In a world of tired or momentarily inattentive drivers, vehicle safety features that mitigate against those problems can reduce the number of road traffic accidents. Work in the International Journal of Computational Vision and Robotics looks at how convolutional neural networks can be used to analyse input from external cameras on a vehicle to detect obstacles in the road and assist the driver in avoiding them.
Ramzi Mosbah and Larbi Guezouli of the University of Batna 2 in Batna, Algeria, have focused their attention on a forward-facing camera on a car that can see the road ahead. Images acquired from the camera are fed to the system which then determines whether an unexpected object is present in the path of the vehicle. They suggest that such an intelligent driving assistance system could alert the driver or be used to apply appropriate controls to the car directly to avoid a collision.
The system uses the Canny edge detector (considered the best algorithm of its kind) and the Hough line transform to identify the edges of the road as well as the horizon and so limit the area of each image that needs to be analysed. This reduces total computation time by allowing the neural network to ignore irrelevant parts of the image. The YOLO neural network is then used to detect any objects on the road ahead in real time.
The team adds that a second camera inside the car and pointed out the driver’s face adds a second safety feature to their overall system. This second system monitors the driver’s eyes and determines whether or not they remain closed for a significant period suggesting drowsiness or that the driver has fallen asleep. At this point, the system could sound a wake alarm as well as take control of the vehicle until the driver is back in complete control, perhaps warning them that they need to stop safely and rest before continuing their journey.
High-end vehicles already have some such systems. However, there is a pressing need for this functionality to be available to the broader general consumer market. The team is also now working to develop the detection system into a mobile device, which would ultimately make it more accessible at lower cost and without the need for the extra expense that fitting cameras to lower-budget vehicles would entail. Indeed, a standalone system that use the forward- and back-facing cameras on the driver’s smartphone might be all that is required to assist.
Mosbah, R. and Guezouli, L. (2022) ‘Convolutional neural networks for obstacle detection on the road and driving assistance’, Int. J. Computational Vision and Robotics, Vol. 12, No. 6, pp.573–594.