Early detection of herding behaviour during emergency evacuations

Social scientists have observed a number of irrational behaviours during emergency evacuations, caused by a range of possible cognitive biases. One such behaviour is herding - people following and trusting others to guide them, when they do not know where the nearest exit is. This behaviour may lead to safety under a knowledgeable leader, but can also lead to dead-ends. We present a method for the automatic early detection of herding behaviour to avoid suboptimal evacuations. The method comprises three steps: (i) people clusters identification during evacuation, (ii) collection of clusters' spatio-temporal information to extract features for describing cluster behaviour, and (iii) unsupervised learning classification of clusters' behaviour into 'benign' or 'harmful' herding. Results using a set of different detection scores show accuracies higher than baselines in identifying harmful behaviour; thus, laying the ground for timely irrational behaviour detection to increase the performance of emergency evacuation systems.