Table 1. Factors for analyzing empirical studies on activity recognition.

F.latent.infty	Method allows inference in latently infinite state spaces (typically employing a computational action language).
F.plan.synth	Plan synthesis is supported. Otherwise, the approach requires to create plan libraries by explicitly enumeration.
F.duration	Durative actions are supported. (This will significantly increase inference complexity, as the starting time for an action becomes another state variable, which has a large value space. See Appendix S4)
F.action.sel	Explicit mechanisms for modeling human action selection based on opportunistic and/or goal driven features are supported.
F.probability	Method provides (an approximation of) the posterior probability distribution over states (or actions, depending on the mechanism). This is a prerequisite for selecting assistive interventions using decision-theoretic methods (i. e., that aim at maximizing the expected utility).
F.struct.state	The state maintained by inference provides a structured representation of the environment state. This allows the formulation of state predicates and the dynamic synthesis of contingency plans. (Otherwise the state typically represents the action currently executed.)
F.non.monoton	Non-monotonous action sequences are considered, that – temporarily – may increase goal distance. (This affects the number of plans that need to be considered. Methods using explicit plan enumeration usually avoid non-monotonicity.)
F.complexity	Filter step complexity (computational complexity for the filtering step from to ). If greater than , for instance , then online filtering is essentially intractable.
Method	Type of inference method used.
Scenario	Scenarios considered in experimental tasks.
N.states	Number of states considered. (See text for further explanation.)
N.plan.length	Lengths of plans considered in study.
N.classes	Number of classes in classification target used for performance evaluation.
N.subjects	Number of subjects participating in trials (or “sim” in case evaluation is based on simulated observations).
M.accuracy	Accuracy is provided as performance measure.
M.conf.based	Other quantities based on confusion matrices (true–positive rate, precision, etc.) are provided as performance measures.