REDO-SLAM: Robust Efficient Dynamic Optical Flow-Based SLAM with Deep Reinforcement Learning

We introduce REDO-SLAM, a robust visual SLAM system specifically designed for dynamic environments, which addresses a major drawback in many existing SLAM techniques. Our approach optimally selects a model for segmenting dynamic objects by combining a computationally efficient optical flow neural network with segmentation models, reducing the number of frames needed for applying resourceintensive semantic segmentation models. By using optical flow values, we can predict object masks, enabling the system to effectively manage dynamic objects. We also employ deep reinforcement learning (DRL) to determine the most appropriate frames for accurate mask prediction based on motion-aware rewards. We assessed our method using the TUM and Kitti datasets, showcasing its outstanding performance in dynamic settings. Our method significantly lowers computation time and maintains satisfactory performance in obtaining keyframe trajectories. In comparison to recent visual SLAM systems, our approach is more scalable, efficient, and resilient.