main publications

 [back to top]



[main papers] [journals] [conferences]

My current research addresses several topics of situated vision, that is, specific aspects of vision when the observer is doing a specific task: locomotion (driving, walking, etc.). The main medium, in order to study vision, is images; this is why a number of contribution belong to the computer vision and computer graphics academic fields. A psychological approach of visual perception by drivers and pedestrians is also proposed.

visual perception

visual attention on the road. The allocation of visual attention was investigated at a crossroads, looking for information about the ongoing cognitive processes in eye movement data (fixations and saccades). In the TR-F paper, a leftward bias is demonstrated in a simulated driving task.
 road lighting: ergonomic approach. Starting from current practice in road lighting , we have emphasized several limits of road lighting design based on the Visibility Level (Adrian 1989, LRT). These limits are due to the fact that the road lighting performance is related to visual performance in laboratory conditions (psychophysics, e.g. Blackwell 1946, JOSA), instead of situated performance (such as driving performance).
situated vision: pedestrians and drivers. One key issue in situated vision is the change in visual performance depending on the ongoing task and environmental factors. We give evidence of the difference between the driver, passenger and pedestrian conditions in a very simple experimental framwork, in the VIV paper. In RTS, we conducted a review on pedestrian's models (for simulation purpose), and focused on some weaknesses of perception and anticipation in current models. In the AAP paper, a methodology is proposed for assessing the visual impact of the rainfall on the driver's visual performance.
driving simulators: ecological validity. Using computer graphics images in driving simulators, it is not straightforward that the visual behaviour will be similar to what would happen on the road. This was tested, comparing computer graphics images and video images of the same urban environment. Differences were found when analyzing the oculometric patterns. photometry

photometry of road signals. The recent development of LED-based signalling calls for photometric recommendations, in order both to save energy and avoid glare.
 road lighting: photometric approach. The Visibility Level (VL) proposed by Adrian (1989, LRT) is a reference model for road lighting practitioners, as an attempt to design road lighting based on a reference visual performance (target detection). In my CIE paper, I sho that the definition of a reference visual performance is a tough compromize. In the LRT paper, we focus on one weakness of the psychophysical data at the basis of the VL: in the real world, a target on the road does not appear on a uniform background.
  • Brémond, Dumont, Ledoux & Mayeur (2011). Lighting Research and Technology, 42(1) pdf
  • Brémond (2007), in Proc. 26th session of the CIE (vol. 2), D4 pp. 100-103, n° PO-414, Beijing (China), july 2007 pdf
computer vision

computer vision : image quality. Vision models allow to compare two images, in the sense of a given visual function. Such model are of great interest when images are displayed to observers, in order to achieve psycho-visual experiments (including driving simulation). We have discussed some of these models, and focused on the use of visual attention models to compare image sequences (REFIG paper). In the ECCV paper, we found that usual saliency models are not suited for High Dynamic Range (HDR) images. In the JEI paper, we show that unfortunately, image metrics based on visual performance (VDP), visual appearance (CIECAM) and visual attention (Itti & Koch) are only weakly correlated.
 computer vision : visual saliency of road signs. Bottom-up visual attention models, made popular by Itti et al. in the field of Computer Science, happen to fail to predict visual performance when the observer is engaged in action, which occurs in a driving situation. We have considered, instead, that in situated vision, there is a need for a top-down model of visual saliency. This idea was applied to the detection of road signs, using a Support Vector Machine to learn the task, then to compute the estimated visual saliency based on a priori knowledge. 

computer vision : human visibility. The princeps paper of Marr and Hildreth (1980) is revisited and a new model is proposed for edge detection. The main issue here is not edge localization but the estimation of the edge visibility, from the point of view of a human observer. The proposed algorithm takes the Contrast Sensitivity of the human eye into account. An age-dependent CSF model was also proposed, as well as an ADAS based on this model.
 image processing : geodesic propagation. Geodesic propagation computes the shortest paths between two sets. This idea was applied, first, to compute the diffusion of fluids in porous materials (J. Microscopy), then to compute an image subset where some post-processing (estimation of fog physical properties) could be done with confidence.
computer graphics
computer graphics : tone mapping for road visibility applications. Rendering High Dynamic Range (HDR) images in physical units is not enough to display realistic images. The limitations of current display devices in terms of luminance dynamics makes it necessary to use Tone Mapping Operators (TMO) before displaying an image. In ACM TAP, we proposed that for some applications, such as road visibility studies, the TMO design and evaluation should consider visual performance rather than visual appearance (judgment). In the Visual Computer paper, we also designed a rendering pipeline for real time applications, including HDR rendering, glare simulation and tone mapping.
computer graphics : rendering. Attempt to include a physical computation of road lighting in virtual environments are presented here. The BLPC paper proposes an evaluation of the results based on luminance differences, while the Laval Virtual paper proposed a pre-processing of road lighting, embedded in an interactive virtual environment.  

complex systems

improving the perception function in a multi-agent system. One approach for traffic simulations is to use Multi-Agent Systems (MAS), where distributed entities (e.g. drivers or pedestrians agents) take decisions on the basis of their local perception and system knowledge. We contribute to improve the perception function in such simulation models, on the basis of the state of the art in visual attention.

cellular automata. Lattice gas automata have been proposed by Frish et al (1986) to simulate hydrodynamic equations. One nice thing there is that any bounding condition may be taken into account. This allowed to simulate the filtration of cast-iron in conditions where the bounding conditions change with time.
    • Bremond &  Jeulin (1994), Revue de Métallurgie et Matériaux. 
    • Bremond & Jeulin (1994), in Proc. International Symposium on Mathematical Morphology, Kluwer press, Serra and Soille Editors.
Complexity. From a comparison between transportation systems and biological systems, I suggest that one reason which may explain a number of formal links between the two domains is their complexity, meaning that understanding the system's interactions is more important than understanding its components. Then, I discuss the benefits one could take from system biology approach in the field of transportation.