An adaptive people counting system with dynamic features selection and occlusion handling

• A pair of collaborative Gaussian process models (GP) is used to handle occlusion.
• A principled technique is proposed to measure the level of occlusion in a frame.
• It proposes a method of choosing the best features depending on their environment.
• The system is evaluated using two benchmark datasets, the Mall and UCSD datasets.

This paper presents an adaptive crowd counting system for video surveillance applications. The proposed method is composed of a pair of collaborative Gaussian process models (GP) with different kernels, which are designed to count people by taking the level of occlusion into account. The level of occlusion is measured and compared with a predefined threshold for regression model selection for each frame. In addition, the proposed method dynamically identifies the best combination of features for people counting. The Mall and UCSD datasets are used to evaluate the proposed method. The results show that the proposed method offers a higher accuracy when compared against state of the art methods reported in open literature. The mean absolute error (MAE), mean squared error (MSE) and the mean deviation error (MDE) for the proposed algorithm are 2.90, 13.70 and 0.095, respectively, for the Mall dataset and 1.63, 4.32 and 0.066, respectively, for UCSD dataset.