Haar Wavelet-Enhanced Receptive Field Attention Convolution for 3D Point Cloud Object Detection

Abstract

Point cloud-based 3D object detection holds significant potential in applications such as autonomous driving and robotic perception. However, existing methods face challenges in handling small and occluded objects within complex environments. These limitations primarily stem from the inability of fixed receptive field convolution operations to effectively capture local details and multi-scale features. To address this, we propose a novel convolutional module, Haar Wavelet-Enhanced Receptive Field Attention Convolution (HRFAConv). Building on the classical Receptive Field Attention Convolution (RFAConv), this module integrates a multi-scale feature fusion module based on Haar wavelet transformation (HWFF). By applying Haar wavelet transformations to feature maps, HRFAConv enhances the model's sensitivity to local details, significantly improving its ability to detect small and occluded objects. Experimental results demonstrate that HRFAConv achieves remarkable accuracy improvements in scenarios involving small-scale, occluded, and densely distributed objects, while incurring only minimal computational overhead and parameter increase. This validates its effectiveness and efficiency in complex environments.