Ir. Thomas VANDAMME

"Assessing Intellectual Property Relevant Similarities In Images Through Algorithmic Decision Systems" (Under supervision of Pr. Olivier Debeir and Pr. Julien Cabay)

This PhD Thesis, funded by an ARC (Action de Recherche Concertée) at ULB, is at the crossroads of Deep-Learning and Intellectual Property. This project aims at assessing Intellectual Property (IP) relevant similarities, with a focus on image recognition technologies. Current Algorithmic decisions systems are developed by private companies for the purposes of IP enforcement (monitoring infringing goods online, filtering out content) and registration by IP Offices.

We aim at exposing the biases of such systems, and propose new, unbiased and open-source algorithms for these decisions systems.

The Masaai project, in collaboration with MintT ("Medical Intelligent Technologies"), and funded by Innoviris, aimed at developing neural network models for fall detections in hospital environment, using depth-sensing (Time-Of-Flight) cameras.

During this project, we developed neural networks for fall detection and person segmentation in depth sequences. We developed an automatic labelization method based on RGB-D cameras and pre-trained RGB person detection neural-network. One of the aspects of the project was to use a 3D engine (Unity) to automatically generate large amounts of labeled synthetic sequences, and evaluate the possibility of using such technique to pre-train the networks.

"Person Detection Using Time-Of-Flight Cameras and Machine Learning" (2019)

Some access-control gates used in public transportation stations reduce thefare evasion by counting the number of people present in an airlock, assessing that every user paid. Automatic Systems developed such a gate equipped with a Time-Of-Flight camera located above the gate, in order to evaluate its occupancy.

This thesis develops a detection algorithm meant to evaluate the occupancyof the gate using a sequence of depth-images, based on an image-segmentation neural network, followed by a labelization. The occupancy of each image is sim-ply passed through a mode to obtain the sequence’s global occupancy estimation.

A dataset acquired in Lille, in real-life settings, was split in order to train and evaluate the neural network, as well as evaluate the general performancesof the system. The neural network, based on a U-Net, achieved an Area Under Curve of the Precision/Recall curve of 0.94 with only 160k parameters, while the whole system achieved a 86% accuracy, with as little as 2.6% of sequences under-evaluated, those where fraud may occur. The training was done using 4770 depth-images, the validation of the segmentation network over 1362 images, and the system performance evaluation on 3317 sequences.