engleski

Action Recognition in Handball Scenes

Action recognition in sports, especially in handball, is a challenging task due to a lot of players being on the sports field performing different actions simultaneously. Training or match recordings and analysis can help an athlete, or his coach gain a better overview of statistics related to player activity, but more importantly, action recognition and analysis of action performance can indicate key elements of technique that need to be improved. In this paper the focus is on recognition of 11 actions that might occur during a handball match or practice. We compare the performance of a baseline CNN-model that classifies each frame into an action class with LSTM and MLP based models built on top of the baseline model, that additionally use the temporal information in the input video. The models were trained and tested with different lengths of input sequences ranging from 20 to 80, since the action duration varies roughly in the same range. Also, different strategies for reduction of the number of frames were tested. We found that increasing the number of frames in the input sequence improved the results for the MLP based model, while it didn't affect the performance of the LSTM model in the same way.

Human action recognition, Action recognition in sport, Handball, Inception v3, CNN, MLP, LSTM

nije evidentirano