Visual Object Recognition with 3D-Aware Features in KITTI Urban Scenes

Area of Science:

• Computer Vision
• Robotics
• Autonomous Systems

Background:

• Driver assistance systems and autonomous robotics require robust environment perception.
• Vision sensors offer cost-effective 3D scene understanding compared to LiDAR.
• Detecting road participants like cyclists and pedestrians in urban settings is crucial for navigation.

Purpose of the Study:

• To detect and estimate the orientation of cars, pedestrians, and cyclists using stereo color images.
• To develop 3D-aware features that capture both appearance and depth information.
• To extend the Deformable Part Model (DPM) for improved 3D object detection.

Main Methods:

• Computed 3D-aware features from stereo color images.
• Extended the part-based object detector (DPM) to utilize 2.5D data (color and disparity).
• Conducted a detailed analysis of the training pipeline and evaluated performance on the KITTI dataset.

Main Results:

• Achieved increased detection ratios for cars and cyclists compared to a baseline DPM.
• This work is the first to report results using stereo data for the KITTI object detection challenge.
• The proposed 3D-aware features enhance the understanding of road scenes.

Conclusions:

• The proposed method effectively utilizes stereo vision for 3D object detection in autonomous driving scenarios.
• The enhanced DPM with 3D-aware features shows competitive performance on challenging urban datasets.
• This research contributes to advancing the perception capabilities of autonomous systems.