Unseen Obstacle Detection via Monocular Camera Against Speed Change and Background Noise

This paper proposes a novel obstacle detection system optimized for mobile platforms. The system uses a long-short step Recurrent Neural Network (RNN) for optical flow estimation in various speed scenarios, combined with a global direction filter that filters out background noise, resulting in more robust and accurate obstacle detection. The system is designed to maximize processing speed and resource efficiency on mobile platforms. Performance is demonstrated on own-collected YouTube videos, achieving a precision rate of 95.2%, recall rate of 94.3%, and a frame rate of 75 FPS, surpassing state-of-the-art optical flow techniques. The proposed method is also evaluated on a real robot platform, demonstrating robust performance in detecting and avoiding obstacles of varying sizes and speeds under different lighting and noise conditions. Overall, the proposed system offers a reliable and efficient solution for obstacle detection and avoidance on mobile platforms, with high confidence in obstacle detection and avoidance in various real-world scenarios.