Table 1.
Factors for analyzing empirical studies on activity recognition.
Table 2.
Quantitative and qualitative properties of selected studies on activity and intention recognition.
Figure 1.
Instrumentation and trial setting.
Left: Instrumentation of participants (red points indicate IMU positions). Right: Conceptual spatial layout (view from above) and domain objects of trial setting.
Figure 2.
Boxes represent tuples of random variables. An arc starting/ending at a box ( = a tuple) represents a set of arcs connected to the tuple's components. Nodes with double outline signify observed random variables.
Table 3.
Factors and levels for experimental configurations.
Figure 3.
Baseline classifier accuracies per subject.
Different numbers of principal components in the observation model have been used. An accuracy of .2 (solid gray line) is achieved by selecting the action class with highest prior probability.
Figure 4.
Boxplots of Accuracy vs. CSSM configuration.
= continuous (
) resp. discrete (
) duration model.
=
(restricted),
(complete),
(scripted) distance computation. Details for the O21s configurations marked by triangles are shown in Fig. 5. The orange triangle marks the CMf configuration used in testing
.
Figure 5.
Accuracy comparison of selected CSSM configurations to baseline (HMM), by subject and filter method.
c/d = use of (continuous) or
(discrete) duration model.
/
/
= use of
(complete),
(restricted), or
(script) goal distance model. Observation model O21s, distance weight L1. (Subjects sorted by median performance in all configuration.)
Figure 6.
Per-class performance measures.
Detailed accuracies for the configuration (OS21s, , L1,
).
Table 4.
Cohen's and overall accuracies for selected configurations.
Figure 7.
Color: first factor, X-axis: second factor. X-position “Mean”: main effects of first factor. Grey points: main effects of second factor. Error bars give the 95% confidence intervals due to between subject variance. Effect comparisons are based on within subject differences. Confidence intervals for effects therefore are much smaller than implied by these error bars.
Table 5.
Median SpU and XpS values and ratios, across all runs.
Figure 8.
Observed normalized remaining time to goal (RT) versus normalized goal distances.
Data based on states traversed in observed sequences. These states have been entered at different times by different subjects; due to alternative paths also multiple discoveries by the same subject were possible. In total, 960 different discovery events were recorded (blue points in plot). For each discovery event, we plot normalized goal distance of discovered state (X axis) versus remaining execution time (Y axis). The three plots show the values for different goal distance computation methods/heuristics. The red line is the linear model predicting (RT) from heuristics using
. Table 6 gives properties of linear models.
Table 6.
Properties of linear models in Fig. 8 (for all ,
).
Figure 9.
Estimating the probability of state properties.
Sample values for configuration , plotted for each subject individually.
Table 7.
Median values and IQR (Q1, Q3) for Accuracy and JSD found for predicate estimation based on configurations .