Single-image crowd counting method based on depth information and scale perception information

Abstract

The invention relates to a computer vision technology, and discloses a single-image crowd counting method based on depth information and scale perception information, which improves the prediction capability and reduces the calculation complexity. The method comprises the steps of S1, performing Gaussian mapping on human head center coordinate data corresponding to an input sample picture to generate a preliminary true value density map, and correcting the preliminary true value density map based on depth information obtained by a depth estimation algorithm to obtain a true value density map;s2, predicting a crowd density map of the input sample picture by adopting a density estimation network to generate a predicted density map, calculating a loss error according to the predicted densitymap and a true value density map, adjusting network parameters through gradient back propagation, and generating a density prediction model through iteration; and S3, when crowd counting is carried out on a single image, a density prediction model is utilized to generate a prediction density map of the image, and the total number of people in the image is obtained through calculation.

Description

technical field [0001] The invention relates to computer vision technology, in particular to a single image crowd counting method based on depth information and scale perception information. Background technique [0002] The purpose of crowd counting is to input a picture, after being processed by the network model, output the crowd density map corresponding to the picture, and finally sum the probability of the number of people corresponding to each pixel on the density map to get the final total number of people. Crowd counting tasks are full of challenges due to issues such as occlusions, perspective changes, crowd scale changes, and distribution diversity. [0003] The early methods mainly locate each pedestrian in the crowd through the target detector, and the number of detected targets is the counting result. However, these methods use handcrafted features for classifier training and perform poorly in highly crowded scenes. In order to solve the problem of crowd coun...

AI Technical Summary

Problems solved by technology

Although multi-scale information can be integrated using the Inception structure, as the network passes forward, the features are highly abstracted, and the detailed features of small targets are lost, resulting in a decline in the prediction ability of final small targets.

In addition, using transposed convolution for scale reduction has high computational complexity, and its performance has no outstanding advantages within a certain range of training batches.

Images