Parametric Model for Real Time Identification of Traffic Violations Using Surveillance Videos

Abstract

Road safety is one of the major concerns in any traffic management system. Traffic rules violations and accidents on road are major issues now-a-days. Due to traffic congestion on freeways, it is becoming increasing difficult to identify vehicles violating traffic rules. In order to solve the real world problems, a number of computer vision techniques have increased enormously. One of such area that has gained significant attention is the application of computer vision techniques in automatic traffic surveillance. In this thesis, a novel approach is proposed for detection of a number of traffic violations on highways. The proposed solution first employs machine vision techniques for segmentation of vehicles as moving objects. The technique is used to detect the moving objects from a live video stream as the fundamental component of the algorithm. The development of a suitable approach for detection of vehicle as moving object is a challenging task. Many factors like non-stationary background, unconstrained patterns, moving object patterns, should be taken into account for evaluation of object detection algorithms. The technique then performs homographic transformation for mapping the video coordinates from a real perspective view to flat plane in Google space.The technique also extracts lanes marks on the highway for identification of some of the violations. The lane marks include both straight and curves marks. The proposed technique then uses a mathematical model for translation of lane marks on the road as equation of lines and curves. Moving from frame to frame, a tracking algorithm then tracks each vehicle and generates a trace. These traces are again modelled as parametric equations. As these tracks may be nonlinear, therefore a piecewise linearity is used for the modelling and ease of computation for detection of traffic violation. The parametric equations also caters the speed of each vehicle and its direction. The model then simultaneously solves a number of equations for finding intersection of traces with the traffic lanes to identify the violations and classifies them from the list of violations in the dictionary. To cover larger length of road on highways, camera handoff algorithm is also designed. Camera handoff technique keeps track of all vehicles along with their tracing on Google maps. A vehicle which is violating traffic rule on the boundary of the traffic frame can still be detected. The video is also tagged for replay the exact time a violation incident has happened. This novel modelling approach can help machine based identification of a number of traffic violations and is of great help for ensuring safety on the roads. The proposed technique is compared with the state of the art techniques in the literature. Comparisons are done demonstrating the uniqueness of the algorithms, set of violations they detect and their performance with the other techniques.