Eindhoven University of Technology
EDL P16-25 P2: Deep Learning for 3D Reconstruction, Matching & Recognition (RMR)
Data-efficient deep learning for surveillance applications
In many surveillance applications, training data is limited and collecting sufficiently varied data is challenging. During my PhD research I aim to reduce the impact of this problem in multiple ways and will try to apply the results of my research to real-world surveillance problems wherever possible. So far the topics in which I have applied these techniques are: – Change detection: Using both a current and an old frame of a scene with a Siamese network to vastly reduce the complexity of detecting new objects in the scene, thus reducing the required training data. This research has been succesfully applied in a project to detect roadside bombs, for the ministery of defense, and for the detection of littering on public roads (architecture shown in image). – Classification: Training a rare-vehicle classifier with limited training data and especially limited labeling effort, applied to the detection and classification of mobile cranes on highways. – Object detection (current research): Improving the robustness of object detectors to weakly represented or even unseen domains, such as unseen weather conditions, lighting, or scene types. More briefly: domain adaptation and domain generalization. – Explainable deep learning (future research): Explainability is required for safety-critical applications, for example when creating deep learning solutions for the Ministery of Defense. When training with limited data, overfitting on the training set is a concern. Explainability ensures awareness of situations where the network will work to spec, and when it is likely to underperform.
|Real-time small-object detection from ground vehicles using a siamese convolutional neural network. Sander R. Klomp, Dennis W.J.. van de Wouw, Peer H.N. de With, Journal of Imaging Science and Technology.Open access-gold. DOI:10.2352/j.ImagingSci.Technol.2019.63.6.060402. |
ECDNet: Efficient siamese convolutional network for real-time smallobject change detection from ground vehicles. Klomp, S., van de Wouw, D. & de With, P. Electronic Imaging: Intelligent Robotics and IndustrialApplications using Computer Vision 2019; Conference proceedings. Openaccess-gold. DOI:10.2352/ISSN.2470-1173.2019.7.IRIACV-458
Rare-class extraction using cascaded pretrained networks applied to crane classification. Klomp, S. R., Brouwers, G., Wijnhoven, R. G. J. & de With, P. H. N. Electronic Imaging, Intelligent Robotics and Industrial Applications using Computer Vision 2020; Conference proceedings. Open access-gold. https://doi.org/10.2352/ISSN.2470-1173.2020.6.IRIACV-049