Lightweight Stereo Vision for Obstacle Detection and Range Estimation in Micro-Mobility Vehicles.

Area of Science:

• Computer Vision
• Robotics
• Embedded Systems

Background:

• Micro-mobility vehicles require efficient obstacle detection and range estimation in constrained environments.
• Existing solutions often face limitations in cost, power, and computational resources for embedded applications.

Purpose of the Study:

• To propose HAGVNet, a lightweight stereo matching network for embedded ranging.
• To validate its deployability in a target-level ranging pipeline using YOLOv11n.
• To achieve accurate distance and 3D position estimation for micro-mobility applications.

Main Methods:

• HAGVNet utilizes a hierarchical attention-guided cost volume (HAGV) for modulated cost modeling.
• Employs ConvNeXtV2-style 2D cost aggregation for enhanced stability and boundary consistency.
• Integrates depth statistics within detected regions for target distance and 3D position estimation.

Main Results:

• HAGVNet achieves 0.73 px EPE on SceneFlow with 20.08 G FLOPs, demonstrating an excellent accuracy-complexity trade-off.
• On an embedded Jetson Orin Nano Super, it reaches 46.3 FPS (TensorRT FP16).
• Field tests show 0.5-8.6% relative ranging errors within 2-10 m.

Conclusions:

• HAGVNet provides a computationally efficient and accurate solution for embedded ranging.
• Its performance validates practical feasibility for low-speed target-level ranging in micro-mobility.
• The network offers a promising approach for real-world deployment under strict resource constraints.