PhD position: On-edge learning for smart cameras

     
Employer
Published
WorkplaceLeuven, Flemish Region, Belgium
Category
Position

Description

PhD position: On-edge learning for smart cameras

This PhD is about making smart cameras really smart: we will enable them to learn from their experiences.

In this PhD project, the candidate will perform research on deep on-edge learning. The goal is to enable smart cameras, including an embedded computation platform, to retrain themselves. Indeed, this goes beyond to the classical approach where smart cameras are only used for inference of incoming camera images. When self-learning on edge devices, the advantage is that rich sensor data can stay on board. This would be very interesting in the scope of saving communication bandwidth, as no training images or network weight files need to be transferred. The device boils down to using non-supervised learning, semi-supervised (if some data can be sent over the network and can come back labelled) or using learning strategies that only use high level supervision, like Reinforcement Learning.  

To adapt generically learned models, we will explore self-supervised knowledge transfer from a slow, complex, but generic neural network to a situation-specific, small and fast neural network. For example: Just after installing a sensor in a new (fixed) position, it can use a pre-trained general object detection neural networks such as YOLO or SSD as a "teacher" for pedestrian detection, which are accurate but quite slow on full HD resolution. We can let the smart camera itself use these detections to train a smaller neural network, which is only suited for the fixed position of the camera but is fast enough. The right equilibrium between generalization and overfitting must of course be found. The PhD project comprises of three steps: Implementation of on-edge learning on high-end commercially available embedded GPU platforms, Optimisation of on-edge learning on low-cost small-footprint embedded CPU platforms, On-edge learning implemented on a custom in-house developed embedded HW platform.

The research goal of EAVISE is applying state-of-the-art computer vision techniques as a solution for industry-specific vision problems. In order to meet stringent execution speed, energy footprint, cost and price requirements, the developed algorithms are implemented and optimized on embedded systems, such as GPUs, DSPs and FPGAs. Application areas include industrial automation, product inspection, traffic monitoring, e-health, agriculture, eye-tracking research, microscopic imaging, surveillance and land surveying.

The research goal of EAVISE is applying state-of-the-art computer vision techniques as a solution for industry-specific vision problems. In order to meet stringent execution speed, energy footprint, cost and price requirements, the developed algorithms are implemented and optimized on embedded systems, such as GPUs, DSPs and FPGAs. Application areas include industrial automation, product inspection, traffic monitoring, e-health, agriculture, eye-tracking research, microscopic imaging, surveillance and land surveying.
  • A basic knowledge in deep learning, image processing and/or embedded implementation techniques

    KU Leuven seeks to foster an environment where all talents can flourish, regardless of gender, age, cultural background, nationality or impairments. If you have any questions relating to accessibility or support, please contact us at diversiteit.HR [at] kuleuven[.]be.

Web

In your application, please refer to myScience.be
and reference  JobID 1862.