Section: Research Program

Perception of People, Activities and Emotions

Machine perception is fundamental for situated behavior. Work in this area concerns construction of perceptual components using computer vision, acoustic perception, accelerometers and other embedded sensors. These include low-cost accelerometers [Bao 04], gyroscopic sensors and magnetometers, vibration sensors, electromagnetic spectrum and signal strength (wifi, bluetooth, GSM), infrared presence detectors, and bolometric imagers, as well as microphones and cameras. With electrical usage monitoring, every power switch can be used as a sensor [Fogarty 06], [Coutaz 16]. We have developed perceptual components for integrated vision systems that combine a low-cost imaging sensors with on-board image processing and wireless communications in a small, low-cost package. Such devices are increasingly available, with the enabling manufacturing technologies driven by the market for integrated imaging sensors on mobile devices. Such technology enables the use of embedded computer vision as a practical sensor for smart objects.

Research challenges addressed in this area include development of practical techniques that can be deployed on smart objects for perception of people and their activities in real world environments, integration and fusion of information from a variety of sensor modalities with different response times and levels of abstraction, and perception of human attention, engagement, and emotion using visual and acoustic sensors.

Work in this research area will focus on three specific Research Actions

Multi-modal perception and modelling of activities

The objective of this research action is to develop techniques for observing and scripting activities for common household tasks such as cooking and cleaning. An important part of this project involves acquiring annotated multi-modal datasets of activity using an extensive suite of visual, acoustic and other sensors. We are interested in real-time on-line techniques that capture and model full body movements, head motion and manipulation actions as 3D articulated motion sequences decorated with semantic labels for individual actions and activities with multiple RGB and RGB-D cameras.

We have explored the integration of 3D articulated models with appearance based recognition approaches and statistical learning for modeling behaviors. Such techniques provide an important enabling technology for context aware services in smart environments [Coutaz 05], [Crowley 15], investigated by Pervasive Interaction team, as well as research on automatic cinematography and film editing investigated by the Imagine team [Gandhi 13] [Gandhi 14] [Ronfard 14] [Galvane 15]. An important challenge is to determine which techniques are most appropriate for detecting, modeling and recognizing a large vocabulary of actions and activities under different observational conditions.

We explored representations of behavior that encodes both temporal-spatial structure and motion at multiple levels of abstraction. We will further propose parameters to encode temporal constraints between actions in the activity classification model using a combination of higher-level action grammars [Pirsiavash 14] and episodic reasoning [Santofimia 14] [Edwards 14].

We have adapted this work to construct narrative descriptions of cooking activities from ego-centric vision, in cooperation with Remi Ronfard of the Imagine Team of Inria.

Perception with low-cost integrated sensors

In this research action, we will continue work on low-cost integrated sensors using visible light, infrared, and acoustic perception. We will continue development of integrated visual sensors that combine micro-cameras and embedded image processing for detecting and recognizing objects in storage areas. We will combine visual and acoustic sensors to monitor activity at work-surfaces. Low cost real-time image analysis procedures will be designed that acquire and process images directly as they are acquired by the sensor.

Bolometric image sensors measure the Far Infrared emissions of surfaces in order to provide an image in which each pixel is an estimate of surface temperature. Within the European MIRTIC project, Grenoble startup, ULIS has created a relatively low-cost Bolometric image sensor (Retina) that provides small images of 80 by 80 pixels taken from the Far-infrared spectrum. Each pixel provides an estimate of surface temperature. Working with Schneider Electric, engineers in the Pervasive Interaction team had developed a small, integrated sensor that combines the MIRTIC Bolometric imager with a microprocessor for on-board image processing. The package has been equipped with a fish-eye lens so that an overhead sensor mounted at a height of 3 meters has a field of view of approximately 5 by 5 meters. Real-time algorithms have been demonstrated for detecting, tracking and counting people, estimating their trajectories and work areas, and estimating posture.

Many of the applications scenarios for Bolometric sensors proposed by Schneider Electric assume a scene model that assigns pixels to surfaces of the floor, walls, windows, desks or other items of furniture. The high cost of providing such models for each installation of the sensor would prohibit most practical applications. We have recently developed a novel automatic calibration algorithm that determines the nature of the surface under each pixel of the sensor.

Work in this area will continue to develop low-cost real time infrared image sensing, as well as explore combinations of far-infrared images with RGB and RGBD images.

Observing and Modelling Competence and Awareness from Eye-gaze and Emotion

Humans display awareness and emotions through a variety of non-verbal channels. It is increasingly possible to record and interpret such information with available technology. Publicly available software can be used to efficiently detect and track face orientation using web cameras. Concentration can be inferred from changes in pupil size [Kahneman 66]. Observation of Facial Action Units [Ekman 71] can be used to detect both sustained and instantaneous (micro-expressions) displays of valence and excitation. Heart rate can be measured from the Blood Volume Pulse as observed from facial skin color [Poh 11]. Body posture and gesture can be obtained from low-cost RGB sensors with depth information (RGB+D) [Shotton 13] or directly from images using detectors learned using deep learning [Ramakrishna 14]. Awareness and attention can be inferred from eye-gaze (scan path) and fixation using eye-tracking glasses as well as remote eye tracking devices [Holmqvist 11]. Such recordings can be used to reveal awareness of the current situation and to predict ability to respond effectively to opportunities and threats.

This work is supported by the ANR project CEEGE in cooperation with the department of NeuroCognition of Univ. Bielefeld. Work in this area includes the Doctoral research of Thomas Guntz to be defended in 2019.


[Bao 04] L. Bao, and S. S. Intille. "Activity recognition from user-annotated acceleration data.", IEEE Pervasive computing. Springer Berlin Heidelberg, pp. 1-17, 2004.

[Fogarty 06] J. Fogarty, C. Au and S. E. Hudson. "Sensing from the basement: a feasibility study of unobtrusive and low-cost home activity recognition." In Proceedings of the 19th annual ACM symposium on User interface software and technology, UIST 2006, pp. 91-100. ACM, 2006.

[Coutaz 16] J. Coutaz and J.L. Crowley, A First-Person Experience with End-User Development for Smart Homes. IEEE Pervasive Computing, 15(2), pp. 26-39, 2016.

[Coutaz 05] J. Coutaz, J.L. Crowley, S. Dobson, D. Garlan, "Context is key", Communications of the ACM, 48 (3), pp. 49-53, 2005.

[Crowley 15] J. L. Crowley and J. Coutaz, "An Ecological View of Smart Home Technologies", 2015 European Conference on Ambient Intelligence, Athens, Greece, Nov. 2015.

[Gandhi 13] Vineet Gandhi, Remi Ronfard. "Detecting and Naming Actors in Movies using Generative Appearance Models", Computer Vision and Pattern Recognition, 2013.

[Gandhi 14] Vineet Gandhi, Rémi Ronfard, Michael Gleicher. "Multi-Clip Video Editing from a Single Viewpoint", European Conference on Visual Media Production, 2014

[Ronfard 14] R. Ronfard, N. Szilas. "Where story and media meet: computer generation of narrative discourse". Computational Models of Narrative, 2014.

[Galvane 15] Quentin Galvane, Rémi Ronfard, Christophe Lino, Marc Christie. "Continuity Editing for 3D Animation". AAAI Conference on Artificial Intelligence, Jan 2015.

[Pirsiavash 14] Hamed Pirsiavash , Deva Ramanan, "Parsing Videos of Actions with Segmental Grammars", Computer Vision and Pattern Recognition, pp. 612-619, 2014.

[Edwards 14] C. Edwards. 2014, "Decoding the language of human movement". Commun. ACM 57, 12, pp. 12-14, November 2014.

[Kahneman 66] D. Kahneman and J. Beatty, Pupil diameter and load on memory. Science, 154(3756), pp. 1583-1585, 1966.

[Ekman 71] P. Ekman and W.V. Friesen, Constants across cultures in the face and emotion. Journal of Personality and Social Psychology, 17(2), 124, 1971.

[Poh 11] M. Z. Poh, D. J. McDuff, and R. W. Picard, Advancements in noncontact, multiparameter physiological measurements using a webcam. IEEE Trans. Biomed. Eng., 58, pp. 7–11, 2011.

[Shotton 13] J. Shotton, T. Sharp, A. Kipman, A. Fitzgibbon, M. Finocchio, A. Blake, M. Cook, and R. Moore, Real-time human pose recognition in parts from single depth images. Commun. ACM, 56, pp. 116–124, 2013.

[Ramakrishna 14] V. Ramakrishna, D. Munoz, M. Hebert, J. A. Bagnell and Y. Sheikh, Pose machines: Articulated pose estimation via inference machines. In European Conference on Computer Vision (ECCV 2016), pp. 33-47, Springer, 2014.

[Cao 17] Z. Cao, T. Simon, S. E. Wei and Y. Sheikh, Realtime Multi-person 2D Pose Estimation Using Part Affinity Fields. In 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR 2017), IEEE Press, pp. 1302-1310, July, 2017.

[Holmqvist 11] K. Holmqvist, M. Nyström, R. Andersson, R. Dewhurst, H. Jarodzka, and J. van de Weijer, Eye Tracking: A Comprehensive Guide to Methods and Measures, OUP Oxford: Oxford, UK, 2011.