Automatic Badminton Action Recognition Using CNN with Adaptive Feature Extraction on Sensor Data

With the fast development of sensor technology, a sensor chip can be easily inserted into the badminton racket so that the data of the activity of the badminton player can be recorded continuously. It has become a very challenging problem whether we can use these sensor-based data to effectively recognize the badminton actions. Although there have been some researches on the general human action recognition such as sitting, walking and running, there is few result or public dataset concerning badminton action recognition. In order to investigate the problem of badminton action recognition, we implement a specialized sensor chip being inserted into the badminton racket to collect the data of ten major badminton actions. On such a dataset, a deep convolutional neural network (CNN) can get a much better result of badminton action recognition than the conventional methods, but its recognition accuracy is not good enough for practical applications. The key difficulty is that the attributes recorded as time series in the sensor data have different measures and magnitudes so that the conventional feature normalization cannot work effectively in this complicated situation. In order to overcome this difficulty, we propose a specific block of adaptive feature extraction to enhance the performance of CNN for badminton action recognition. It is demonstrated by the experimental results on the sensor dataset that our proposed CNN with the adaptive feature extraction block can get 98.65% action recognition accuracy and outperforms the competitive methods remarkably.