Analysis of the Conditions for the Occurrence of Sudden Braking using Drive Recorder Videos

– Using the Distance between Vehicles Estimated by Deep Learning –

Abstract

Detecting and predicting abnormal driving behaviors such as sudden braking counts for preventing traffic accidents. Keeping distances from other vehicles is considered one of the most common situations where sudden braking occurs. However, to detect distances, state-of-the-art 3D object detection usually requires advanced sensors such as LiDAR. In this work, we use a monocular visual odometry system to estimate distances between vehicles with only a common drive recorder. As a preparation for predicting sudden braking events, we train a mining model to detect them with the estimated distance information along with probe data. The experiments show that with the estimated distances, the model gains an improved performance on braking detection.