A generic object counting algorithm under partial occlusion conditions

This paper addresses the generic object counting problem with object overlapping occurring at varied levels and degrees. The overall image containing the objects is segmented from the background. Thereafter a combination of parameters is extracted from each of the segments to construct a parameter space. The overall space formed by these vectors contains redundant dimensions due to the existence of correlated information among the selected parameters. Therefore, by performing a principal component analysis (PCA) of the parameter space, the redundant dimensions are dropped. A Gaussian model is constructed for the base cluster containing one object per segment in the reduced space. Through the base model parameters, the model order or the object count per segment of the other clusters are appropriately identified such that they serve as likelihood functions for each model order. Tunable parameters are identified for this model labeling process, under the assumption of a decaying growth of the mean and an enhancing growth of the standard deviation for higher model orders. An important aspect of the proposed method is the ability to estimate the number of objects within a cluster regardless of the shape, colour and orientation of objects.