Vehicle overtaking hazard detection over onboard cameras using deep convolutional networks

Abstract

The development of artificial vision systems to support driving has been of great interest in recent years, especially after new learning models based on deep learning. In this work, a framework is proposed for detecting road speed anomalies, taking as reference the driving vehicle. The objective is to warn the driver in real-time that a vehicle is overtaking dangerously to prevent a possible accident. Thus, taking the information captured by the rear camera integrated into the vehicle, the system will automatically determine if the overtaking that other vehicles make is considered abnormal or dangerous or is considered normal. Deep learning-based object detection techniques will be used to detect the vehicles in the road image. Each detected vehicle will be tracked over time, and its trajectory will be analyzed to determine the approach speed. Finally, statistical regression techniques will estimate the degree of anomaly or hazard of said overtaking as a preventive measure. This proposal has been tested with a significant set of actual road sequences in different lighting conditions with very satisfactory results.