A Multi-Sensor Video/LiDAR System for Analyzing Intersection Safety

Aotian Wu, Tania Banerjee, Anand Rangarajan, Sanjay Ranka

Abstract:

We introduce an integrated video and LiDAR analytics system for analyzing pedestrian and vehicle behavior
at traffc intersections. Subsystems for each modality leverage
advanced deep-learning techniques to detect pedestrians and
vehicles and then use a Kalman-flter-based tracking algorithm
to generate tracks. The video and LiDAR tracks are then
aligned spatiotemporally onto the same coordinate system with
synchronized clocks.
We evaluate the benefts of these two modalities by providing
both qualitative and quantitative comparisons, utilizing low-level
measures such as detection and tracking accuracy, as well as highlevel measures such as severe events. Additionally, we compare
the two modalities at different times of the day and show that
LiDAR is competitive with video during daylight hours and
signifcantly outperforms video at late evening when lighting
conditions are poor. To the best of our knowledge, this study
represents the frst detailed comparison of these two modalities
for observing traffc intersections.