Recognition performance was evaluated on a database of isolated gesture recordings
made by both the teacher and student using three different matching techniques (L1
norm, L2 norm, and histogram intersection). Overall, both features were very effective
for recognition, with average recognition rates in the range of 90.5% to 99.5%, with
the dance feature showing improved robustness (discounting some errors introduced
by derelict/noisy recordings in gestures G4 and G5). The incorporation of posture transitions
as a descriptor shows a marked boost in recognition performance (across all
matching metrics used) and can be attributed to its detection of temporal ordering
of postures. The bag-of-segments approach to all four descriptors offers flexibility and
generalization across instances of movement recorded from a candidate user: recognition
for which, due to the natural variation of the human when repeating movements
and the sensor noise introduced by the Kinect, can be a challenging task.
Recognition performance was evaluated on a database of isolated gesture recordingsmade by both the teacher and student using three different matching techniques (L1norm, L2 norm, and histogram intersection). Overall, both features were very effectivefor recognition, with average recognition rates in the range of 90.5% to 99.5%, withthe dance feature showing improved robustness (discounting some errors introducedby derelict/noisy recordings in gestures G4 and G5). The incorporation of posture transitionsas a descriptor shows a marked boost in recognition performance (across allmatching metrics used) and can be attributed to its detection of temporal orderingof postures. The bag-of-segments approach to all four descriptors offers flexibility andgeneralization across instances of movement recorded from a candidate user: recognitionfor which, due to the natural variation of the human when repeating movementsand the sensor noise introduced by the Kinect, can be a challenging task.
การแปล กรุณารอสักครู่..