The thesis presents a computer algorithm for automatic image analysis of handball match videos obtained from the YouTube channel of the Handball Federation of Slovenia. The presented approach involves the detection of two different objects - the ball and the goal on the playground, which represents the basis for an advanced, computer-aided analysis of handball matches. The initial step of the approach consisted of exporting images from video clips, creating an image database. Balls and goals present in the images were then manually annotated following the formatting style of the widely used dataset for object detection COCO. The prepared database was then used to train DETR to detect labelled objects. In the testing phase, it was found that DETR was able to detect the goals without errors. The detection of the balls was on the other hand less accurate and strongly dependent on the intensity of blurriness of the objects, which occurs as a consequence of the ball's speed of movement.
|
