Shutter Speed

Traffic accidents are an all too prevalent and serious global issue, with millions occurring each year, resulting in significant casualties and economic losses. According to the World Health Organization, automobile accidents are a leading cause of injury and death worldwide, particularly among young people. In light of these statistics, the automotive industry has been increasingly focusing on integrating advanced safety systems, driver assistance technologies, and autonomous driving features to enhance overall driver safety.

Driver assistance systems have become integral components of modern automobiles. Adaptive cruise control is one such technology that automatically adjusts a vehicle's speed to maintain a safe following distance from the car ahead. Lane departure warning and lane-keeping assistance systems help prevent unintentional lane departures by providing visual or haptic alerts and steering assistance when necessary. Forward collision warning and automatic emergency braking are designed to detect potential collisions with vehicles or obstacles ahead, alerting the driver and autonomously applying the brakes if necessary.

In recent years, autonomous driving technologies have also started to gain traction, promising to revolutionize road safety further. These systems, ranging from basic driver assistance to full autonomy, aim to reduce human error — a significant factor in many accidents. But if you have been browsing at dealerships lately, you know very well that these systems do not come cheap. They rely on expensive sensing and computing elements, and huge amounts of development time are required to get them operating acceptably.

Some relief may eventually find its way to our pocketbooks thanks to the work of a team led by engineers at the University of Malaga in Spain. One of the most fundamental tasks involved in modern automobile safety systems revolves around locating other vehicles, and determining their direction of travel and rate of speed. Traditionally, this requires pricey equipment, but the researchers managed to show that it is possible using only a single windshield-mounted camera, which many vehicles are now already equipped with.

The team leveraged a deep learning object detection neural network, paired with a tracking algorithm, to inspect images captured by a single camera. This component is able to locate other nearby vehicles, and by looking at the changes over time, calculate their trajectories. Next, a linear regression model determines the vehicle speed by analyzing its position and size in the image. This system is capable of tracking multiple vehicles simultaneously, and if any safety criteria are violated, it can provide an alert to the driver.

In addition to being relatively simple and inexpensive, the system is also ready to use right out of the box. No initial calibration process is required.

To validate their approach, the team leveraged the Prevention dataset, which consists of video sequences of real-world traffic scenes. The video is paired with LiDAR measurements to provide ground truth measurements of other vehicles' speeds. The mean absolute error between the new technique and the LiDAR readings was about 1.8 kilometers per hour, but some outliers were also detected. These outliers resulted from factors like lighting and occlusions, which caused problems for the object detection model. The researchers noted that improvements in object detection algorithms could significantly enhance the overall accuracy of their system.

At present, the team is working to better deal with outliers, which could create serious problems under real-world conditions. If they can solve these issues, they hope to integrate their system into real vehicles.