Depth camera-based virtual rehabilitation systems are gaining attention in occupational therapy for cerebral palsy patients. When developing such a system, domain-specific exercise recognition is vital. To design such a gesture recognition method, some obstacles need to be overcome: detection of gestures not related to the defined exercise set and recognition of incorrect exercises performed by the patients to compensate for their lack of ability. We propose a framework based on hidden Markov models for the recognition of upper extremity functional exercises. We determine critical compensation mistakes together with restrictions for classifying these mistakes with the help of occupational therapists. We first eliminate undefined gestures by evaluating two models that produce adaptive threshold values. Then we utilize specific negative models based on feature thresholding and train them for each exercise to detect compensation mistakes. We perform various tests using our method in a laboratory environment under the supervision of occupational therapists.