Scene-dependent proposals for efficient person detection

In this paper, we present a new method that provides a substantial speed-up of person detection while showing high classification accuracy. Our method learns a Gaussian Mixture Model of locations and scales of the persons in the scene under observation. The model is learnt in an unsupervised way from a set of detections extracted from a small number of frames, so that each component of the mixture represents the expectation of finding a target in a region of the image at a specific scale. At runtime, the windows that most likely contain a person are sampled from the components and evaluated by the classifier. Experimental results show that replacing the classic sliding window approach with our scene-dependent proposals in state of the art person detectors allows us to drastically reduce the computational complexity while granting equal or higher performance in terms of accuracy.