Crowd counting model training method based on automatic domain division and application

Abstract

The invention relates to a crowd counting model training method based on automatic domain division and application. The training method comprises the following steps: 1) establishing and training to obtain a preliminary training model; 2) verifying the preliminary training model by using a test set, and obtaining an identification error corresponding to each image in the test set; 3) dividing theimage with the identification error greater than or equal to a set threshold value into a target domain, and dividing the image with the identification error less than the set threshold value into a source domain; and 4) adding a second loss function between the source domain and the target domain, and performing optimization based on the second loss function and the original loss functions of thetwo domains to obtain a final model.Compared with the prior art, the method has the advantages that the adaptability is high, the counting accuracy is improved and the like.

Description

technical field [0001] The invention relates to the field of robot vision, in particular to a crowd counting model training method and application based on automatic domain division. Background technique [0002] Crowd counting is of great significance in intelligent security, urban planning, etc. Crowd monitoring can effectively avoid stampede incidents, and at the same time adjust urban infrastructure construction and urban layout according to historical crowd flow and crowd distribution. [0003] The existing crowd counting algorithms are mainly divided into the following categories: [0004] 1. The method based on target detection: including detection based on the whole and detection based on parts of the body - by locating and identifying each pedestrian or head on the image, and then counting the number of people based on the results. The advantage is that it can detect accurate pedestrians or head positions, but the disadvantage is that for high-density crowd images,...

AI Technical Summary

Problems solved by technology

The advantage is that it can detect accurate pedestrians or head positions, but the disadvantage is that for high-density crowd images, the detection effect is poor, and it is difficult to deal with serious occlusion problems between crowds.

[0005] 2. Regression-based method: that is, the estimation of the number of crowds does not accurately locate the position of pedestrians, but gives an estimate of the approximate number of crowds. MCNN and CSRNet are more representative deep learning crowd counting algorithms, which are based on density maps. regression algorithm

The advantage of the regression-based method is that it is better than the target detection method for high-density crowd images, but the disadvantage is that there is no accurate positioning of pedestrians, and the existing network model cannot adapt to the changing distribution of pedestrians.

[0006] When using deep learning to solve the problem of crowd counting, the training data is often relatively single, and the obtained model cannot meet good adaptability; when multiple data sets (multiple domains) with different data distributions are used for simultaneous training, the model is greatly reduced. Therefore, taking into account the accuracy and data adaptability is an urgent problem to be solved in the practical application of crowd counting.

In addition, the common network is only trained on a certain data set first, and can only have higher accuracy on a data set that satisfies the same data distribution. For example, the training results of CSRNet on ShanghaiTech PartA perform poorly on PartB.

Images