A Deep Learning Framework for Cycling Maneuvers Classification

In recent years, cycling has become increasingly popular globally, which takes up little space and leads to nearly no environmental damage. Bicycles permit daily commuters to travel in an efficient manner through frequent traffic congestion. Mixed traffic conditions and a complex physical environment pose difficulties to a bicyclist's activities and safety, especially when the bicyclist is engaged in more risky cycling maneuvers. To gain a better understanding of the risks inherent in various cycling maneuvers and assist in road safety assessments, an efficient system for identifying cycling maneuvers is needed. This paper proposes a new set of definitions of cycling maneuvers specific to Chinese bicyclists. The cycling maneuvers were categorized into passing, avoiding, carriageway-occupied, sidewalk-occupied, and regular riding maneuvers. In addition, a convolutional neural network (CNN)-based method was developed to classify these five cycling maneuvers. Field data from a video survey in the urban area of Xi' an, China (998 records) was used to evaluate the performance of the proposed model. The data includes human-related features, road-related features, and traffic-related features (for example, the gender of the bicyclist, cycling speed, vehicle-bicycle separation, bicycle-sidewalk separation, the width of the bicycle path, and traffic volume). A promising CNN model was identified by optimizing the model configuration and adjusting the model parameters. Five prevailing methods including multi-Logit, artificial neural network, support vector machine, random forest, and gradient boosting decision tree, were selected to conduct comparison analysis with the proposed CNN model. It is found that the proposed CNN model exhibited superior performance in cycling maneuver classification task.