Street Detection Gaussians: An Approach for Efficient Real-Time 3D Scene Reconstruction

In this work, we present Street Detection Gaussians (SDGs), a unified framework for real-time 3D reconstruction and dynamic object detection in urban environments. Our method combines 3D Gaussian splatting with a streamlined detection pipeline to address the limitations of existing NeRF based and point cloud approaches in handling large-scale dynamic scenes. By integrating tracked poses for dynamic objects and leveraging depth-aware segmentation via Grounded-SAM, our model achieves 135 FPS rendering speeds on the Waymo dataset while maintaining photorealistic reconstruction quality (PSNR: 34.92 dB, SSIM: 0.940). Experimental results demonstrate state-of-the-art performance in both static scene rendering and dynamic traffic element detection, outperforming 3DGS by 25% in mAP for object localization. This work bridges the gap between high-fidelity reconstruction and real-time urban monitoring applications.