Virtual Reality (VR) has nowadays become a very useful tool for therapists in the treatment of phobias. Indeed, it allows the simulation of scenarios which are difficult to reproduce in real life. It also allows for a situation to be repeated as much as one wants. Moreover, it allows for a complete control over the situation. The simulation can be stopped if the patient cannot handle it. It can also be tweaked for gradual exposure. Virtual Reality Exposure Therapy (VRET) has proven to be efficient in the context of phobias such as acrophobia or the fear of flying. Social phobia, however, are much harder to deal with. Indeed, as humans, we are experts in human representations and behaviors; it makes it much harder to obtain credible and immersive environments. In this thesis, we describe a set of tools and applications which we have developed to be used in VRET of social phobia and agoraphobia with crowds. We first describe how we create different scenarios for VRET of social phobia. We then expose the application we have developed which allows for elaborate interactions between a user and virtual characters. In particular, we have designed and implemented a software which allows for virtual characters to change behavior depending on the user's eye contact behavior. It allows them to seem interested when being looked at and distracted when not. We then describe the model we have implemented to simulate gaze attention behaviors for crowds of virtual characters. This consists of a method that automatically detects where and when each virtual character in a crowd should look. Secondly, it consists of a dedicated gaze Inverse Kinematics (IK) solver in order for the virtual characters to satisfy the constraints defined by the automatically detected points to be looked at. This allows for the characters to perform the looking motion in a natural and human like way. We then describe the architecture we have developed to combine the work we have done in the domain of social phobia and this model of attention behaviors for crowd characters. We thus use our model of looking behaviors to allow for crowd characters to look at each other. We also use eye-tracking and optical motion capture to determine where a user is looking in a CAVE environment. The virtual characters then respond by either looking at the user, looking at what the user is looking at, or looking at other characters in the crowd. We thus obtain an immersive and interactive environment for VRET in the domain of agoraphobia with crowds. The third part of this thesis describes various experiments we have conducted in order to validate our applications. Our first study consists of using VR in a head-mounted display (HMD) for the treatment of social phobia. In this study, we also use eye-tracking in order to analyze eye contact avoidance behaviors before and after therapy. We then discuss the use of eye-tracking as a tool to help assess and diagnose social phobia. Since eye contact avoidance behaviors are frequent in people suffering from such phobias, eye-tracking can certainly be a helpful tool. We describe an experiment in which we tested eye-tracking as a diagnosis and assessment tool on a phobic population and on a control group. We also describe an experiment to evaluate the potential of our proposed interaction loop in the context of social phobia. Finally, we describe the experiment we have conducted to evaluate our application in the context of agoraphobia with crowds.