Community-Based Multi-layer Analytics Architecture for Civic Violations

Civic violations such as public drinking and public smoking are common in the streets and have become a norm in the community. Recent efforts fail to automatically detect, alert, and make analytics for civic violations, specifically public smoking and alcohol drinking, for computing architectures adapted for local communities. This paper proposes a community-based multi-layer analytics architecture for civic violations. The multi-level civic violation recognition is implemented in a mist-fog-edge architecture. The architecture implements decentralization to support fault-tolerance in a community-based environment. The system implements a series of deep learning models for human clustering and scenario isolation using YOLOv5 with intersection of union (IoU) and non-maxima suppression (NMS) algorithms, and violation recognition using Vision Transformers with violation amplifiers. Experimental results show that the human clustering and scenario isolation has a mAP of 0.46 using IoU as 0.3 and the violation recognition with an accuracy score of 86.96 %. The overall architecture has a total latency of 3.06 seconds for end-to-end AI inference operation.