MTData is a leading Australian fleet telematics provider. With more than 20 years of experience, this Australian company supplies a fleet management system tailored to specifically address customer's business needs.
As part of their latest ADAS (advanced driver assistant system) and fleet management solution the 7000AI, they have partnered with RidgeRun.ai. The partnership has integrated computer vision and video analytics for real time safety monitoring, while ensuring the transport industry's concerns about privacy are achieved in particular avoiding facial recognition of individuals.
The following sections describe some of the features we've helped MTData integrate into their system.
Face Landmarks
Drowsiness is one of the most critical conditions to detect in a driver. Early detection of sleepiness may avoid accidents and even save lives. Some of the current detection techniques make use of facial expressions to evaluate the tiredness level of a driver.
Detecting accurate face landmarks is essential for correctly assessing drowsiness from facial expressions. While there are some off-the-shelve models available, MTData's use case presented some extra challenges, namely:
Images came from a night vision camera
Motion blur due to car movement
Real-time requirements
RidgeRun helped MTData integrate accurate facial landmarks detection, in real time, for night-vision cameras.
The final model provides 106 high-precision face landmarks in the LaPa format. Further models consume these points and detect unwanted conditions, such as potential drowsiness.
Gaze Tracking
Driver distraction is another critical condition to detect to ensure trip safety. Accurately predicting where the driver is looking at can help assess if the driver's attention is compromised. From the same night-vision camera, posing the same challenges, MTData monitors the gaze of the drivers to aid in fleet security.
RidgeRun integrated real-time monocular gaze tracking within MTData's safety vision pipeline.
The final model provides pitch and yaw angles for each eye that approximate the driver's gaze direction, which allows further components to analyze if a distraction is taking place.
NVIDIA TensorRT and DeepStream Deployment
The models developed above are only useful if they can be efficiently deployed to operate in real time. As part of the work done for the 7000AI, RidgeRun.ai optimized and deployed the models as TensorRT engines. This format is the most optimal way to make efficient use of the NVIDIA GPUs.
Furthermore, these TensorRT engines were integrated into NVIDIA DeepStream. DeepStream makes use of GSteamer, a highly efficient multimedia framework. This ensures not only fast inference, but also optimal video pre-processing and real time data flow.
RidgeRun deployed deep learning models to DeepStream to be able to run efficiently and in real-time.
Real Time Computer Vision for your Application
Modern hardware and optimization techniques make it possible to compute video analytics in real time and with a small resource footprint. MTData's fleet safety system is the perfect example on how to efficiently add video intelligence to your product by making the most out of your hardware.
Are you wondering if your idea is too far-fetched? Let's talk! At RidgeRun.ai we will be happy to meet, get to know your requirements and design the appropriate solution for you.
Comments