Automatic Mapping of Finish Line Videos for the Objective Analysis of Sprint Behavior

Abstract

This paper proposes a computer vision-based methodology to generate riding line maps from bird’s eye view sprint videos. These maps can be used to objectively evaluate dangerous sprint behavior or to perform sprint performance studies. In order to generate the maps, our automatic workflow first extracts the road and riders from the video images using state-of-the-art object detection models. Next, feature points in the remaining part of the images are used to estimate the homography parameters and to stitch the overlapping images into a map of the finish zone. The same homography parameters are also used to reproject the riders onto the sprint map. Based on their positions on the map and the timing info from the video metadata, we get a spatio-temporal description for each riders’ sprint. These descriptions are stored in JSON format and can be used for further analysis. As a demonstrator, we present some examples of objective evaluations of dangerous sprint behavior. Those evaluations are based on outlier and overlap detections of the riding lines.