Human detection from fish-eye image by Bayesian combination of probabilistic appearance models

This paper presents a method for automated human detection using fisheye lens camera. We introduce a probabilistic model to describe the wide variation of human appearance in hemispherical image. In our method, a human is modeled as variable shape features of body silhouette and head-shoulder contour. These features are extracted from the human images taken at various distance and orientation with respect to the camera, and form the training data set for template modeling. Non linear template models are build by the combination of Principal Component Analysis (PCA) and Kernel Ridge Regression (KRR). Finally, the problem of human detection is formulated as maximum a posteriori (MAP) estimation using above model. Experiments are conducted on indoor space where a fisheye lens camera is installed on the ceiling. The robustness and accuracy of our method is discussed through the experimental results.