, respectively. In this work, we introduce two key contributions: a heatmap-guided model that integrates key point information into a learning-based framework, significantly enhancing robustness in challenging environments with lighting variations, and a stable pose learning model that utilizes a novel rotational manifold loss function and an adaptive gradient clipping technique called BBClip to improve the stability and accuracy of training. The effectiveness of our method was demonstrated on the KITTI dataset and the CVPR Visual SLAM Challenge dataset. Compared with state-of-the-art learning-based methods, our model achieved a 7.6% reduction in translation error and a 1.4% reduction in rotation error on the KITTI sequence. On the CVPR Challenge dataset, our method achieved 82% higher trajectory accuracy than did traditional geometrical methods. Notably, the proposed approach has superior performance in complex and real-world environments; it is therefore a promising solution for next-generation autonomous systems.