People Counting Based on Multi-scale Region Adaptive Segmentation and Depth Neural Network

People counting based on surveillance camera is the basis of the important tasks, such as the analysis of crowd behavior, the optimal allocation of resources and public security. Aiming at the low accuracy of the people counting method based on object detection, a people counting method based on multi-scale region adaptive segmentation and deep neural network is proposed in this paper. The idea originates from the analysis and research of multi-scale objects, and it is found that the detection accuracy will be improved if the multi-scale objects match the size of multi-scale anchors. In this method, K-means is used to cluster the detection results of Faster-RCNN model. Then the image is segmented adaptively according to the clustered results. Finally, Faster-RCNN model is used to detect the segmented images. The experimental results show that the average accuracy of this method is 45.78% on mall dataset, which is higher than Faster-RCNN about 3.59%.