Project Description
The PhD student within this project will be trained in state-of-the-art computer vision algorithms based on deep learning and will study how to adapt and specialize them for real-time analysis of videos and sensor data collected by dashcams (camera on vehicle) installed on VERIZON customer vehicles. The objective of this project will be the development and field testing of algorithms to compute high level semantic understanding of road scenes suitable for resource-constrained platforms, like the Qualcomm Snapdragon 888, Ambarella CV25, or Raspberry Pi 3 by using frameworks like Tensorflow Lite, CVFlow or PyTorch Live. Example applications of interest are danger anticipation, driver drowsiness detection and surveillance, while example tasks to be optimized are detection of road objects and their position in space or sensor fusion of video, inertial measurements and GPS. ESR11 will propose simplifications of existing models to make them meet the real-time and low power constraints on embedded devices and will test their accuracy on publicly available datasets as well as real customer data. Collaboration and a stay in Sorbonne will provide training on heterogeneous computing architectures and optimal resource allocation, which will then be applied to the domain of this project. During a stay and collaboration at the University of Manchester, this PhD student will use simulated datasets within the Open Data Project to deliver a starting point for deep learning techniques in HEP triggers, where the initial testing ground will be using energy deposits left in the detector by hadronic jets as images. This work will also have connections with the project of the PhD student based in Helsinki for identification of jets from heavy quarks.
Host country: Italy
Host beneficiary: Verizon Connect Italy, Florence
PhD-awarding institution: University of Bologna
Experiment affiliation: Industry-based / inter-experiment
Planned collaborations and secondments: Sorbonne, University of Manchester