This thesis is concerned with the design of a complete framework that allows the real-time recog- nition of humans in a video stream acquired by a static camera. For each stage of the processing chain, which takes as input the raw images of the stream and eventually outputs the identity of the persons, we propose an original algorithm. The first algorithm is a background subtraction technique named ViBe. The purpose of ViBe is to detect the parts of the images that contain moving objects. The second algorithm determines which moving objects correspond to individuals. The third algorithm allows the recognition of the detected individuals from their gait.
Our background subtraction algorithm, ViBe, uses a collection of samples to model the history of each pixel. The current value of a pixel is classified by comparison with the closest samples that belong to the collection. We propose an innovative model update technique which allows to obtain an appropriate modeling of the history of the pixel with a reduced number of samples. Furthermore, our model update policy ensures the spatial coherence of the background model and enables the use of a background model composed exclusively of background samples. We show that ViBe outperforms other state-of-the-art techniques while being faster than most of them. Our algorithm is actually fast enough to run in real-time on the low speed processor of a Canon camera.
We then introduce an algorithm that processes the silhouettes of the moving objects detected by ViBe. The purpose of this algorithm is to detect and locate humans. The silhouettes of the moving objects are classified as being either human or non-human. This classification is performed on the basis of the cover by rectangles of the silhouettes, which is a new morphological operator introduced in this thesis. Rectangles from the cover by rectangles of a silhouette are classified individually. The silhouette is then classified according to a majority vote policy among its rectangles. We show that the detection of the persons is robust and can be computed in real-time.
The last stage of the processing chain recovers the identity of the detected persons from their gait. Our gait recognition technique is a logical follow-up to our person detection algorithm. This algorithm is based on the classification of a gait signature computed from the covers by rectangles of the silhouettes of a walker. The purpose of a gait signature is to capture the dynamics of the time series of the silhouettes of a walking person. Experiments on a public database show that the results of our gait recognition technique are on par with those of other techniques described in the literature. In the last part of this thesis, we apply our gait recognition algorithm to an original application: the intelligent control of access to a secure area.