Research Lab


Cascade Classifier for People Detection in Surveillance Videos

Introduction firefox For the purpose of people detection in surveillance videos, generic detector may be degraded in specific scenes due to variant illuminations, backgrounds, camera viewpoints and video quality. To overcome some of these difficulties, a three-layer cascade classifier is developed. Specifically, two independent linear SVMs based on histogram of oriented gradients (HOG) are used in the first layer to efficiently generate front and side candidates to the next layer. The second layer are kernelized SVMs using Multiscale HOG (MSHOG) to reject most of false positives that passed the first layer. To avoid the curse of dimensionality, the partial least squares analysis (PLS) is employed for dimemsionality reduction. As a result, the high-dimensional MSHOG descriptor is projected on to a much lower dimensional subspace. Finally, in addition to including background information,a more discriminative descriptor combining MSHOG and color self-similarity (CSS) and projected by PLS is used in the third layer to filter out dificult negatives passed the previous layers. firefox


Sequence 1


Sequence 2