A crowd counting method and system based on a spatial perception attention refinement framework

Abstract

The invention discloses a crowd counting method and system based on a spatial perception attention refinement framework. The method comprises the following steps: generating feature maps, initial crowd density maps and density grade information of all training images by utilizing a convolutional neural network; inputting the information into an iterative spatial perception refining module, carrying out dynamic localization on the crowd density map in a spatial perception mode by utilizing a spatial regression mapping network, generating a corrected local density map, and carrying out a dynamiclocalization updating strategy for next iteration by combining a long-short-term memory module; Encoding the density level information to generate a density level graph, integrating the local densitygraph with the density level graph to serve as local and global information of each round of iteration, and inputting the local and global information into a local refining network; the local refinement network adjusts the density distribution of the input area, performs reverse space regression mapping on the input area, and updates and generates a crowd density map in a residual error learningmode; and iteratively carrying out the training process for multiple times to obtain a refined crowd density map.

Description

technical field [0001] The present invention relates to the field of computer vision based on deep learning, in particular to a crowd counting method and system based on a space-aware attention refinement framework. Background technique [0002] The crowd counting problem aims to analyze the number and area density of crowds in a scene. In recent years, due to the brilliant performance of deep learning methods in the field of computer vision, under the joint promotion of data and algorithms, the problem of crowd counting has attracted more and more researchers' attention due to its wide application prospects and subject exploratory nature. [0003] Many methods in the early years were mostly based on the framework of pedestrian detection, and completed the task of crowd counting through the idea of ​​detection first and then counting. Such methods generally analyze the appearance and motion patterns of pedestrians or extract texture features of multiple unit areas of the ima...

AI Technical Summary

Problems solved by technology

However, limited by hardware resources and computing resources, these methods cannot exhaust all scenes; most importantly, these methods completely ignore the perspective distortion caused by different camera viewpoints in different s

Images