In Parkinson's disease, precise analysis of gait disorders remains essential for the diagnostic or the evaluation of treatments. During a gait analysis session, a series of successive dynamic gait trials are recorded and data involves a set of continuous curves for each patient. An important aspect of such data is the infinite dimension of the space data belong. Therefore, classical multivariate statistical analysis are inadequate. Recent methods known as functional data analysis allow to deal with this kind of data. In this paper, we present a functional data analysis approach for solving two problems encountered in clinical practice: (1) for a given patient, assessing the reliability of the gait curves corresponding to the different trials (2) performing intra individual curves comparisons for assessing the effect of a therapy. In a first step, each discretized curve was interpolated using cubic B-splines bases in order to ensure the continuous character of data. A cluster analysis was performed on the smoothed curves to assess the reliability and to identify a subset of representative curves for a given patient. Intra individual curves comparisons were carried out in the following way: (1) functional principal component analysis was performed to describe the temporal structure of data and to derive a finite number of reliable principal components. (2) These principal components were used in a linear discriminant analysis to point out the differences between the curves. This procedure was applied to compare the gait curves of 12 parkinsonian patients under 4 therapeutic conditions. This study allowed us to develop objective criteria for measuring the improvements in a subject's gait and comparing the effect of different treatments. The methods presented in this paper could be used in other medical domains when data consist in continuous curves.