The feasibility of real life affective detection using physiological signals is usually limited by biosensor noise and artifact. This is challenging in extracting the representative emotion features. In this paper a quasi-homogeneous segmentation algorithm based on Top-Down homogeneous splitting and Bottom-Up Merging using Bhattacharyya distance is proposed to partition the signal and remove artifacts. Furthermore, since physiological responses may also vary within one emotion elicited period, features extracted from segmented segments can better describe recent physiological patterns. In this paper a constraint-based clustering analysis based on estimating best seed of K-means is developed to discover representative emotion-elicited segments at all cross subject partitions which include labeled and unlabelled feature vectors.