Visual Sensing for Comprehensive Granular Traffic Data Collection in Support of Intelligent Transportation Applications

Comprehensive granular traffic data collection has been identified as one of the most essential tasks in the development of intelligent transportation systems (ITS). With the recent advances in electro-optical sensors and computational capacities, machine vision is regarded as cost-efficient for traffic data collection. However, the existing vision-based methods are only applicable for collecting limited types of traffic data, and a unified method allowing for comprehensive traffic data collection is still missing in the existing literature. Moreover, there are still challenges that exist for the use of the existing vision-based methods in real settings, including low accuracy and compromised applicability due to shadows, irregular appearances of road user silhouettes, and vehicle occlusions, all of which are critical for their adoption in real traffic settings. To address this hurdle, this dissertation aimed to develop a versatile vision-based method that allows for comprehensive granular traffic data collection by use of the power of computer vision and artificial intelligence. In the proposed method, a deep-learning and view-geometry-based pipeline was developed to retrieve the road user’s 3D spatial information in the sequence of 2D frames. Based on this temporal-spatial dataset, the traffic data regarding road user’s type, location, dimension, speed, volume, and so on can be conveniently obtained thereof. In addition, to record the collected traffic data and facilitate the data interpretation and analysis, an improved structure from motion framework was developed for 3D road infrastructure reconstruction, from which the granular road infrastructure information, such as main road, crosswalk, road marking, lane division, traffic sign, and lamp pole can be obtained. Furthermore, an unmanned aerial vehicle (UAV)-based approach was developed as a benchmark for the evaluation of the traffic data from surveillance cameras, in which the measurement error caused by depth relief of road users and the perspective distortion of the onboard camera is handled. Several potential intelligent transportation applications were designed and conducted to evaluate the performance of the proposed method. The results showed that the comprehensive granular traffic data can be successfully extracted by the proposed method with promising accuracy, and effectively addressing the practical challenges regarding vehicle shadows, vehicle occlusions, and irregular appearance of road user silhouettes. The contribution of this dissertation lies in the creation of a new computer vision-based method and its demonstration for the reality capture of traffic scenes in support of the development of ITS