Deep transfer learning-based anomaly detection for cycling safety

Despite the general improvements in road safety, with the growing number of bicycle users, cycling safety is still a challenge as demonstrated by the fact that it is the only road transport mode with an increase in the number of fatalities in EU cities.

Moreover, to analyze the problem to improve the road transport system, the traditional network screening based on crash statistics is a reactive approach and less effective due to the lack of suitable bicycle data availability, as well. In such a framework, new opportunities for data collection in smart cities and communities are emerging as proactive approaches to identify critical locations where safety treatments can be effectively applied to prevent bicycle crashes.

This research applied a deep transfer learning model to detect anomalies in cycling behavior that can be associated with traffic conflicts or near-miss crashes.

The paper presents how to build a users' tailored riding model named DTL AD to detect and localize riding anomalies by using a set of data in the National Marine Electronics Association (NMEA) string of Global Navigation Satellite System (GNSS) recorded with instrumented bicycles by different cyclists.

More specifically, DTL AD exploits a convolutional autoencoder (CAE) with transfer learning to reduce data labelling and training effort.

A case study demonstrates the identification of anomalies in cycling behavior visually represented on Geographic Information Systems (GIS) maps, showing how data clustering is well located in high-risk areas.