On-line human action detection using space-time interest points

On-line human action detection using space-time interest points

en

The on-line human action detection is an important task in human-machine interaction and related applications. One of the possible approaches to the detection is exploitation of space-time interest points. Such points are typically identied using feature extractor and then they are processed and classified. The classification can be performed using codebooks built based on feature vectors statistics. The individual feature vectors are transformed into bag of words representation using such codebooks and then the code words are classied using SVM. The proposed approach improves the training process and extends the known approaches. The training part of the dataset is split into shorter shots with equal duration and these are annotated and classied using a SVM classier. This ensures that the time-local motion is captured by the SVM while the longer time behavior is left on further processing mechanisms, such as, e.g. HMMs. In the proposed approach, the output of the SVM classier is simply compared to a threshold and the presence of a value above the threshold indicates that the desired human activity occurred. The contribution describes the approach summarizes the achieved results and draws conclusions.

The on-line human action detection is an important task in human-machine interaction and related applications. One of the possible approaches to the detection is exploitation of space-time interest points. Such points are typically identied using feature extractor and then they are processed and classified. The classification can be performed using codebooks built based on feature vectors statistics. The individual feature vectors are transformed into bag of words representation using such codebooks and then the code words are classied using SVM. The proposed approach improves the training process and extends the known approaches. The training part of the dataset is split into shorter shots with equal duration and these are annotated and classied using a SVM classier. This ensures that the time-local motion is captured by the SVM while the longer time behavior is left on further processing mechanisms, such as, e.g. HMMs. In the proposed approach, the output of the SVM classier is simply compared to a threshold and the presence of a value above the threshold indicates that the desired human activity occurred. The contribution describes the approach summarizes the achieved results and draws conclusions.