Deep Learning in computer vision, and in particular for Action Detection, is an effective solution for studying human behaviors of large population, and could be applied to children with autism. It allows capturing, in a non-intrusive and continuous way over time, behavioral patterns. Action detection from live video streams is an important task for monitoring patients, building robots for assisted living and other healthcare applications. Although several approaches, including Deep Convolutional Neural Networks (CNNs), have significantly improved performance on action classification, they still struggle to achieve precise spatio-temporal action localization in untrimmed video streams.
The PhD candidate involved in this project will design novel algorithms for detecting actions, taking advantage of the latest research in Deep Learning. These algorithms will be validated on various international video benchmarks and on a new video database on autism spectrum disorders (ASD) and be published in most prestigious conferences (e.g. CVPR). The early detection of ASD is a crucial issue because it makes it possible to set up intensive and early appropriate care management when certain developmental processes can still be modified.