FF-VO: Fast Visual Odometry for Dynamic Environment Based on FastestDet

Visual Odometry (VO) faces a critical challenge in real-world deployment: striking a balance between enhancing reliability in dynamic surroundings and reducing computational resource consumption. To address this challenge, we propose FF-VO, a stereo visual odometry system that leverages a compact object detection algorithm, FastestDet, to provide essential semantic information. The front-end of FF-VO utilizes the FAST operator and optical flow tracking to ensure system lightweightness. We deploy this system on an ARM platform and evaluate its performance on two standard datasets across three different scenarios. Additionally, we conduct real-world online evaluations with the system running at 30 FPS. The results show that the system can alleviate the computational resource overhead while ensuring accurate localization and robustness in dynamic environments.