Automatic Multi-Class Collective Motion Recognition Using a Decision Forest Extracted from Neural Networks

Abstract

This paper presents an approach to machine recognition of multiple classes of collective motion behaviours. Previous work has demonstrated that it is possible to distinguish structured collective motion from random, unstructured motion. However, it has proved difficult to use such techniques for automatically recognising specific collective motion variants such as moving in a line versus moving in a group. To enable a knowledge base to recognise multiple classes of collective motion, this paper proposes a decision forest approach. The proposed approach extracts machine-understandable knowledge from a neural network trained to automatically recognise collective motions. The main advantage of this approach is that besides being automatic, it is fast, accurate and easy to use. We show that our deep neural network achieves 90.30% accuracy for multi-class labelling of collective motion behaviours, which is more accurate than shallow neural networks for this problem. Furthermore, a knowledge base extracted using the decision forest on the deep neural network can recognise the class of random behaviour and the eight classes of collective motion behaviours with 88.81% accuracy in just 0.03 seconds, which is only 1.49% less accurate than the original deep neural network, but over 100 times faster.