Keypoint-Aware Single-Stage 3D Object Detector for Autonomous Driving.

Area of Science:

• Computer Vision
• Machine Learning
• Robotics

Background:

• Current single-stage 3D object detectors rely on single points in feature maps for confidence scores, lacking boundary and inner features.
• This approach fails to establish a clear association between regression boxes and confidence scores, leading to detection inconsistencies.

Purpose of the Study:

• To introduce a novel single-stage 3D object detector, the keypoint-aware single-stage 3D object detector (KASSD).
• To enhance feature representation and address the limitations of point-based feature extraction in existing detectors.

Main Methods:

• Designed a lightweight location attention module (LLM) with feature reuse strategy (FRS) and location attention module (LAM) for efficient multi-level feature representation.
• Introduced a keypoint-aware module (KAM) to model spatial relationships and learn semantic information by representing objects as keypoints.

Main Results:

• KASSD achieved a competitive performance of 79.74% Average Precision (AP) on the moderate difficulty level of the KITTI dataset.
• The proposed method maintained a high inference speed of 21.8 FPS.

Conclusions:

• The keypoint-aware approach effectively models spatial relationships and semantic information for improved 3D object detection.
• KASSD offers a promising solution for efficient and accurate single-stage 3D object detection in real-world applications.