Abstract : With the fast increase in computing power, numerical simulations of physical phenomena can nowadays rely on up to billions of elements. To extract relevant information in the huge resulting data sets, engineers need visualization tools permitting an interactive exploration and analysis of the computed fields. The goal of this thesis is to improve the visualizations performed by engineers by taking into account the characteristics of the human visual perception, with a particular focus on the perception of space and volume during the visualization of dense 3D data. Firstly, three psychophysics experiments have shown that direct volume rendering, a technique relying on the ordered accumulation of transparencies, provide very ambiguous cues to depth. This is particularly true for static presentations, while the addition of motion and exaggerated perspective cues help to solve part of these difficulties. Then, two algorithms have been developed to improve depth perception during the visualization of complex 3D structures. They have been implemented on the GPU, to achieve interactive renderings independently of the geometric nature of the analysed data. EyeDome Lighting is a new non-photorealistic shading technique that relies on the projected depth image of the scene. This algorithm enhances the perception of shapes and relative depths in complex 3D scenes. Also, a new fast view-dependent cutaway technique has been implemented, which permits to access otherwise occluded objects while providing cues to understand the structure in depth of masking objects.