15-09-2017, 12:09 PM
Significant research has been devoted to the detection of people in images and videos. This increase results in an extremely high dimensional space (more than 170,000 dimensions). In such large-scale spaces, classical machine-learning algorithms like SVMs are almost intractable with respect to training. In addition, the number of training samples is much less than the dimensionality of the feature space, at least one order of magnitude. Finally, the extraction of characteristics from a densely sampled grid structure leads to a high degree of multicollinearity. In order to circumvent these data characteristics, we use the partial least squares (PLS) analysis, an efficient dimensionality reduction technique that preserves significant discriminative information to project the data into a subspace of much smaller dimensions (20 dimensions, reduced by the 170,000). Our human detection system, using PLS analysis on the enriched descriptor set, is shown to outperform the most advanced techniques in three varied datasets including the popular INRIA pedestrians data set, the low-resolution Daimler-Chrysler pedestrians data set grayscale and the ETHZ pedestrian data set consisting of full length videos of crowded scenes.
It can be understood in the following video:
It can be understood in the following video: