Tiny Obstacle Discovery by Occlusion-aware Multilayer Regression.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Detecting tiny obstacles with monocular cameras is challenging due to weak and inconsistent edge cues.
• Small object size and similar appearance to free space hinder accurate obstacle capture.

Purpose of the Study:

• To propose an occlusion-based multilayer approach for robust tiny obstacle detection.
• To enhance the accuracy and completeness of obstacle discovery in monocular vision.

Main Methods:

• An obstacle-aware occlusion edge generation fuses cues from multilayer regions to intensify object characteristics.
• A multistride sliding window strategy captures proposals enclosing tiny obstacles.
• A novel obstacle-aware regression model with a primary-secondary regressor generates an obstacle-occupied probability map.

Main Results:

• The proposed method improves accuracy by approximately 19% over FPHT and PHT.
• Achieves comparable performance to the MergeNet approach.
• Experimental validation on two datasets demonstrates effectiveness across different scenarios.

Conclusions:

• The occlusion-based multilayer approach effectively addresses challenges in tiny obstacle detection.
• The developed regression model and edge detection significantly enhance detection capabilities.
• The method offers a promising solution for obstacle avoidance in monocular vision systems.