A Pillar-based Lightweight 3D Detector with Density-Enhanced Pillar Encoder

To address the challenges in the field of 3D object detection, such as complex network architectures, large network parameter counts, and the prevalent use of sparse convolution which is not conducive to edge deployment, we design a lightweight and real-time 3D object detection neural network, which consists of three parts. In the pillar encoding section of the network, we resolve the issue of insufficient point cloud data within some pillars, which makes element-wise feature aggregation difficult, by introducing a Density-Enhanced Pillar Encoding module (DEPE). DEPE expands the perceptual range of distant pillars based on density information, allowing them to include more potential neighboring points and more fully leverage the feature extraction capabilities of the pillar encoding module, without significantly increasing computational load. In the backbone of the network, in line with the requirements for lightweight and real-time performance, we utilize reparameterization techniques and depth-wise separable convolutions, enabling the backbone to have the advantages of a single branch and fewer parameters during inference, while also avoiding the use of sparse convolution and thereby reducing the difficulty of network deployment on edge devices. Experiments on the KITTI dataset demonstrate that our proposed method offers significant advantages in terms of parameters and inference latency, with a mere 1.87M parameters and an inference frame rate of up to 18Hz, achieving a balance between efficiency and accuracy.