Vision-based traffic monitoring system with hierarchical camera auto-calibration

Resumen

In recent decades, traffic has become a great problem in most of the cities around the world, due to the increment of the number of vehicles and the transport infrastructures demand. It represents a social, economical and environmental phenomenon which involves all the society. Therefore it is crucial to consider it as a key area to improve. Along these lines, and to guarantee a safe, fluid and sustainable mobility, it is important to analyse the behaviour and interaction of vehicles and pedestrians in different scenarios. Not long ago this task was performed only by human operators at traffic control centres. However, the advances in technology, suggest an evolution in the methodology towards the automation of the surveillance and control. The presented work describes a target detection system on transport infrastructures, for applications in the framework of Intelligent Transportation Systems (ITS). Particularly as a monitoring system to detect and predict incidents (traffic accidents, dangerous manoeuvres, traffic jams, etc.) on critical areas of transportation infrastructures, like intersections or roundabouts. To achieve this objective, a monocular vision-based approach with hierarchical camera auto-calibration is proposed. It is able to measure parameters of vehicles and pedestrians, as an input of a future incident detection system, traffic control system, etc. The common problem of computer vision in this kind of applications, and where the proposed solution puts special emphasis, is the adaptability of the algorithm to external conditions. Accordingly, illumination or weather changes, occlusions, instabilities due to wind or vibrations, etc. are compensated. Furthermore the algorithm is independent of the position of the camera, and it is able to work with variable pan-tilt-zoom cameras in fully self-adaptive mode. One of the contributions of this thesis is the extraction and use of vanishing points, through structured elements of the image, to obtain an automatic calibration of the camera without any prior knowledge. This calibration provides an approximate size of the searched targets, improving the performance of the detection steps. To segment the image, a background subtraction method, based on Gaussian Mixture Models (GMM), image stabilization and shadow detection algorithms, is used. Finally about tracking, the traditional idea of considering objects as a whole is rejected. Instead, characteristic target features are extracted and analysed to achieve an optimal clustering which deals with occlusions. In the document, the results obtained in real traffic conditions are presented and discussed, without any prior knowledge of the scene or the camera.