Video-based dense crowd counting method and system

Abstract

The invention provides a video-based dense crowd counting method, and the method comprises the steps: calculating a video superposed optical flow according to original video data; generating motion features according to the video superimposed optical flow; generating appearance features according to the original video data; cascading the motion features and the appearance features to obtain mergedfeatures, and inputting the merged features into an expansion convolutional network to generate crowd active region segmentation; dividing the crowd active region, inputting the crowd active region into a spatial attention mechanism network, generating a spatial attention graph, and weighting the appearance features by using the spatial attention graph to obtain the appearance features after attention weighting; and inputting the attention weighted appearance features into an expansion convolutional network to obtain crowd density estimation. Meanwhile, the invention provides a video-based dense crowd counting system. According to the method, density estimation based on spatial attention is realized, so that the purpose of reducing background interference is achieved, and the robustness of crowd counting is improved.

Description

technical field [0001] The present invention relates to the technical field of dense crowd counting, in particular to a video-based dense crowd counting method and system. Background technique [0002] The dense crowd counting problem is the problem of using computer vision technology to determine the total number of people in an image or video sequence. [0003] (1) Analysis of recent patented technologies on crowd counting: [0004] 1. The Chinese invention patent application "Method and Device for Counting Dense Crowds" with the application number CN201810986919.0 proposes a crowd counting method that uses CNN to perform density regression on still images one by one. This method only uses a single frame for each density regression The data does not consider the relationship between frames; [0005] 2. The Chinese invention patent application with the application number 201710514797.0 "Method and device for calculating dense crowd flow based on video" uses the frame diff...

AI Technical Summary

Problems solved by technology

For complex background situations, it is difficult for this method to distinguish between moving backgrounds and moving pedestrians;

[0006] 3. The application number is 201310136575.1 Chinese Invention Patent Application "A Crowd Density Estimation Method and People Flow Statistics Method Based on Video Analysis". Although the optical flow is also used for people flow statistics, the optical flow is only used to calculate the crowd Flow velocity instead of weak supervision for crowd active area segmentation

Its shortcoming is that the ranking mechanism of crowd density depends heavily on the data set, resulting in poor transfer learning performance;

The disadvantage is that this method requires intensive time annotation and cannot be widely used

[0010] The following conclusions can be drawn from the analysis of relevant domestic and foreign patents and related research: At present, there is no combination of motion information for weakly supervised active area segmentation, and then combined with image CNN appearance features for the application of video dense crowd counting methods

Images