Analyzing Trajectories on Grassmann Manifold for Early Emotion Detection from Depth Videos

Taleb Alashkar (1) , Boulbaba Ben Amor (1) , Stefano Berretti (2) , Mohamed Daoudi (1)

1 LIFL - FOX MIIRE
2 DSI - Dipartimento di Sistemi e Informatica

— This paper proposes a new framework for online detection of spontaneous emotions from low-resolution depth se-quences of the upper part of the body. To face the challenges of this scenario, depth videos are decomposed into subsequences, each modeled as a linear subspace, which in turn is represented as a point on a Grassmann manifold. Modeling the temporal evolution of distances between subsequences of the underlying manifold as a one-dimensional signature, termed Geometric Motion History, permits us to encompass the temporal signature into an early detection framework using Structured Output SVM, thus enabling online emotion detection. Results obtained on the publicly available Cam3D Kinect database validate the proposed solution, also demonstrating that the upper body, instead of the face alone, can improve the performance of emotion detection.

Domaines: Vision par ordinateur et reconnaissance de formes [cs.CV]

Taleb Alashkar, Boulbaba Ben Amor, Stefano Berretti, Mohamed Daoudi. Analyzing Trajectories on Grassmann Manifold for Early Emotion Detection from Depth Videos. IEEE International Conference on Automatic Face and Gesture Recognition, FG 2015, May 2015, Ljubljana, Slovenia. ⟨hal-01109468⟩