Activity Recognition in Nature using Sound and AI

The coronavirus pandemic has forced the Dutch government to put indoor activities to a hold. Therefore, many people seek alternatives outdoors. Especially, natural environments such as parks have become a popular place to visit. This left municipalities, park owners and nature conservers to wonder whether they could get concrete insight in the actual usage of the area. Based on this yet unknown data, parties responsible for the park's management, would be able to optimise how an area is being utilised. This paper evaluates a sound-based approach for human activity recognition in nature by using a convolutional neural network (CNN) with spectral image input. The current model is able to recognise four activities with 85% accuracy: walking, running, cycling and null (environmental noise only). Three phenomena that could affect the perceived acoustic data were investigated: sound attenuation, overlapping sounds and noise interference.