A robust method for automatic player detection in sport videos

In the last years, pattern recognition approaches using machine learning techniques to construct object detectors have encountered great success in the domain of visual detection. However, the application of automatic and robust player detection in sport videos is a particularly challenging task because of the small size, and high variability in shape and appearance of the objects to detect, and the influence of noise due to motion blur and video coding. In this paper, we present a sport player detection method based on convolutional neural networks and compare it to the well-known Adaboost approach which is known to perform well in other contexts. Compared to traditional methods, these approaches do not use background substraction after camera motion estimation. Experiments showed that both methods are very robust and effective and that convolutional neural networks perform better with an average gain of 4% on various datasets.