Adriel Isaiah V. Amoguis

As levels of road traffic congestion increase relative to population density, it is becoming increasingly necessary for traffic managers to have awareness of road situations in real-time to keep up with traffic management. There are already existing techniques and applications in computer vision that traffic managers use to collect real-time telemetry, such as but not limited to vehicle counting algorithms. However, these algorithms and applications may not be lane-aware. Enabling lane awareness to these systems allows them to be more granular, which enables more in-depth telemetry such as lane usage, driver pattern recognition, and anomaly detection, among others. Lane awareness in these systems are enabled by performing lane segmentation. This study investigates two approaches to this. The first approach uses vehicle trajectories to generate aggregated trajectory maps, which are then clustered to determine trajectory lane membership and to generate representative trajectories that describes the lane. On the other hand, the second approach takes an end-to-end method and uses road lane features such as demarcation lines to segment lanes. The first approach proved to be more viable as a lane segmentation algorithm compared to the second approach as it was able to segment lanes more reliably, given enough vehicle trajectories are present.

Insights derived from surveillance-based road telemetry are vital for traffic engineers, managers, and policymakers to make well-informed decisions regarding traffic policies. However, road telemetry tends to be general for any given road scene, leading to non-granularity and generalization of insights. This can be improved by isolating telemetry based on road lanes through lane segmentation. Knowing and segmenting lanes allows for the analysis of each lane’s individual telemetries which may contribute to various fields such as road vehicle navigation, traffic violation detection, road wear and tear detection for preventive maintenance, and even vision-based road anomaly detection, among others. This paper demonstrates how \textitfunctional lanes are empirically determined given a video recording of road scenes from fixed traffic surveillance cameras, explores its differences from \textitideal lanes derived from lane demarcation lines, and evaluates the fine-grained analysis taken from functional lane telemetry along with comparisons with ideal lanes. As a proof of concept, some example road telemetry is extracted from both types of lanes. Both systems achieved respectable performance. The ideal lane segmentation system achieved a mean pixel-wise mAP of 0.7760 with a mean F1-Score of 0.9789. The functional lane segmentation system achieved a mean silhouette score of 0.4260 with a mean V-Measure of 0.5777. Lastly, proof-of-concept road telemetry was achieved, showing feasibility.