Faster AI for Advanced Driver Assistance Systems

Read how Deeplite helped E-SMART to dramatically reduce the size of its YOLOv3 PyTorch computer vision model so it could be deployed to existing ADAS systems.

smaller dnn model
X 16.8 Smaller Model
faster AI
44% Faster AI
cost savings per vehicle
20 USD per vehicle saved


As an industry leader in intelligent speed management solutions, E-SMART faced a critical roadblock in rolling out its latest innovation to customers: a computer vision algorithm that goes beyond existing speed ‘warning’ technologies to actively control a vehicle’s speed.

The challenge for E-SMART was that computer vision software is typically large and complex, requiring expensive, power-hungry graphics processing units (GPUs) to operate effectively. As a result, it was unclear how the active speed control technology could be deployed in newly-built systems – not to mention the thousands of vehicles already in the field with legacy, resource-constrained CPUs – while maintaining the cost-efficient standards that E-SMART’s customers counted on.


Mathieu Boivin“As the transportation market accelerates toward smart, tech-assisted and eventually autonomous vehicles, partnering with Deeplite will turn our vision of proactive speed control and other new technologies into a reality for fleet managers. With Deeplite, we are delivering cutting-edge tech to our customers at a price point that is highly accessible. It’s a win for everyone.” said Mathieu Boivin, President & CEO of E-SMART.



To avoid the time and expense required to upgrade hardware across customers’ fleets, E-SMART instead looked for ways to maintain the accuracy of its AI model while optimizing it to run on the existing hardware already in its customers’ trucks and turned to Deeplite and its Neutrino™ AI optimization software.

Deeplite’s Neutrino uses AI to make other AI models smaller, faster, and more energy-efficient at the edge and in other resource-constrained environments. After training E-SMART’s Convolutional Neural Network (CNN) to the acceptable level of performance (mAP) to recognize speed limit signs from the LISA data set and the ability to run on various hardware to maximize deployment flexibility, Deeplite achieved exceptional results while maintaining E-SMART’s desired level of accuracy.