Enhancing 3D Object Detection for Edge Devices: A PointPillars Approach with Attention Networks and Range Adjustment

The advancement of 3D object detection on edge devices poses significant challenges, especially for autonomous robots and vehicles requiring computational efficiency within constrained hardware environments. This study addresses these challenges by proposing an optimized 3D detection framework that integrates the Jetson Xavier AGX platform with LiDAR sensors, enhancing the PointPillars model to balance accuracy with efficiency. The approach employs a Range Expert mechanism to dynamically adjust detection range and voxel size, reducing processing load while sustaining high precision. A Lightweight Attention Pyramid Network further improves feature extraction, enhancing the model's ability to detect objects in diverse scales and conditions. Evaluation experiments have demonstrated that the system can achieve a 3D object detection task within 10 meters with a frame rate of 23.39 FPS, a mean Average Precision (mAP) value of 0.969, and a recall value of 0.875. The results underscore advantages of the model in functionally relevant tasks such as navigation, collision avoidance, and path planning scaling a real-world 3D detection system on autonomous assets with limited resources.