Deep Learning for Automated Crime Scene Reconstruction from 3D Imagery: Enhancing Forensic Accuracy and Urban Safety through Computer Vision

Abstract

This research introduces a combined deep learning system for automated three-dimensional (3D) reconstruction of crime scenes, aimed at improving forensic precision, efficiency, and clarity. The system combines Convolutional Neural Networks (CNNs), Transformer-based attention methods, and point-cloud encoders to handle multimodal data from LiDAR, structured-light, and photogrammetric origins. A collection of 500 reconstructed crime scenes was utilized to train and assess the model using cross-validation methods. Quantitative findings indicate a mean segmentation accuracy of 0.80, an average reconstruction error of 7.79 mm, and a mean Intersection over Union (IoU) of 0.813, exceeding conventional photogrammetric techniques that attained merely 0.68 accuracy and 12.84 mm error, reflecting a 39.3% enhancement in total reconstruction quality. The model additionally accomplished a decrease in processing time from 264 seconds to 128 seconds per scene while sustaining robustness in varying environments, with segmentation accuracy staying above 0.75 even in high occlusion and low-light situations. A comparative analysis of scan modalities indicated the best results for LiDAR-based reconstructions (0.812 accuracy, 7.12 mm error), whereas structured-light and photogrammetry reached accuracies of 0.754 and 0.802, respectively. Assessments of explainability through Grad-CAM, SHAP, and attention-weight visualization produced interpretability scores of 8.4, 8.9, and 9.2 (on a 10-point scale), validating the model’s transparency for forensic purposes. These findings show that the suggested CNN–Transformer fusion system offers a reliable, high-quality, and legally understandable process for digital forensic reconstruction, in line with the United Nations Sustainable Development Goals (SDGs) 11 and 16 regarding safe and just institutions.