High-Accuracy Crowd Counting Method Based on Mixed Labeled Dataset

Existing crowd counting technology mainly uses one single detection model or density map regression model to obtain the number of people in an image, but a single model is difficult to adapt to different camera position, viewing angle, resolution, and different crowd density. We believe that fusion of two different methods for prediction can greatly improve the accuracy of prediction in various situations. In this paper, we propose a method that improve the existing counting method of detection counting method, and then develop a network to fuse the results of regression prediction and detection prediction, so that not only can we make full use of the existing different labeled dataset but also the network generalize stronger by fusing completely two model and the counting accuracy will be higher. Experiments show that with the same model and parameters, the counting accuracy of the method proposed in this paper is much higher than any single model.