Demographic statistics device and method based on boundary selection

Abstract

The invention discloses a demographic statistics device and method based on boundary selection, and belongs to the field of deep learning and image processing. The invention comprises the following steps of by improving the YOLO neural network by increasing the division of the YOLO unit from 7*7 to 9*9, that is, increasing the number of detections per unit to three, and then replacing the YOLO PCneural network with the Fire module in SqueezeNet; replacing The 16th, 18th, and 24th 3*3 convolutional layers with Dire modules in SqueezeNet and decreasing the number of compressed portion of the Fire module from 128 to 96; retraining network to get a new S-YOLO-PC neural network, using the S-YOLO-PC neural network for boundary selection of human flow statistics, using the new neural network, making it more accurate in the case of greatly reduced models, can be used in a variety of occasions Human flow detection.

Description

technical field [0001] The invention relates to a device and method for people counting based on boundary selection, belonging to the fields of deep learning and image processing. Background technique [0002] In the video surveillance screen, pedestrians are very important detection targets. Pedestrian detection, crowd density estimation and people flow statistics are all key components of smart security and smart buildings. The changing and complex background in the video poses a challenge to how to distinguish and detect pedestrians and other types of objects in the surveillance screen, and how to effectively distinguish pedestrians from their backgrounds. In specific scenarios, this work often has problems such as inaccurate detection, inaccurate counting, long delays in detection and counting results, and large storage space occupied by deep models, which is not conducive to transmission and downloading. [0003] In recent years, deep learning technology has set off a ...

AI Technical Summary

Problems solved by technology

[0004] YOLO is an advanced real-time target detection algorithm with a relatively high accuracy rate, but it still encounters many problems in the actual application environment, such as taking up a large storage space, which is not conducive to transmission, download, and detection in actual applications. Seriously restricting its practical application, the YOLOv2 updated after 2017 and the later YOLOv3 version borrowed the idea of ​​FasterR-CNN and introduced the anchor detection mechanism, which made the new YOLO neural network unable to use the boundary selection method for pedestrian counting

Images