Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

This work presents a solution for autonomous vehicles to detect arbitrary moving traffic participants and to precisely determine the motion of the vehicle. The solution is based on three-dimensional images captured with modern range sensors like e.g. high-resolution laser scanners. As result, objects are tracked and a detailed 3D model is built for each object and for the static environment. The performance is demonstrated in challenging urban environments that contain many different objects.

mapping --- self-localization --- object detection --- laser scanner --- tracking

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In several fields of application, multi-gigapixel images must be analysed to gather information and take decision. This analysis is often performed manually, which is a tedious task given the volume of data to process. For instance, in cytology, branch of medical sciences which focuses on study of cells, cytopathologists analyse cell samples microscope slides in order to diagnose diseases such as cancers. Typically, malignancy is assessed by the presence or absence of cells with given characteristics. In geology, climate variations can be analysed by studying the concentration 
of micro-organisms in core samples. The concentration is usually evaluated by smearing the samples onto microscope glass slides and counting those micro-organisms.

In those situations, computer sciences and, especially, machine learning and image processing provide a great alternative to a pure-human approach as they can be used to extract relevant information automatically. Especially, those kind of problems can be expressed as object detection and classification problems.

This thesis presents the elaboration and assessment of a generic framework, extit{SLDC}, for object detection and classification in multi-gigapixel images. Especially, this framework provides implementers with a concise way of formulating problem dependent-components of their algorithm (i.e. segmentation and classification) while it takes care of problem-independent concerns such as parallelization and large image handling. 

The performances of the framework are then assessed on a real-world problem, thyroid nodule malignancy. Especially, a workflow is built to detect malignant cells in thyroid cell samples whole-slides.

Results are promising: the effective processing time for an image containing 8 gigapixels is less than 10 minutes. In order, to further reduce this execution time, some improvements are proposed.

The framework implementation can be found on GitHub: https://github.com/waliens/sldc.

machine learning --- cytomine --- image processing --- cytology --- object detection --- Ingénierie, informatique & technologie > Sciences informatiques

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

In the recent years, object detection models have leveraged deep learning architectures to improve performance in many problems. However, these techniques require
a large amount of high quality labelled data in order to reach their full potential,
and obtaining such data may prove to be an arduous task. In this context, this
work explores the possibility of using entirely synthetically generated and labelled
images to train an object detection model. In particular, we examine which factors
of variations in the synthetic data best transfer to real data. Unsurprisingly, models
trained on synthetic data only perform significantly worse than models trained on
real data. We explore whether the synthetic images can be enhanced using filtering
and generative models, but find the results to be inconclusive. In a setting where
both real and synthetic data are available, we experiment to find out how these
should be combined to improve performance in the real domain. We find that the
synthetic and real datasets should be combined into a single training dataset, and
that the object detection model trained in this fashion significantly outperforms the
model trained on real data only.

Deep learning --- Object detection --- Synthetic data --- Ingénierie, informatique & technologie > Sciences informatiques

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Unmanned Aerial Vehicles (UAVs) equipped with video cameras are a flexible support to ensure civil and military safety and security. In this thesis, a video processing chain is presented for moving object detection in aerial video surveillance. A Track-Before-Detect (TBD) algorithm is applied to detect motion that is independent of the camera motion. Novel robust and fast object detection and segmentation approaches improve the baseline TBD and outperform current state-of-the-art methods.

Automatische Videoüberwachung --- Object Detection --- Independent Motion Detection --- Bewegungsdetektion bei bewegter Kamera --- MultiobjekttrackingAutomatic Video Surveillance --- Multiple Object Tracking --- Track-Before-Detect --- Objektdetektion

Choose an application

• Reference Manager
• EndNote
• RefWorks (Direct export to RefWorks)

Monitoring and census of wild animal populations are among the key elements in nature conservation. The use of UAV (Unmanned Aircraft Vehicle) or light aircraft as aerial image acquisition system is a more suitable and cheaper alternative to traditional census methods. However, the manual localization and identification of species within these images can quickly become time-consuming and complex. Detection algorithms, based on Convolutional Neural Networks (CNNs), have shown a good capacity for animal detection based on aerial images. Nevertheless, most of the work is focused on binary detection cases. The main objective of this study is to compare the performances of three recent detection algorithms to detect and identify seven African mammals (Alcelaphinae, buffaloes, elephants, hippopotamuses, kobs, warthogs, and waterbucks) based on high-resolution aerial images of various African landscapes. To do so, the performances of the multi-class CNNs Faster-RCNN, Libra-RCNN and RetinaNet to detect these seven animal species in aerial images from four different datasets were evaluated. The algorithms tested were able to detect 91.8 to 95.5% of the animals, with a ratio of 2.8 to 13.8 false positives per true positive. All three algorithms have generally met the challenges that aerial images can present in animal detection. Libra-RCNN showed the best mean Average Precision (mAP=0.68), the lowest degree of inter-species confusion and a lower sensitivity to variation in prediction thresholding. Hippopotamuses and warthogs were the most difficult species to identify and detect (low precision) by all three algorithms. However, these algorithms present themselves as good future semi-automatic detection tools and each has interesting specificity for a potential practical implementation. La surveillance et le recensement des populations animales sauvages font partie des éléments clefs dans la conservation de la nature. L'utilisation de drones ou d'avions légers comme système d'acquisition d'images aériennes se présente comme une alternative plus adaptée et moins chère que les méthodes traditionnelles d'inventaire. Cependant, la localisation et l'identification manuelles des espèces au sein de ces images peuvent rapidement devenir chronophage et complexe. Des algorithmes de détection, basés sur des réseaux de neurones convolutifs (RNCs), ont montré une bonne capacité à la détection animale sur base d'images aériennes. Néanmoins, la majorité des travaux ont porté sur des cas de détection binaire. L'objectif principal de cette étude est de comparer les performances de trois récents algorithmes à détecter et identifier sept mammifères africains (Alcelaphinae, buffles, éléphants, hippopotames, cobs, des phacochères et cobs à croissant) sur base d'images aériennes à haute résolution de paysages africains variés. Pour ce faire, les performances des RNCs multi-classe Faster-RCNN, Libra-RCNN et RetinaNet à détecter ces sept espèces animales au sein d'images aériennes provenant de quatre jeux de données différents, ont été évaluées. Les algorithmes testés ont réussi à détecter 91,8 à 95,5% des animaux, avec un rapport allant de 2,8 à 13,8 faux positifs par vrai positif. Les trois algorithmes ont globalement relevé les défis que peuvent présenter les images aériennes en détection animale. Libra-RCNN est celui qui a montré la meilleure mean Average Precision (mAP=0.68), le moins de confusion entre les espèces et une moins forte sensibilité à la variation du seuillage des prédictions. Les hippopotames et les phacochères ont été les espèces les plus difficiles à identifier et détecter (faible précision) par les trois algorithmes. Toutefois, ces algorithmes se présentent comme de bons futurs outils de détection semi-automatique et possèdent chacun une spécificité intéressante pour une potentielle implémentation pratique.

object detection --- wildlife monitoring --- unmanned aircraft vehicle --- aerial images --- African mammals --- convolutional neural networks --- computer vision --- deep learning --- détection d'objets --- surveillance de la faune sauvage --- drone --- images aériennes --- mammifères africains --- réseaux de neurones convolutifs --- vision par ordinateur --- apprentissage profond --- Ingénierie, informatique & technologie > Multidisciplinaire, généralités & autres