Improved 3D Object Detector Under Snowfall Weather Condition Based on LiDAR Point Cloud

Abstract

LiDAR sensors are now used to supplement structure information and depth information for 3D object detection in automated driving. In adverse weathers, however, LiDAR tends to collect many noisy points in rainy or snowy days, which may disturb the results of object detection. In order to enhance the performance of the detector, we improve existing LiDAR-only 3D object detectors from two aspects under real snow weather condition. Firstly, double-attention block including point-wise attention and channel attention is applied to reweight the input feature of stacked pillars for crucial information extraction. Secondly, a lightweight and effective global context based pillar feature refinement extraction block is employed to capture long-range contextual information. It aims to filter local noisy information in the feature map, especially for the data collected in adverse weather conditions. Moreover, most of the previous works tend to focus on dataset under normal weather condition, so driving scenarios in adverse weather will bring challenges to the generalization of the model. Hence, to adapt our network to diverse domains better, we design a maximum mean discrepancy (MMD) block to get the distribution of domain feature representations as well as calculate the MMD loss in training process. Accordingly, the distribution discrepancy of two domains is narrowed. The performance evaluated on Canadian Adverse Driving Condition (CADC) Dataset collected in snowfall weather condition and KITTI dataset verifies the improvement of our approach. Code is available at <uri xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">https://github.com/jiajia0408/i3detector_snowfall</uri> .