Crowd statistical method based on depth learning

Abstract

The method discloses a crowd statistical method based on depth learning. The method extracts a motion foreground from a video, ensures camera angle of view and perspective invariance by using a humanbody region model, and finally determines the human body region statistical number through preprocessing, extraction, and detection. The method can not only reduce the search area of the sliding window and improve the search efficiency, but also overcome the deformation defect of the monitoring video, caused by view angle, monitoring scene distance, and the like, and installation and deployment ofthe system are simple, a detection model based on a depth learning convolution neural network improves the accuracy of human detection, the use of a non-maximum suppression method eliminates redundant sub-regions to reduce repeated counting, so that the results of human detection, human location and population statistics are more accurate.

Description

Technical field [0001] The method belongs to the field of video intelligent monitoring, and specifically relates to a crowd counting method based on deep learning. Background technique [0002] With the popularity of video surveillance systems, cameras have spread all over the city. First of all, in the face of such a large number of cameras and surveillance videos, it is unrealistic to manually identify the behavior and attributes of the crowd in the surveillance scene. Secondly, in the face of complex scenes such as rainy days, snowy days, night scenes or super-dense crowds, it is still difficult to identify the people in them with the naked eye, let alone count the number of people in them. [0003] At present, crowd counting methods applied to video surveillance systems are mainly divided into three categories: the first category uses detectors to slide in the image to determine and count human bodies one by one; It is the result of crowd counting; the third category u...

AI Technical Summary

Problems solved by technology

However, the above methods all use manual feature extraction, which is not suitable for more complex scenes, or does not introduce perspective and observation angle invariance, so that the method cannot deal with object deformation caused by perspective and perspective, and cannot be well applied to wide field of vision. In some scenarios, or the method of solving the invariance of perspective and viewing angle is adopted, but the accuracy is too dependent on the user's manual measurement of parameters such as camera shooting angle and viewing distance, which complicates the installation and configuration of the system

However, purely using a detector to process an image depends on the quality of the detector, and using a sliding window to traverse the entire image requires a huge amount of calculation, and real-time performance is difficult to guarantee

Images