More and more applications that might benefit from a inference algorithms are popping up. For example people counting, seat occupancy detection, track anomalies, bearing wearout,...
On the other hand, bringing all the sensor data to a single node in the train and to a cloud takes a lot of power and bandwidth resources. So sharing the processing load over all sensor nodes (edge computing) has a better power efficiency and because heat dissipation is shared it also improves reliability.
Today, more and more actors come with solutions to run neural networks on a microcontroller based devices :
STM32 API for ST MCUs
Google Edge Computing (edge TPU)
TensorFlow Lite for MCUs
Renesas MCUse-AI translator for TensorFlow or Caffe
e-IQ API & NXP Vision AI for i.MX RT or 7
This thesis will be two-fold:
Create a machine vision model
To train a machine vision model for robust inference of train seat occupancy status detection, a somewhat large variety of pictures with different ambient conditions & people/luggage/seat color will be needed (100-1000 pictures).
Both students will perform a supervised machine learning with existing Open Source Software (TensorFlow or other to be selected) in a format that can be translated later to the selected MCU platform.
The data needs to be tagged for 3 states (for each of the pair of seats) depending on seat state (empty, used by people or luggage). Kids should also be differentiated from luggage.
Evaluate the model
The trained model will then be translated to run on the selected MCU platform development kit and its accuracy will be evaluated depending on :
RAM/CPU per FPS needed for inference (for appropriate MCU selection)
The development kit, selected and bought in advance at the very beginning of the thesis, should support inference on it's MCU with existing translating tools and have a SD card port for inference data input and be close to the final needs (low power & low cost MCU)
Nature of the work
Level: Academic Master, PhD
Specialty: AI / Machine Learning, Electronics / Hardware, Embedded Software
Type of work: Research: 33%, Implem.: 33%, Experim.: 33%
Type of activities: Design, Experimenting, Implementation, Literature study, Measurements, Programming