Computer Vision for Road User Detection and Localisation in Intelligent Transportation System Infrastructure

Demand for intelligent transportation systems is on the rise. Intelligent features have the potential to solve numerous problems related to congestion, pollution, as well as safety of transportation. In order to implement different intelligent transport system applications, a great amount of real-time traffic information is needed for robust and reliable decision-making. This thesis explores solutions for utilising roadside cameras as real-time data acquisition tools for smart traffic applications. The research of this thesis was focused on solving problems related to practical applicability of computer vision for road user detection and localisation in roadside camera views. Experimental testing was carried out to validate the proposed methods, as well as produce results highlighting the capabilities of computer vision solutions for intelligent road user monitoring applications. A fast road user detection algorithm based on fusion of background subtraction and a convolutional neural network classifier was proposed. The detector was shown capable of running at over 30 frames per second on a low-cost single-board computer. The accuracy of the detector was only slightly lower than those of compared state-of-the-art object detectors that performed detection at single digit processing speeds. The developed detector is effective in edge computing applications that perform real-time road user detection. The sensitivity of roadside camera-based road user localisation was studied by quantifying errors caused by different inaccuracies in the measurement. The experiments highlighted that minor errors in road user detection can propagate in to large errors in the localisation. Furthermore, slight deviation in the camera extrinsic parameters can render the localisation results unusable. Careful calibration of a roadside camera is required for reliable localisation results. An automated calibration approach for roadside cameras was proposed, based on utilisation of global navigation satellite system positioning information of connected vehicles. Compared to previous automated methods, the proposed method is more generalisable, includes a feedback loop for verifying the calibration, and calibrates cameras to a global coordinate system. Experimental testing was carried out to validate the performance of the method. Future development should focus on further tuning the proposed methods with increased robustness for a wider variety of operating environments. Roadside cameras monitor highly different road environments with different types of traffic. Camera views can have immense variation in road user types, traffic density, as well as appearance and orientation of road users. Extensively robust algorithms are required to successfully deploy roadside computer vision solutions in intelligent transportation applications.