Resumen de Building planning action models using activity recognition

Javier Ortiz Laguna

• Activity recognition is receiving a special attention because it can be used in many areas. This field of artificial intelligence has been widely investigated lately for tasks such as following the behavior of people with some kind of cognitive impairment. For instance, elderly people with dementia. The recognition of the activities that these people carry on permits to offer assistance in case they need it while they are performing the activities. Currently, there are many systems capable of recognizing the activities that a user performs in a specific environment. Most of these systems have two problems. First, they recognize states of the activities instead of the entire activity. For instance, instead of recognizing an activity that starts at the moment ????? and ends at ?????, these systems split the time line in fixed-length temporal windows that are classified as belonging to an activity or another. These windows sometimes overlap two activities which makes classifying the activities more difficult. Also this prevents the system from detecting the states of the system before and after each activity. These states are needed to build behavioral models. Second, most of these systems recognize complete high-level activities such as cooking or making tea but they can not recognize the low-level activities that compose the high-level activities. For example, pick-up the fork or switch on the oven. For this reason, most of the systems in the literature can not be used to assist people during the activities since they recognize the activity itself and they can not provide the low-level activities that the user has to execute to complete the high-level activity. For these reasons, in this thesis we have three objectives. The first objective is the development of a new activity recognition algorithm capable of overcoming the problems that the fixed-length temporal windows cause and, also, capable of extracting the states that the system can traverse. The second objective is to automatically generate an automated planning domain able to represent the user behavior using the activity recognition system. In order to do that, we will use the activity recognition system developed in the previous step to recognize the activities and the states of the environment before and after the activities. Once the system is capable of performing this task, the planning domain is generated using that information. Then, the automated planning domain will be used by an automated planner to generate sequences of actions to reach the goal of the user. That way, those sequences will be used to assist users by telling them the next action or actions to accomplish their goals. The third objective is to use automatically generated planning domains for guiding users to accomplish the task they pursue. In addition, we want to check whether the generated plans can be used to recognize the activities alone or to help a sensors system to improve its results. Also, the generated plans will be used to predict the next activities that the user may perform. This way, we will test the planning domains and the plans generated by the planner to check if they are capable of offering information to recognize the activity that the user performs or, at least, offering information for the activity recognition system to improve its results. --------------------------