High-Precision Map Reconstruction and Object Detection Methods for Autonomous Driving Scenarios Based on Image and Point Cloud Fusion

In the field of autonomous driving, high-precision (high-definition)maps (HDMap) are essential to obtain accurate vehicle location and path planning. An advanced HDMapNet architecture is proposed in this paper, where object detection functionality is embedded to significantly enhance the network's efficiency. Detection capabilities with the SD-Detection Head designed by myself. This method significantly improves detection accuracy without increasing detection time. To further optimize detection performance, we introduce attention mechanisms and GSConv (Ghost Shuffle Convolution). The attention mechanism effectively focuses on important features, improving the efficiency and accuracy of feature extraction. GSConv combines the advantages of standard convolution and depthwise separable convolution, reducing computational complexity while retaining more feature information. Experimental results demonstrate that the enhanced HDMapNet outperforms other methods on standard datasets, particularly in small object detection and real-time performance. This method provides a new solution for efficient object detection in autonomous driving systems.