**Acknowledgements**

*Deep Learning Applications*

**9. Conclusion**

**Figure 15.**

A threshold of activity measure can be used to filter active from inactive players, since the players that perform sports actions should make more sudden movements

Looking at activity measures in a sequence of frames, the ranking of players' activity can change between frames. So, to choose the players that are active throughout the sequence, the Active player score is calculated as the average activity measure of the player along the trajectory of the player's bounding boxes. The result is a set of player trajectories with corresponding player activity scores (**Figure 15**).

*Detected leading player (white box) and his trajectory through the whole sequence (yellow line).*

In this chapter, the applications of deep learning methods on typical CV tasks such as object detection, object tracking and action recognition are presented on

Handball is a team sport, played with a ball, with well-defined player's roles, goals and rules. During the game, the athletes move quickly throughout the field, change positions and roles from defensive to offensive and vice versa, use different techniques and actions, and doing so often get partially or completely occluded by another athlete, making player detection, tracking and action recognition challeng-

For detection, the algorithm must be able to locate an object in relation to its environment and, define that object. It is important for the detector to be as accurate and fast as possible especially if the real time detection is needed. State of the art deep learning-based detectors such as YOLOv3 and Mask R-CNN, prove to be successful for player detection, while the performance on ball detection still lags due to the combination of its small size, great speed and occlusion by the players. Once objects such as players are detected, they can be tracked. Here, the Hungarian assignment algorithm and SORT with a deep association metric (Deep SORT) are considered for tracking. The goal of a tracker is to assign the same unique track ID to the same player in consecutive frames, which is complicated by the changes of appearance and sudden motions of players. Thus, the trackers can model this motion or the changing appearance to help the association process. The Deep SORT adds an appearance model based on deep neural network features. This appearance model allows the Deep SORT method to re-identify players that have been temporarily occluded or left the scene much more successfully than the other tested methods, making it more appropriate for use in the handball domain.

For the action recognition task, LSTM network is used, as it is suited to deal with both image information contained in a single video frame and its temporal evolution during the performance of actions. The obtained action recognition results are promising, however due to dependence of the action recognition model on the performance of previous stages, i.e. object detection and tracking, the challenge remains to improve all three tasks. As in all deep learning tasks, an important factor

videos from the handball domain, recorded during training and matches.

ing problems of ongoing research interest.

corresponding to higher activity measures than other players.

**162**

This research was fully supported by the Croatian Science Foundation under the project IP-2016-2106-8345 "Automatic recognition of actions and activities in multimedia content from the sports domain" (RAASS) and by the University of Rijeka under the project number uniri-drustv-18-222.
