3D pedestrian detection based on hybrid multi-scale cascade fusion network

Abstract

Improving pedestrian safety on the road is one of the essential tasks of autonomous driving. LiDAR-based intelligent perception systems can provide necessary guarantees for pedestrian safety in autonomous driving by accurately detecting pedestrians in real time. However, the detection performance suffers from the small-scale issue and blurred boundary of pedestrian point clouds. This work proposes a novel PillarHMCNet, which focuses on enhancing the feature representation of pedestrian point clouds to improve the 3D detection performance, tackling the issues mentioned above. Concretely, a Hybrid Encoder (HE) module is proposed to extract sparse and dense features of pedestrians through customized encoders, enhancing the feature representation of small-scale objects. Afterward, a Multi-scale Cascaded Feature Fusion (MCFF) module is introduced to fuse multi-layer sparse and dense features, improving the pedestrian contour representation. Finally, a dense head is used to conduct 3D detection based on the output of the MCFF module. Moreover, a direction-sensitive loss is leveraged to improve the model’s positioning accuracy by introducing the angle and distance-IoU (DIOU) losses. Quantitative and qualitative evaluations are conducted on the KITTI dataset, and in the detection of pedestrians and cyclists in 3D mode, our model outperforms PillarNet by 4.22% and 1.17%. The results verify the effectiveness and universality of the proposed method in intelligent perception of autonomous driving. The code will be available at https://github.com/CCZU-Myan/PillarHMCNet.