Humans are prepared to understand others’ emotional expressions from subtle body movements and facial expressions, changing the way they communicate in function of those interactions. To transit to a “factual” human-machine collaboration, where the machine delivers relevant information and functionalities in a timely and appropriate manner, machines need to be equipped with such capabilities. This thesis aims to develop a body affective classifier (emotions and sentiments), that can adapt (be discriminative) on-the-fly to different peculiarities of key variables: gender, age, culture, and environment. More specifically, it will be proposed to (i) compile a dataset tailored for body-based affective analysis that encompasses the mentioned key variables. Develop classification models for human body (ii) language, and for human body (iii) emotion and sentiment. (iv) Construct a body affective classifier framework, that incorporates and combines the aforementioned models, contextualized by the key variables. Finally, (v) evaluate the performance using real-time data.