The accurate identification of traffic conflicts is critical for video-based safety analyses applying Artificial Intelligence techniques. The first step of this process is the identification of road user trajectories. Without accurate trajectories all subsequently derived data and insights are unreliable.

Determining accurate and reliable conflict measures (e.g. time to collision, post encroachment time, deceleration required to avoid an accident) is only possible if we minimize the error of the extracted trajectories between the real-world coordinates and coordinates within the camera image frames. Accurate coordinates are particularly important to determine the encounter type of conflicts (rear-end, side-swipe, angle, etc.) as this process uses the known latitude/longitude coordinates of roadway lane markings.

AMAG has developed a state-of-the-art camera calibration technique to convert trajectories observed in pixel coordinates to real-world latitude and longitude coordinates. This process performs best if the trajectory points of road users are detected as close to the ground as possible.

Conventional computer vision approaches annotate road users using a rectangular upright bounding box through object detection algorithms. Trajectories are determined by reducing the bounding box to a single point that represents the position of a road user at each instance in time. These single points may be chosen as the bounding box centroid or other points on the rectangle. All of these choices however introduce errors that affect the accuracy of the identified conflict measures.

To improve the accuracy of this vital step, the AMAG team has recently developed and implemented the capability to acquire more accurate locations of road users by flattening the bounding box around road users. This improvement relies on the developed capability through our camera calibration process to extract camera extrinsic and intrinsic parameters and utilize them to create a polygon surrounding the road user boundaries parallel to the ground. This cutting-edge innovation ensures more accurate real-world coordinates, producing more reliable detection of conflicts and further derived safety insights.