ODTFormer: Efficient Obstacle Detection and Tracking with Stereo Cameras Based on Transformer

Area of Science:

• Robotics and Artificial Intelligence
• Computer Vision
• Machine Learning

Background:

• Autonomous navigation in robots heavily relies on accurate obstacle detection and tracking.
• Existing methods often face challenges in efficiency and performance, particularly in complex environments.

Purpose of the Study:

• To introduce ODTFormer, a novel Transformer-based model for integrated obstacle detection and tracking.
• To achieve state-of-the-art performance in obstacle detection while maintaining computational efficiency for tracking.

Main Methods:

• Utilizes deformable attention to create a 3D cost volume for obstacle detection.
• Employs voxel matching between consecutive frames for obstacle tracking.
• The model is optimized end-to-end for seamless integration.

Main Results:

• Achieved state-of-the-art performance on the DrivingStereo and KITTI obstacle detection benchmarks.
• Demonstrated comparable accuracy to existing tracking models with 10-20x less computational cost.
• The Transformer-based architecture proves effective for integrated detection and tracking.

Conclusions:

• ODTFormer offers a highly efficient and effective solution for obstacle detection and tracking in autonomous navigation.
• The model's performance and reduced computational requirements make it a promising advancement for real-world robotic applications.
• The integrated approach simplifies the pipeline and improves overall system efficiency.