CTM-Net: 3D object detection from LiDAR point clouds for autonomous driving

3D object detection is critical for real-time environment perception in autonomous driving. However, existing LiDAR-based methods often suffer from insufficient feature representation and limited robustness to geometric variations, leading to orientation and localization errors. To address these challenges and ensure computational efficiency, this paper proposes CTM-Net, a systematic optimization scheme for PointPillars that aims to achieve efficient inference on high-performance computing platforms, providing a feasible basis for future migration to vehicle-mounted edge computing platforms. The framework integrates modern convolutional designs with attention mechanisms to enhance multi-scale feature modeling without incurring heavy computational overhead. Specifically, a ConvNeXt module is embedded into the backbone to strengthen bird’s-eye view (BEV) semantic expressiveness and contextual modeling with low latency. Furthermore, a Multi-transform Feature Augmentation (MTFA) mechanism is introduced to explicitly improve feature robustness against scale and translation variations by adaptively fusing features from multiple transformed states. A modular residual feature chain is also constructed to perform residual fusion between attention-enhanced features and base features, mitigating information loss during optimization while preserving real-time performance. Extensive experiments on the KITTI benchmark demonstrate that, compared to the PointPillars baseline, CTM-Net achieves a 7.12 percentage point increase in the mAP for 3D detection and a 5.72 percentage point increase in the mAP for BEV detection, representing a significant improvement in accuracy. With an inference speed of 17.4 FPS, it strikes a competitive balance between accuracy and efficiency.