Vision-based vehicle behaviour analysis: a structured learning approach via convolutional neural networks

With the rapid development of artificial intelligence, the study of intelligent transportation is getting more and more attention and vision-based vehicle behaviour analysis has become an active research field. Most existing methods label vehicle behaviours with discrete labels and then use the vehicle trajectories or motion characteristics to train classifiers which identify vehicle behaviours. However, a simple discrete label cannot contain detailed information about the vehicle behaviour. So, inspired by structured learning, the authors design a structured label which is used to characterise the instantaneous behavioural state based on the vehicle image, including behaviour trend and degree simultaneously. A structured convolutional neural networks model is constructed to learn and predict structured representation of transient vehicle behaviour and preliminary experimental results justify the feasibility of vehicle behaviour structural analysis model, but it achieves only 53.3% prediction accuracy. To reduce the risk of overfitting to small-scale training data, the authors further propose an overfitting-preventing deep neural network, which exploits transfer learning and multi-task learning to achieve a much higher prediction accuracy of 91.1%.