3D reconstruction of traffic scene images using neural radiance fields (NeRF) and target association

Challenges such as multi-target motion, complex occlusion and drastic perspective changes in traffic scenes often cause target aliasing and semantic inconsistency in 3D reconstruction using Neural Radiance Fields (NeRF) in dynamic environments. This study proposes a spatiotemporal consistency NeRF method based on dynamic instance decomposition. A foreground-segmentation network first semantically decouples each frame into static background and dynamic instances. Cross-frame instance correspondence is established using optical flow constraints, while a temporal consistency module estimates target motion parameters for continuous dynamic mapping. Independent instance index encoding is embedded in body density and radiation functions to enable NeRF to distinguish objects’ implicit radiation distributions during training. The method jointly optimises volume rendering errors for static and dynamic regions, introducing a spatiotemporal smoothing regularisation term to maintain geometric continuity and temporal consistency. Experimental results demonstrate that static background reconstruction achieves an SSIM of 0.982, 0.034 higher than standard NeRF. Dynamic object reconstruction improves vehicle Object Association Accuracy (OAA) to 0.958 and Temporal Consistency Index (TCI) to 0.953. Overall reconstruction error averaged over all valid rendering pixels is reduced to 0.012, with improved geometric continuity. These results confirm that the proposed approach enables high-precision 3D reconstruction and target association in multi-object dynamic traffic scenes.