Automated pedestrian safety analysis using data from traffic monitoring cameras

Abstract

Transportation agencies often own extensive networks of monocular traffic cameras, which are typically used for traffic monitoring. However, the information captured by such cameras can also be of great value for transportation planning and operations applications, particularly when large data sets may be systematically analyzed. In this paper, we propose an approach to use data collected by existing monitoring cameras to automatically identify locations where pedestrian safety may be a concern. Our methodology utilizes a convolutional-neural-network-based method to recognize pedestrians in traffic camera feeds. Results are stored and aggregated, and may be queried for further analyses. The proposed computational approach may leverage hardware such as GPUs and distributed computing clusters to enable the analysis of large volumes of data. The post recognition analysis utilizes unsupervised learning methods to identify the spatial and temporal patterns of pedestrian positions, which are then correlated to specific scenarios such as usage of crosswalk, compliance with traffic signals, and pedestrian-vehicle interactions. Applications include the identification of potential safety concerns, measuring the effectiveness of proposed safety strategies, and identifying the need for improvements. This work provides preliminary results based on data from cameras owned by the City of Austin. We also discuss outputs such as pedestrian volume estimation and crossing hot-zones identification in the context of Smart Cities, and identify potential challenges and limitations.