An investigation into the performance and robustness of intelligent transportation system infrastructure node sensors

This thesis investigates the performance and robustness of sensors deployed in Intelligent Transportation System (ITS) infrastructure nodes to enhance the safety of all road users, particularly in the context of intelligent and increasingly autonomous vehicles. Despite the ubiquity of surveillance cameras, phone cameras, and ADAS sensors, infrastructure sensor nodes capable of robust and reliable environmental perception are not yet commonplace, prompting the central research question: why? One hypothesis explored in this thesis is that there are still several open questions about the individual sensor or combination of sensors that will enable safety-critical Infrastructure-to-Vehicle (I2V) ITS applications. A comprehensive review of the literature on infrastructure nodes reveals that, as yet, no single sensor exists that is capable of meeting the diverse requirements of all I2V applications. Contemporary autonomous vehicles predominantly deploy one or more RGB cameras, RADAR, or LIDAR sensors for safety-critical perception tasks; consequently, these sensors set the benchmark for infrastructure node sensors in this thesis. However, this thesis also examines thermal and event-based cameras to evaluate their potential for I2V systems. Event-based cameras exhibit particular promise due to their capacity to capture asynchronous events associated with moving objects, thus reducing bandwidth while preserving dynamic range at low latency, whereas thermal sensors have been shown to significantly improve night-time performance. Given the experience of the automotive and security industries to date, RGB frame-based cameras will most likely be one of the core sensors of ITS infrastructure, and this thesis examines a number of the challenges associated with their deployment. For example, subtle issues such as the variability in image signal processing (ISP) parameters across different RGB surveillance datasets raise concerns regarding the generalisability of deep learning models trained on such data. Another example is performance degradation in camera perception performance due to low-quality, inexpensive lenses or temperature-induced misalignment between the sensor and lens, resulting in blurred images. The impact of lens blur on deep learning perception performance is characterised by employing a physically realistic lens blur model to evaluate the correlation between image sharpness and perception accuracy. Results demonstrate that perception performance correlates significantly with sharpness, a factor that can be taken into account in determining the expected performance or capabilities of downstream perception algorithms. In conclusion, this thesis addresses several key challenges in safety-critical perception from an infrastructure perspective, focusing on sensor selection, configuration, and tolerances in the context of object detection performance on an ITS node.

Funder

Taighde Eireann - Research Ireland
European Commission
Valeo Vision Systems

Publisher

University of Galway

Rights

CC BY-NC-ND

cbn

Collections

University of Galway Theses (PhD Theses)