Neural Network and Prior Knowledge Ensemble for Whistle Recognition

Whistle recognition is becoming an increasingly crucial aspect of RoboCup. Therefore neural networks are being utilized in this field more frequently. They are typically more effective than straightforward conventional approaches but still have flaws in fields that require prior knowledge, as conventional approaches do. In this work, we present an approach that can outperform standalone variants of both methods by fusing prior knowledge of traditional methods with a neural network. Additionally, we were able to keep the composite system runtime efficient on the integrated hardware of the NAO Robot.