The Effect of Tracker and Control Point Distributions on the Accuracy of Vehicle Position and Speed Estimates from Dash Camera Video

Dash cameras (dashcams) can provide collision reconstructionists with quantifiable vehicle position and speed estimates. These estimates are achieved by tracking 2D video features with camera tracking software to solve for the camera position history and then calculating speed from the position-time history. Not all scenes have the same geometric features in quality or abundance. In this study, we quantified how the accuracy of vehicle position and speed estimates from dashcam video varied with large differences in 2D tracker distributions that mimicked the continuum between barren environments and feature-rich environments. We used video from a dashcam mounted in a vehicle undergoing straight-line emergency braking. The surrounding environment had abundant trackable features, including road markings, street lights, signs, trees, and buildings on both sides. We first created a reference solution using SynthEyes, a 3D object and camera tracking program, and all of the available features, and then created eight test conditions by eliminating the available 2D trackers and 3D control points in large volumetric regions to mimic less feature-rich environments. To quantify accuracy, we calculated the bias and uncertainty in the camera position and speed estimates between the test conditions and the reference condition. For six of the test conditions, we found that removing large regions of the available tracker points and control points generated camera solutions with relatively small position errors (mean absolute errors <120 mm) and small speed errors (<±1 km/h) compared to the reference solution. One test condition did not reach a solution and another test condition with a poor distribution of few control points had larger errors. Overall, these findings showed that the match-moving methodology used here for calculating a dashcam's position and speed remained accurate even in some relatively feature-poor environments.