News release

From: Edith Cowan University

Facing the music: Detecting dangerous driving through AI facial analysis

Researchers from Edith Cowan University (ECU) are developing new technology that could change how drunk and dangerous drivers are identified.

Using a single 3D deep learning model, researchers are able to detect three major causes of road accidents simultaneously; blood alcohol concentration, fatigue and expression, such as anger.

Blood alcohol concentration was identified with almost 90 per cent accuracy while drowsiness was recognised with 95 per cent accuracy. The technology can also determine the level of intoxication, classifying impairment into three categories; sober, moderate or severe.

Led by ECU PhD candidate Abdullah Tariq, the findings of  ‘Jack of Many Faces: A Step Towards Facial Expression and Physiological State Analysis with a Single Network’ was presented at the prestigious British Machine Vision Conference (BMVC25).

“Drink driving is a major public safety challenge across the globe and the number one contributing factor of crashes in Australia,” Mr Tariq said. “Approximately 30 per cent of accidents are due to drink driving.”

While the public safety challenge was the major motivation for the researchers, Mr Tariq said he wanted to explore alternatives to traditional detection methods.

“Traditional blood alcohol detection through breathalysers and blood tests are highly accurate, however they do have their own challenges - they can be intrusive, require active subject cooperation and don’t allow continuous, real-time monitoring,” Mr Tariq said.

“Human faces encode a wealth of information, such as emotion, cognitive behaviour or physiological state, but most AI models are task-specific – this motivated me to explore if we could develop a single model capable of handling multiple facial tasks.”

Jack of Many Faces

Dr Syed Zulqarnain Gilani from ECU’s Centre of AI and Machine Learning said this was thought to be the first study to identify fatigue, expression and blood alcohol simultaneously.

“Using a single algorithm from a video, we can detect whether the person is tired, intoxicated, and their mood,” Dr Gilani explained. “The interesting thing is in psychology literature, we know these three are related. So, if you’re very highly fatigued, it is almost as if you are drunk.

“Likewise, if your expression or mood is angry, this can lead to road rage and extremely dangerous driving.”

Dr Gilani explained that the 3D network looked at certain facial features to capture expression recognition and assessment of physiological state.

“This algorithm is smart, because it can tell the difference between whether a driver is sleepy, just making a facial expression, or affected by alcohol. By separating these factors, it can better understand the driver’s real physical state,” Dr Gilani said.

BiFuseNet

In another study, the researchers also investigated how combining infrared (IR) with normal colour video (RGB) data can further improve analysis. Their findings showed that integrating IR with RGB helps in analysing blood alcohol concentration more effectively in challenging environments, particularly under poor or low-lighting conditions.

BiFuseNet: A Multimodal Network for Estimating Blood Alcohol Concentration via Bidirectional Hierarchical Fusion, was presented at the International Conference on Multimodal Interaction (ICMI25).

Mr Tariq said the system automatically captures diverse facial dynamics, including eye blinking, subtle facial movements, and progressive facial feature changes that are critical for distinguishing between different states.

“Our rational was to develop a fully automated framework for estimating blood alcohol concentration by using RGB and IR video stream,” Mr Tariq said. “Previously this was done manually by looking at pupil dilation and eye closing ratios, but these methods may not perform well with certain factors, such as varying lighting conditions.

“BiFuseNet will automatically extract all facial features and facial geometry to estimate whether a person is intoxicated or not. Because it combines different kinds of information in a smart way, it performs better than older methods that only used one type of video.”

Facing the Future

Dr Gilani said the research could lead to an innovative, non-invasive way to help combat drink driving.

“Extensive experiments have shown this technology achieves a classification accuracy of 88.41 per cent, offering the potential for establishing a new, state-of-the-art estimation for blood alcohol concentration.”

Advancements in computer vision means the technology can even determine the level of intoxication.

“The system can classify alcohol impairment levels into three categories – sober, moderate and severe,” Mr Tariq said.