Behaviour-based AI Algorithm for Crime Against a Person and Inchoate Crime in Sri Lanka

Abstract

The growing integration of digital technologies into daily life has led to a steady rise in crimes that leave behind digital footsteps in the form of text logs, voice recordings, and observation footage. In Sri Lanka, crime investigation systems such as the On-Scene Crime Reporting System (OSCRS- LK) have built up evidence collection practices and improved documentation at crime scenes. although, these systems remain limited in their ability to performance automated behaviour analysis and often rely heavily on manual explanation by investigators, which can be time- consuming and prone to human error. This research addresses this gap by proposing a behaviour- based artificial intelligence (AI) algorithm capable of analyzing multimodal evidence including text, audio, and video to detect unsure behaviour associated with crimes against a person and inchoate crimes, such as threats, harassment, and attempted violence. The proposed system employs machine learning models tailored to each data type; LLM for text processing, CNN for image processing, and YOLOv8 for video-based behavioural recognition. Extracted features from these models are combine into a unified representation, which is then used to calculate a risk score for each incident. To ensure transparency, fairness, and accountability in decision-making, SHAP explainability techniques are integrated, enabling law enforcement officers to explain the model’s predictions and understand the underlying reasoning. By combining multimodal evidence with understandable AI, this research aims to improve early threat detection, increase digital forensic capabilities, and accelerate criminal investigations in the Sri Lankan context. The proposed framework not only strengthens OSCRS-LK but also provides a scalable foundation for future AI-driven crime solutions.