Proactive safety monitoring: A functional approach to detect safety-related anomalies using unmanned aerial vehicle video data

Abstract

Abstract The advent of smart cities has motivated the field of transportation safety to transition towards a proactive approach that anticipates and mitigates risks before crashes occur. One of the stepping-stones to complete this transition is to monitor the risk levels of the transportation roadway system continuously. With the advancement in emerging technologies, computer-vision-based safety analysis approaches that focus on risk assessment using traffic conflicts quantified by surrogate safety measures extracted from traffic videos have been proposed. As the detailed movements of all the vehicles are logged in the video data, the quantified traffic conflicts can depict the whole picture of risk levels for a specific location. However, the capability of using traffic videos and the quantified traffic conflicts for proactive safety monitoring has yet to be investigated widely. In addition, one of the limitations of past safety research is that safety-related data are mostly aggregated into summary statistics (such as the total number of crashes/conflicts), which constitutes a major simplification of the real risk levels that change over time. Therefore, this study contributes to the literature by proposing a novel functional approach that models time series of a variant of the time exposed time-to-collision safety indicator directly to detect safety-related anomalies from traffic video data. Functional data smoothing methods that can mitigate overfitting and account for the nonnegative characteristics of safety indicator are applied. Nine commonly used functional anomaly detection methods are summarized and compared using approximately one-hour of traffic video data recorded by an unmanned aerial vehicle at a signalized intersection. Ground truth safety-related anomalies are identified by manually reviewing the video data, which is used to validate the performances of anomaly detection methods. Good separation between safety-related anomalies and non-anomalies (0.85 area under the receiver operating characteristics and 0.8 area under the precision-recall curve) has been found. By applying the proposed safety-related anomaly detection method, significant amount of human resources can be liberated.