Dense crowd estimation method based on deep learning

Abstract

The present invention relates to a method for estimating the number of dense people based on deep learning, comprising the following steps: selecting an image of a dense scene as a test image, and then performing a block operation on the test image, and the ratio of the blocks should be guaranteed to be the same as the width and height of the original image The ratio is approximately the same; the divided image blocks are normalized and normalized into 32×32 pixel blocks as our test samples, and the corresponding real number of people labels are attached; the pixel blocks are sent in batches to the already trained In the deep network of , for each pixel block, the network will feed back a prediction result; the prediction results of each pixel block are summed, and the result obtained is the total number of people in the test image that we need to estimate. The beneficial effects of the present invention are: the method of deep learning is introduced into the specific problem of people counting; the constructed regression model including two-way signals reduces the possibility of over-fitting to a certain extent.

Description

technical field [0001] The invention belongs to the field of image processing and computer vision, and relates to a method for estimating dense numbers of people based on deep learning. Background technique [0002] Estimating people in dense settings has many potential practical applications, including surveillance (e.g., detecting unusually large crowds, or controlling the number of people in an area), security management (recording the number of people entering or leaving an area) , urban planning (for example, analyzing the flow of people in a certain area), etc. Therefore, people counting is an important research topic in computer vision and related fields. [0003] There are two main difficulties in population estimation in dense scenes: [0004] (1) Factors such as mutual occlusion of objects in the scene, perspective distortion of the scene, visual blur caused by lighting conditions, and complex crowd activities. [0005] (2) When the crowd tends to be crowded, th...

AI Technical Summary

Problems solved by technology

[0005] (2) When the crowd tends to be crowded, the number of people is difficult to use traditional features (such as HOG, Haar wavelet, gray level co-occurrence matrix) to represent, making it difficult to apply the method based on feature extraction and location detection to more than one hundred people in the scene

[0008] Shortcomings of prior art one: (1) This method can only be used to detect the scene of dozens of people, when the number of people in the scene exceeds one hundred, it is difficult to detect effectively, such as figure 1 (2) detection-based methods need to use sliding windows, which is a time-consuming process

[0011] Disadvantage of prior art 2: only the density level of people in the scene can be estimated, but no specific estimated value of the number of people can be given

Images