Jahnavi V

DOI: https://doi.org/10.5281/zenodo.23032511

Modern crime scene investigation involves substantial evidence collection and analysis, including CCTV footage and images requiring manual and automated inspection. The proposed AI Crime Scene Evidence Analyzer is a full-stack forensic intelligence web platform with an integrated suite of six AI modules, including YOLOv8 weapon and evidence detection trained on a domain-specific weapons dataset, frame-level surveillance video analysis with UCF-Crime anomaly timestamps, Scene Change Detection using image alignment (ORB) and RANSAC homography estimation, followed by localization of altered regions using SSIM dissimilarity, and a novel multi-factor weighted risk score generation with 0-100 case-level risk indices, MongoDB-based audit trail generation of evidence custody steps, and a case-aware forensic LLM chatbot named ARIA with Claude API context injection. Additionally, the platform automatically generates a PDF forensic report for the submitted case within five seconds. Evaluation demonstrates that the weapon detection module achieves mAP@0.5 > 75% on the Kaggle Weapons Detection dataset and the Scene Change Detection module reliably localizes altered regions in aligned before-after crime scene image pairs with dissimilarity < 0.75 global SSIM. The contribution of this work is a forensic intelligence platform leveraging six individually published forensic computer vision techniques alongside a novel risk score engine, evidence comparison pipeline, and chain-of-custody logging, implemented in React.js, Python Flask, Node.js, and MongoDB, and evaluated on UCF-Crime and Kaggle Weapons Detection benchmarks.

Forensic Intelligence; YOLOv8; Crime Scene Analysis; Scene Change Detection; Structural Similarity Index; Image Alignment; Video Anomaly Detection; Chain of Custody; Large Language Models; Risk Scoring; ARIA Chatbot; UCF-Crime Dataset; Evidence Management

The AI Crime Scene Evidence Analyzer is a full-stack forensic intelligence platform which employs six AI modules – YOLOv8 object detection, Scene Change Detection using ORB alignment and SSIM comparison, video surveillance analysis, multi-factor risk scoring, MongoDB-based chain of custody logging, and a Claude API-injected LLM chatbot – to provide accelerated evidence processing and management. Evaluated on the UCF-Crime and Kaggle Weapons Detection benchmarks, the system demonstrates mAP@0.5 > 75% for the YOLOv8 weapon detection module and reliable alteration localization for the Scene Change Detection module when the global SSIM of two aligned images is below 0.75. The core contribution of this work is the design of a novel multi-factor risk scoring system which enables the combined six-module system to estimate a 0-100 case risk score, which is useful for triaging of active cases. Overall, the work demonstrates that the integration of AI into the forensics workflow can reduce manual processing time of evidence, reduce opportunities for human error in chain of custody, and provide an objective risk scoring system which can be utilized in high-volume evidence management scenarios. Future directions include the addition of blockchain-based custody logging, live CCTV stream processing, deep learning-based change detection for more significant viewpoint variations, and multilingual support for the ARIA chatbot. ACKNOWLEDGMENT The authors would like to acknowledge the esteemed faculty members and project guides of the Department of Computer Science and Engineering (Artificial Intelligence and Machine Learning), Don Bosco Institute of Technology, Bengaluru for their valuable suggestions, motivation, and support throughout the project on AI Crime Scene Evidence Analyzer. We also express our sincere gratitude to the team members for their contributions towards completion of this work.

[1] A. Thakur, A. Shrivastav, R. Sharma, T. Kumar, and K. Puri, “Real-Time Weapon Detection Using YOLOv8 for Enhanced Safety,” arXiv:2410.19862, Amity University, Oct. 2024.
[2] C. L. Huang and B. Y. Liao, “A Robust Scene-Change Detection Method for Video Segmentation,” IEEE Trans. Circuits Syst. Video Technol., vol. 11, no. 12, pp. 1281–1288, Dec. 2001, doi: 10.1109/76.974682.
[3] W. Chen, and M. Shah, “Real-World Anomaly Detection in Surveillance Videos,” in Proc. IEEE/CVF CVPR, Salt Lake City, UT, USA, 2018, pp. 6479–6488.
[4] S. Farber, “AI as a Decision Support Tool in Forensic Image Analysis: A Pilot Study on Integrating Large Language Models into Crime Scene Investigation Workflows,” J. Forensic Sci., DOI: 10.1111/1556-4029.70035, Apr. 2025.
[5] P. Shanthi and V. Manjula, “Weapon Detection with FMR-CNN and YOLOv8 for Enhanced Crime Prevention and Security,” Sci. Rep., Nature, DOI: 10.1038/s41598-025-07782-0, 2025.
[6] A. Germanov, A. Khanova, and V. Nabiyev, “Anomalous Weapon Detection for Armed Robbery Using YOLOv8,” in Proc. IEEE 6th Eurasia Conf. IoT, MDPI Eng. Proc., 2024.
[7] Z. Wang, A. C. Bovik, H. R. Sheikh, and E. P. Simoncelli, “Image Quality Assessment: From Error Visibility to Structural Similarity,” IEEE Trans. Image Process., vol. 13, no. 4, pp. 600–612, Apr. 2004, doi: 10.1109/TIP.2003.819861.
[8] E. Rublee, V. Rabaud, K. Konolige, and G. Bradski, “ORB: An Efficient Alternative to SIFT or SURF,” in Proc. IEEE Int. Conf. Computer Vision (ICCV), Barcelona, Spain, Nov. 2011, pp. 2564–2571, doi: 10.1109/ICCV.2011.6126544.

Jahnavi V (2026). AI Crime Scene Evidence Analyzer. International Journal of Computer Science Engineering Techniques, 10(5), 191–194. ISSN: 2455-135X. DOI: https://doi.org/10.5281/zenodo.23032511