Multiple Object Tracking From LiDAR Point Cloud Data
Manninen, Matias (2025)
2025
Sähkötekniikan DI-ohjelma - Master's Programme in Electrical Engineering
Informaatioteknologian ja viestinnän tiedekunta - Faculty of Information Technology and Communication Sciences
Hyväksymispäivämäärä
2025-06-09
Julkaisun pysyvä osoite on
https://urn.fi/URN:NBN:fi:tuni-202506076911
https://urn.fi/URN:NBN:fi:tuni-202506076911
Tiivistelmä
Automated driving technologies increasingly rely on accurate environmental perception to enhance safety and efficiency in complex urban environments. While most research has focused on vehicle mounted sensors, infrastructure based sensing has emerged as a valuable complement, particularly in areas with frequent interactions between light traffic and vehicles.
This thesis investigates the use of infrastructure-based LiDAR sensing for analyzing pedestrian–tram interactions near a tram stop in Tampere, Finland. A stationary LiDAR sensor was used to capture three-dimensional point cloud data, from which light traffic and trams were detected and tracked using a Detection-Based Tracking pipeline. The tracking employed a clustering-based detection method followed by data association, enabling computational assessment of interactions.
To evaluate traffic safety and interaction dynamics, two established traffic conflict metrics of Time To Collision (TTC) and Post Encroachment Time (PET) were calculated from the tracked trajectories. A quantitative evaluation of the tracking accuracy was first conducted using an external annotated dataset, allowing for the computation of standard Multiple Object Tracking metrics. In the absence of ground truths in the actual deployment data, qualitative evaluation was performed through visualization of tracking and detection results.
The results showed that PET was a more reliable metric than TTC in complex real-world scenarios, primarily due to TTC’s assumptions of constant velocity and its forward-looking nature. Most interactions fell into “safe” or “critical” categories based on established threshold values from literature, validating the applicability of the system. These findings demonstrate the potential of LiDAR-based infrastructure perception systems to enhance safety assessment in the development of automated public transport environments. Furthermore, the results offer useful insights for transitioning the system from offline post-processing toward real-time implementation in practical deployments.
This thesis investigates the use of infrastructure-based LiDAR sensing for analyzing pedestrian–tram interactions near a tram stop in Tampere, Finland. A stationary LiDAR sensor was used to capture three-dimensional point cloud data, from which light traffic and trams were detected and tracked using a Detection-Based Tracking pipeline. The tracking employed a clustering-based detection method followed by data association, enabling computational assessment of interactions.
To evaluate traffic safety and interaction dynamics, two established traffic conflict metrics of Time To Collision (TTC) and Post Encroachment Time (PET) were calculated from the tracked trajectories. A quantitative evaluation of the tracking accuracy was first conducted using an external annotated dataset, allowing for the computation of standard Multiple Object Tracking metrics. In the absence of ground truths in the actual deployment data, qualitative evaluation was performed through visualization of tracking and detection results.
The results showed that PET was a more reliable metric than TTC in complex real-world scenarios, primarily due to TTC’s assumptions of constant velocity and its forward-looking nature. Most interactions fell into “safe” or “critical” categories based on established threshold values from literature, validating the applicability of the system. These findings demonstrate the potential of LiDAR-based infrastructure perception systems to enhance safety assessment in the development of automated public transport environments. Furthermore, the results offer useful insights for transitioning the system from offline post-processing toward real-time implementation in practical deployments.