Instantaneous time to collision estimation using a constant jerk model and a monocular camera

Abstract

We present the application of a constant-jerk kinematic model to assess collision risk using a monocular camera (MC) for car-following (CF) scenarios. First, we redefined the metric of Instantaneous Time-to-Collision (ITTC). By employing the Kalman filter (KF) and data extracted from the frames, we estimated the equivalent parameters for distance, velocity, acceleration, and jerk on the image plane. These parameters serve as coefficients for the kinematic model on the image plane and allow for the calculation of the ITTC. Deep convolutional neural networks (DCNN) were employed for object detection and tracking in the experimental setup. The results of these experiments confirm the effectiveness of employing a constant-jerk model for evaluating the risk of collision, in contrast to the constant acceleration and constant-speed models. Furthermore, the results underscore the pivotal parameters for optimization to boost the effective utilization of MC data frames in ITTC estimations using uncalibrated MC.