AI powered armed robbery detection and real time alert system.
| dc.contributor.author | Kabanda, Simon | |
| dc.contributor.author | Ayella, Patrick Ojok | |
| dc.date.accessioned | 2026-10-19T08:44:14Z | |
| dc.date.available | 2026-10-19T08:44:14Z | |
| dc.date.issued | 2026 | |
| dc.description | Undergraduate research report | |
| dc.description.abstract | Armed robbery continues to pose a serious threat to public safety, economic stability, and community well-being, particularly in rapidly developing urban regions. Traditional surveillance systems in Uganda, such as standard CCTV installations, rely heavily on human monitoring, which is prone to fatigue, delayed reaction, and missed detection of critical events. More critically, authorities are often notified too late victims under attack or held at gunpoint are unable to call for help, meaning law enforcement only learns of an incident after robbers have already fled the scene. This dangerous notification gap is the core problem this system addresses. To address this, the project proposes an AI-Powered Armed Robbery Detection and Real-Time Alert System that autonomously analyses live video feeds to detect weapons, aggressive behaviours, and robbery-related activities without human intervention. Upon detecting a threat, the system instantly dispatches SMS or mobile alerts directly to security authorities, reducing the longer time for the authorities to get notified about the incident and therefore take the required action in time. Computer vision was chosen as the core technology due to its unmatched ability to interpret visual information with speed and consistency beyond human capacity. Specifically, the Vision Transformer (ViT) model was adopted for its self-attention mechanism, which captures long-range spatial relationships across an entire video frame making it far superior to traditional CNNs for detecting complex, dynamic threat scenarios such as weapons or aggressive confrontations, while minimising false alarms. The system was trained on the CamNuvem Dataset one of the most realistic and reliable benchmarks for robbery detection, distinguished by its use of real-world surveillance footage sourced directly from YouTube, Facebook, and social media, rather than staged or synthetic scenes. Of its 972 total videos, 486 capture actual robbery events, carefully edited for annotation precision. Its weak-label protocol video-level labels for training and precise frame-level labels for 49 test videos mirrors real deployment conditions, making model evaluation highly credible. The ViT model was evaluated using accuracy, precision, recall, and F1-score, confirming strong detection performance and low false-alarm rates, validating the system's suitability for real-world deployment across Uganda's urban security landscape. | |
| dc.description.sponsorship | Dr. Odongtoo Godfrey ; Dr. Godfrey K Mirondo ; Busitema University | |
| dc.identifier.citation | Kabanda, S. & Ayella, P. O. (2026). AI powered armed robbery detection and real time alert system. [Unpublished undergraduate research report]. Busitema University. | |
| dc.identifier.uri | https://bdears.busitema.ac.ug/handle/123456789/11102 | |
| dc.language.iso | en | |
| dc.publisher | Busitema University | |
| dc.title | AI powered armed robbery detection and real time alert system. | |
| dc.type | Other |