Rafal Malewski, Head of the Graphics Technology Engineering Center at NXP Semiconductors, and Sébastien Taylor, Vision Technology Architect at Au-Zone Technologies, present the "Implementing an Optimized CNN Traffic Sign Recognition Solution" tutorial at the May 2017 Embedded Vision Summit.
Now that the benefits of using deep neural networks for image classification are well known, the challenge has shifted to applying these powerful techniques to build practical, cost effective solutions for commercial applications. In this presentation, Malewski and Tayler explain how their two companies designed, implemented and deployed a real-world embedded traffic sign recognition solution on a heterogeneous processor.
They show how the two companies used the TensorFlow framework to train and optimize an efficient neural net classifier within the constraints of a typical embedded processor, and how they designed a modular vision pipeline to support that classifier. They explain how the two companies distributed the vision pipeline across the compute elements of the embedded SoC, present their methods for evaluating and optimizing each stage, and summarize the tradeoffs made in the final implementation.
They also touch on solutions to some of the practical challenges of deploying CNN-based vision products, including support for multiple concurrent networks on a single device and efficiently managing remote network model updates.