A hybrid model of convolutional neural networks and deep regression forests for crowd counting

Abstract

Real-time monitoring variation of crowd via video surveillance plays a significant role in the new generation of technology in a smart city. We propose a crowd counting algorithm based on deep regression forest, named CountForest. First of all, according to the correlation among frames, the crowd counting problem is transformed into a label-distribution-learning problem. Then we combine convolutional neural networks(CNN) and deep regression forest to make a hybrid model. CNN is introduced for the task of feature learning and deep decision forest is extended to address label distribution learning problem in crowd counting. Thereinto, the proposed network replaces its softmax layer with the aforementioned probabilistic decision forest in order to better establish a mapping relationship between image features and crowds’ number so as to implement an end-to-end hybrid model for crowd counting problem. Our method demonstrated in the final experiments not only attains the high accuracy in crowd counting but has comparable robustness and instantaneity in selected public datasets as well.