A ffects carry important information in human communication and decision making, and their use in technology have grown in the past years. Particularly, emotions have a strong e ect on physiology, which can be assessed by biomedical signals. This signals have the advantage that can be recorded continuously, but also can become intrusive. The present work introduce an emotion recognition scheme based only in photoplethysmography, aimed to lower invasiveness. The feature extraction method was developed for a realistic real-time context. Furthermore, a feature normalization procedure was proposed to reduce the daily variability. For classi cation, two well-known models were compared. The proposed algorithms were tested on a public database, which consist of 8 emotions expressed continuously by a single subject along diff erent days. Recognition tasks were performed for several number of emotional categories and groupings. Preliminary results shows a promising performance with up to 3 emotion categories. Moreover, the recognition of arousal and emotional events was improved for larger emotion sets.