"... In PAGE 13: ...able 4-1. Statistics of ROI detection and tracking result. ..................................82 Table5 -1.... In PAGE 13: ...able 5-1. Ten types of activities for recognition. .................................................91 Table5 -2.... In PAGE 104: ... The motion parameters computed from ROI video output and the associated virtual camera control parameters are used for activity recognition. Our experiments consist of ten activities as shown in Table5 -1. We separate these activities into three groups.... In PAGE 123: ... Table5 -2 summarizes recognition results. 5% step size is used for accuracy computation.... In PAGE 124: ... It can be seen from experiment that in general PMO method outperforms the PCA method. Activity Group1 Group2 Group3 Virtual Camera Control Parameters 100% ---- ---- 6 bases ---- 75% 60% PCA 10 bases ---- 80% 75% 6 bases ---- 90% 80% Optic Flow PMO 10 bases ---- 90% 85% 6 bases ---- 40% 30% PCA 10 bases ---- 45% 40% 6 bases ---- 50% 50% Affine Model PMO 10 bases ---- 60% 50% Table5... ..."

"... In PAGE 3: ... In a last step, the result is communicated to other sensor nodes. Table3 shows the design space of the activity recogni- tion algorithm optimizations. To assess the influences of the different parameters, we have asked 9 test subjects to perform each activity 30 times.... In PAGE 4: ...line context recognition for the kNN classifier at 32 Hz ture information gain must again be related to the cost, or computational complexity of computing it. Table3 lists the evaluated features. The mean, variance, and energy values of the signal of the window have been selected for implemen- tation and have been computed for all of the 3 accelerometer axes and the light sensor.... ..."

"... In PAGE 7: ... Algorithm 2 slightly outperforms Algorithm 1 for this problem because the dataset used in this experiment consisted of teamwork activities performed by four agents where each agent al- ways play a unique role in the team. Table5 illustrates the confusion matrix for teamwork recognition with Algorithm 2 when = 2. In this case the mean accuracy is 92.... ..."

"... In PAGE 3: ... The similarity score between the two event probability sequences is computed using the dynamic time warping algorithm. Table1 summarizes the recognition results using cumulative match scores (CMS) as the performance measure. CMS is com- puted by accumulating recognition rates from rank a29 onwards.... ..."

"... In PAGE 9: ... Finally the bottom most block shows the activity of the user. Table3 summarizes the retrieval results for all three meetings. The models of the speaker identification algo- rithm are trained for every meeting separately using the other two recordings as the training set.... ..."