Deep Learning-Based Monocular 3D Object Detection with Refinement of Depth Information.

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Monocular 3D target detection using pseudo-LiDAR shows promise but suffers from robustness issues.
• Key limitations include inaccurate target positioning and depth distribution uncertainty, stemming from imprecise depth estimation.

Purpose of the Study:

• To address the limitations of pseudo-LiDAR methods in monocular 3D target detection.
• To improve the accuracy of target depth estimation and reduce uncertainty in depth distribution.

Main Methods:

• A novel method combining image segmentation and geometric constraints for accurate target depth prediction and confidence measurement.
• Utilizing normalized target scale as a priori information to mitigate depth distribution uncertainty (long-tail noise).
• Converting refined depth maps into pseudo-LiDAR point clouds for input into LiDAR-based detection algorithms.

Main Results:

• The proposed framework significantly outperforms state-of-the-art methods on the KITTI dataset.
• Achieved improvements of over 12.37% (easy) and 5.34% (hard) on the KITTI validation subset.
• Demonstrated superior performance on the KITTI test set with gains of 5.1% (easy) and 1.76% (hard).

Conclusions:

• The developed approach effectively enhances monocular 3D target detection by refining depth estimation.
• The method successfully overcomes limitations in target positioning and depth uncertainty, leading to substantial performance gains.
• The framework offers a robust and accurate solution for 3D object detection using pseudo-LiDAR data.