Objective In the field of autonomous driving, high-precision object detection is crucial for ensuring safety and efficiency. A common approach is to use voxel-based methods, which are susceptible to the quantization grid size. Smaller grid sizes make the algorithm more computationally intensive, while larger grid sizes increase quantization loss, leading to the loss of precise position information and fine details. Successive convolution and down-sampling operations can further weaken the precise localization signals in the point cloud. To improve the orientation perception and accuracy of object detection, we propose a frustum transform-based method that uses RGB images to extract features and fuses them with distance information from LiDAR. This approach optimizes the strategy for extracting orientation features from the 3D point cloud. To reduce the model's dependence on annotated data, we also design a semi-supervised learning architecture that employs an adaptive pseudo-labeling method, thereby further reducing the false alarm rate of the group voting-based method.
Methods We propose a LiDAR-RGB fusion network based on the frustum transform (Fig.1). Specifically, texture information is extracted from the RGB image by a deep network and fused with distance information from the LiDAR to maintain the integrity of the 3D spatial features (Fig.2). Subsequently, the weights of the voxel spatial features are optimized using the channel attention module (Fig.3). Finally, a semi-supervised learning architecture (Fig.4) is employed to reduce the false alarm rate by utilizing the spatial feature fusion module (Fig.5) and the group-based voting module. The comparative learning module is used to improve the reliability of the detection.
Results and Discussions The proposed method was evaluated on the KITTI dataset (Tab.1). Our method achieved 56.30% accuracy in pedestrian detection and 75.88% accuracy in vehicle detection, with a detection rate of 21 FPS. In the ablation study of the LRFN (LiDAR-RGB Fusion Network) model (Tab.2), the RVFM (RGB Voxel Feature Module) improved the accuracy in recognizing occluded objects (Fig.6-7). The channel attention module was analyzed in comparison with other fusion modules (Tab.3, Fig.8). In the semi-supervised learning experiments, the teacher model of this study was compared with the 3DIoUMatch model (Tab.4), and the results validated the effectiveness of our teacher model. In the ablation study (Tab.5), the baseline was improved by 8.61% using the full model. These results show a significant improvement over existing methods, highlighting the detection performance of the RVFM and the teacher model.
Conclusions In this study, we propose a 3D object detection technique based on frustum transform and semi-supervised learning architecture. This method maps 2D image features to 3D space, generates homogeneous RGB image voxel features using LiDAR depth distribution information, adaptively selects the voxel space, and optimizes the fusion feature characterization capability through the Channel Attention Module. Finally, targets are detected using the 3D Region Suggesting Network Module. In the ablation experiments (Tab.2), the detection accuracy of the baseline model improved when using the RGB image feature module. The RVFM effectively solved the orientation and proximity problems in visual sample analysis (Fig.6-7). Additionally, the SFF (Spatial Feature Fusion) and GBV (Group-based Voting) modules were proposed to reduce the false alarm rate, and the comparative learning module was introduced to improve the consistency of output results from different views of the student model. The experimental results (Tab.1) show that the LRFN-S (LiDAR-RGB Fusion Network-SLL) method proposed in this paper achieved significant performance, with 75.88% and 56.30% accuracy on the KITTI dataset for automobile and pedestrian detection benchmarks, respectively.