Traffic Violation Detection via Depth and Gradient Angle Change

Abstract

The paper aims to develop an effective traffic violation detection system to detect traffic violations automatically from videos reported by the public, especially for case footage recorded by dashcams facing forward mounted on vehicles or helmets. We aim to address two types of traffic violations: (1) running a red light and (2) turning on a red light. The proposed traffic violation detection system includes two main parts: Violation Target Tracking (VTT) and Target Action Analysis (TAA). First, we detect red traffic lights and vehicles and obtain their locations and depths in VTT. Next, we model the vehicles’ depth and gradient angle changes to catch traffic violations. The experimental results show that our violation detection system can achieve 76.1% true accuracy and 81.9% conditional accuracy on average for all the violation cases.