Scene Detection in Skiing Videos

Abstract

Background
In alpine skiing, video motion analysis provides high level physiological and performance data. However, interpreting this data correctly requires a good eye, knowledge and experience. An artificial assistant could support inexperienced skiers as well as trainers, athletes and sports journalists to discern areas of possible improvement. For the artificial assistant to achieve the best results, a high quality input such as videos from broadcasted ski races is required.

These events are usually transmitted as a series of runs, each caught from different camera perspectives per competing athlete. The video segment from one athlete and one camera view point has position and feature continuity between its frames. Without this continuity, performance analysis algorithms based on concluded sequences such as joint movement or trajectory tracking would not work. This raises the need for a scene detection tool.

Objective
In the following bachelor's thesis, an application capable of detecting scene cuts in broadcasted ski races is developed, optimised and its results evaluated.

Approach
First, a representative broadcast is labeled manually to define the ground truth for the three occurring transitions: Cut, wipe and dissolve. To ensure the detection of a scene change, each transition needs to be covered by a separate algorithm. The cut algorithm makes use of the frame differences between two camera angles by calculating the overall pixel change in the HSV colour space. By developing a valid measure of quality an optimal threshold value for the calculated pixel change is determined.

Due to the complexity and differences in occurrence, wipes and dissolves will not be taken into consideration in this thesis.

Results
Although the developed algorithm takes nine times longer than Blackmagic Design's Da Vinci Resolve auto scene detector, it is still competitive in terms of quality. On top, the application can be extended and used fully automatically where the competition requires manual input and an additional parsing script for further processing.

Conclusion
The objective is achieved. Both the developed and the professional application detect scenes to an acceptable degree. With the scripted parsing tools and additional manual input during processing, Da Vinci Resolve can be used for the artificial assistant's scene detection while the developed application is extended by wipe and dissolve detectors.