{[ promptMessage ]}

Bookmark it

{[ promptMessage ]}

Table 1 shows the results of classi cation for 39

Info iconThis preview shows page 1. Sign up to view the full content.

View Full Document Right Arrow Icon
This is the end of the preview. Sign up to access the rest of the document.

Unformatted text preview: b, the manner in which the head moves forwards horizontally and then dips slightly in the last frame. Table 1 shows the results of classi cation for 39 test sequences. There were 16 action sequences in the front view FV and 23 sequences in the lateral view LV. Table 2 shows the results of classi cation for the individual action sequences. 6 Conclusion In this paper, we have presented a system that can accurately recognize ten di erent human actions in the frontal or the lateral views. The ten actions are sitting down, standing up, bending over, getting up, walking, hugging, bending sideways, squatting, rising from a squatting position and falling down. The system is not sensitive to variations in the gait of the subject or the height or physical characteristics of the person. Our system was able to correctly recognize subjects of varying height and weight. Thus, it has an advantage over systems that use template matching, in which variations in physical dimensions can produce erroneous results. Further, by modeling the system on the difference in co-ordinates of the head, we do not need to construct three-dimensional models of the subject as a prerequisite to recognition, which is a separate problem in itself. Our system does, however, have its limitations. So far we have used hand segmentation to isolate the head. Before we can consider recognition in real time, we need to be able to automatically detect and segment the head. Further, our system has had only a limited number of trials. We need to test it on a larger number of sequences to ensure its robustness. Also, thus far it is able to recognize only one action in a sequence. If a person enters a scene and then sits down, it is unable to identify both the walking and the sitting actions. We would like to be able to recognize sequences in which Test Correct Incorrect  sequences Classi cation Classi cation success FV LV Total FV LV Total FV LV Total FV LV Total 16 23 39 14 17 31 26 8 87.5 73.91 79.76 Table 1: Results of Classi cation Type of Total Correctly  Sequence Number Classi ed Success Standing 4 3 75 Sitting 5 4 80 Bending down 4 4 100 Getting up 5 3 60 Walking 3 1 33 Type of Total Correctly  Sequence Number Classi ed Success Squatting 4 4 100 Rising 4 4 100 Hugging 2 1 50 Falling 4 3 75 Bending sideways 4 4 100 Table 2: Classi cation of the individual action sequences several actions are concatenated. We intend to work on these problems in the next phase of our implementation and expand our system to be able to identify more complex actions and recognize sequences involving combinations of actions. However, we believe that our system provides a starting point for more complex action recognition. We have, towards this end, also experimented with trying to recognize two actions in a single sequence. The results seem to be promising, however more work is called for before we can present any results. Acknowledgements We would like to thank Ms. Debi Paxton for her generous help in editi...
View Full Document

{[ snackBarMessage ]}