Complex human activity recognition for monitoring wide outdoor environments

The problem of automatic recognition of human activities is among the most important and challenging open areas of research in computer vision. This work presents a new approach to automatically recognize complex human activities embedded in video sequences acquired with a large scale view in order to monitoring wide area (car parking, archeological site, etc) with a single static camera. The recognition process is performed in two steps: at first the human body posture is estimated frame by frame and then the temporal sequences of the detected postures are statistically modeled. Body postures are estimated starting from the binary shapes associated to humans, selecting as features the horizontal and vertical histograms and supplying them as input to an unsupervised clustering algorithm. The Manhattan distance is used for both clusters building and run-time classification. Statistical modeling of the detected postures is performed by discrete hidden Markov models. The system has been tested on image sequences acquired in an outdoor archaeological site. Four kinds of activities have been automatically classified with high percentage of right decisions.