Method for counting video objects in real time based on any scene

Abstract

The invention discloses a method for counting video objects in real time based on any scene. The method comprises the following steps of extracting motion foreground pixels from a certain number of video scenes which are marked in advance by adopting a hybrid Gaussian background modeling method blending a frame difference method and a block tracking method, and performing shadow detection to remove shadow areas to obtain a more accurate foreground area; and extracting each communicated sub-block by adopting a connected component algorithm, calculating characteristics of the sub-blocks, training the sub-blocks in all the scenes and object data corresponding to the sub-blocks, and performing a foreground extraction algorithm and extraction of the characteristics of each sub-block on video objects in any unknown scene to count the numbers of the video objects in real time after the training. According to the method, object number information under the video monitoring of a public place can be acquired in real time, so that control over public safety and local scene traffic is greatly facilitated, the labor cost is effectively saved, and the monitoring intelligence is improved.

Description

technical field [0001] The invention relates to the fields of video monitoring, pattern recognition and image processing, in particular to a method for real-time statistics of video objects based on any scene. Background technique [0002] Due to the increasingly prominent security issues, it is necessary to conduct video surveillance in areas with a large traffic flow and people flow, such as stations, squares, and shops. Video object count detection can obtain object count information in real time, issue timely alarms for abnormally dense crowds in some areas, or help subways, shopping malls, roads, etc. to obtain people or vehicle flow information in a timely manner, so as to better guide the flow of people and vehicles to ensure safety. More and more attention has been paid to the real-time statistics technology of video objects. [0003] At present, there are still many problems to be solved in the video object statistics technology at home and abroad: [0004] 1) Th...

AI Technical Summary

Problems solved by technology

[0004] 1) The foreground extraction is not accurate enough and the calculation is large

Current foreground detection algorithms include frame difference method, mixed Gaussian modeling method, optical flow method, motion detection algorithm, etc., often have missed detection and false detection, and do not use the relationship between adjacent frames to improve the extraction accuracy, or the amount of calculation Large, or the extraction accuracy is not enough

If the amount of calculation is large, it is difficult to meet the requirements of real-time processing, and if the extraction accuracy is not enough, it will affect the accuracy of counting

[0005] 2) For the change of light intensity, there is no good adaptability, there are large false detection areas and the effect of shadow detection is not good

The change of light in the video is very common. In the outdoor scene, the sunlight is blocked by clouds, and the indoor light switch or light is blocked, which often cause a large area of ​​false detection and have a great impact on the object statistics.

At present, there is no good method for false detection caused by lighting effects such as ground reflections and object projections.

[0006] 3) There is currently no scene-independent video object statistics system

After installing the camera, a certain number of frames of images must be obtained for calibration and training before object statistics can be performed, which cannot be used at any time, so it is extremely inconvenient and limits the application of video object statistics

Images