Picture-based crowd counting method

Abstract

The invention relates to the technical field of crowd counting methods, and discloses a picture-based crowd counting method, which comprises the following steps of S1, inputting an original picture, and extracting picture features through a VGG-16 network; S2, carrying out 2 * 2 average pooling operation on the extracted VGG feature values to obtain blocks with the size of kxk, and then carrying out convolution on the blocks and a convolution layer with the kernel of 1; the convolution operation with the kernel of 1 has the greatest advantage that the dimension of an original feature value isnot changed, and then a result is input into a Normalization Laye (normalization layer); and S3, re-up-sampling to recover to the feature size of the previous VGG, and finally decoding and outputtingthe density map. According to the picture-based crowd counting method, an end-to-end trainable network architecture is adopted, multi-scale transformation is adapted by learning how to configure weights for different features at an independent pixel level, the influence of perspective distortion on a density map can be effectively reduced, and the accuracy of a crowd calculation result is improved.

Description

technical field [0001] The invention relates to the technical field of crowd counting methods, in particular to a picture-based crowd counting method. Background technique [0002] The early crowd counting method is mainly based on the statistics of the number of pedestrian detection, the position of pedestrians in the crowd is obtained through human detectors or some human body detectors, and then the number of positions is added to get the final crowd counting result. This method needs to obtain human body features through a large amount of human body data, such as: head and shoulder features. When counting, the number of detection frames is accumulated to obtain the current number of pedestrians. This method can get good results for ordinary scenes, but for high Density crowd counting, due to the mutual occlusion between people or the size of the target is too small, it is easy to cause the target detection to be inaccurate, and eventually cause the error of the number st...

AI Technical Summary

Problems solved by technology

[0005] The invention provides a picture-based crowd counting method, which has the advantages of assigning weights to different features at the independent pixel level to adapt to multi-scale transformation, effectively reducing the influence of perspective distortion on the density map, and improving the accuracy of crowd counting results. The problems mentioned in the background technology

Images