A Vision-based System for Traffic Anomaly Detection using Deep Learning and Decision Trees

TL;DR

This paper presents a real-time traffic anomaly detection system combining deep learning and decision trees, capable of accurately identifying and timing traffic incidents from camera footage.

Contribution

It introduces a novel approach integrating YOLOv5-based detection with decision trees for precise anomaly detection and timing in traffic monitoring.

Findings

F1 score of 0.8571 achieved

S4 score of 0.5686 achieved

Effective real-time anomaly detection demonstrated

Abstract

Any intelligent traffic monitoring system must be able to detect anomalies such as traffic accidents in real time. In this paper, we propose a Decision-Tree - enabled approach powered by Deep Learning for extracting anomalies from traffic cameras while accurately estimating the start and end time of the anomalous event. Our approach included creating a detection model, followed by anomaly detection and analysis. YOLOv5 served as the foundation for our detection model. The anomaly detection and analysis step entail traffic scene background estimation, road mask extraction, and adaptive thresholding. Candidate anomalies were passed through a decision tree to detect and analyze final anomalies. The proposed approach yielded an F1 score of 0.8571, and an S4 score of 0.5686, per the experimental validation.