We consider the problem of automatic video-based ski-jump distance measurement.
The procedure is split into two subproblems: predicting the landing and determining the distance of the jump. To predict the landing, we use a convolutional neural network which takes an image of the ski-jump video as input and predicts whether the ski-jumper is in the air or on the ground. To determine the distance of the jump, we use classical computer vision methods which first find the location of the jumper's feet in the image and then use measurement lines to output the precise distance.
The convolutional neural network achieves a classification accuracy of 93%. The complete procedure achieves a mean absolute error of 0.785 meters in the relevant landing area. The predicted landing and the actual landing differ by approximately one frame.
The results of the thesis contribute to the development of modern real-time ski-jump distance measurement systems.