Modeling Pedestrian Near-Crash Events at a Rectangular Rapid Flashing Beacon-Controlled Intersection Using Video Analytics and Long Short-Term Memory Neural Network

Abstract

Pedestrian safety is a long-standing issue in urban areas, where pedestrian near-crash events are more frequent than in suburban or rural areas. To address the pedestrian safety problem, a proactive approach was explored to assess and predict the severity of these events, which are valuable indicators of potential crashes. Object detection and tracking techniques were used to establish the temporal relationship of pedestrian near-crash events involving vehicles at an intersection controlled with rectangular rapid flashing beacons. A long short-term memory (LSTM) neural network model is proposed to warn a driver 2 s before the vehicle reaches the conflict zone. However, this scenario can be considered optimistic, as the 2 s interval represents an ideal driver’s reaction time, which is more likely to happen in a connected and automated vehicle environment where vehicles receive real-time information about their surroundings and perform some basic tasks such as braking without waiting for the driver reaction. The results demonstrate the effectiveness of the proposed LSTM neural network model, with an area under the curve value of 78.5% on the training data and an overall recall of 71.1% on the test data. The practical significance of this model is its potential to provide timely information about near-crash events, thereby enhancing pedestrian safety at critical points such as intersections.