CN-121982074-A - Target detection method based on Kalman filtering

CN121982074ACN 121982074 ACN121982074 ACN 121982074ACN-121982074-A

Abstract

The invention belongs to the technical field of computer vision detection and provides a target detection method based on Kalman filtering, which comprises the steps of obtaining at least two frames of images continuously collected by infrared imaging equipment, carrying out differential operation on two adjacent frames of images and taking absolute values to obtain differential images; the method comprises the steps of respectively establishing a Kalman filter for each pixel point of a differential image, initializing the Kalman filter based on the differential gray values of the corresponding pixel points, carrying out frame-by-frame filtering updating on the differential gray values of the pixel points in the differential image by the Kalman filter, constructing a threshold image, resetting the Kalman filter, comparing the differential image with the threshold image to divide the differential image to obtain a binary image, processing the binary image to obtain candidate target information, and carrying out track association processing on the candidate target information to obtain target point trace information. The invention improves the credibility of the target detection result and suppresses the false alarm caused by background disturbance.

Inventors

QIU CHENG
HAN JIANTAO
He Zunliang
LI YANG

Assignees

湖南傲英创视信息科技有限公司

Dates

Publication Date: 20260505
Application Date: 20260409

Claims (10)

1. The target detection method based on Kalman filtering is characterized by comprising the following steps of: s1, acquiring images continuously acquired by infrared imaging equipment, performing differential operation on two adjacent frames of images, and taking absolute values to obtain a plurality of frames of differential images; S2, establishing an independent Kalman filter for each pixel point of the differential image, and initializing each Kalman filter by taking the differential gray value of the corresponding pixel point as an initial value; s3, carrying out frame-by-frame filtering updating on the differential gray values of all pixel points in the differential image by using the Kalman filter; s4, constructing a threshold image according to a filtering updating result; S5, after the threshold image is constructed, differentiating the newly acquired two adjacent frames of images, and comparing the obtained differential image with the threshold image to divide the differential image so as to obtain a binary image reflecting the moving target area; S6, obtaining candidate point tracks according to the binary image, and carrying out track association processing on the candidate point tracks based on a set number of frames to obtain target point track information.
2. The kalman filter-based object detection method according to claim 1, wherein in S2, the initialization parameters involved in the initialization process of the kalman filter include a process noise covariance matrix and an observation noise covariance matrix, and the process noise covariance matrix and the observation noise covariance matrix of each kalman filter take the same preset parameter values during the initialization.
3. The kalman filter-based object detection method according to claim 1, wherein in S3, performing frame-by-frame filter update on the differential gray values of each pixel point in the differential image using the kalman filter specifically includes: s31, inputting the differential gray values of all pixel points in the differential image into a corresponding Kalman filter frame by frame according to an image frame sequence; S32, adopting a corresponding Kalman filter to carry out filtering update of the preset frame number on the differential gray values according to a judging rule, wherein the judging rule is to compare two adjacent frames of differential images and carry out filtering update on a larger value of the differential gray values of the pixel points.
4. The kalman filter based target detection method according to claim 3, wherein the decision rule includes: When the difference gray value of the pixel point of the current frame of difference image is larger than the difference gray value of the pixel point corresponding to the previous frame of difference image, the difference gray value of the current frame of difference image is used as an observation value to be input into the Kalman filter; and when the differential gray value of the pixel point of the differential image of the current frame is smaller than the differential gray value of the pixel point corresponding to the differential image of the previous frame, inputting the differential gray value of the differential image of the previous frame into the Kalman filter as an observation value.
5. The kalman filter based object detection method according to claim 3, wherein the counter is incremented by 1 while S32 is performed.
6. The kalman filter based object detection method according to claim 3, wherein the preset frame number is 60 frames.
7. The kalman filter-based object detection method according to claim 1, wherein in S4, after obtaining a threshold image, the kalman filter is reset to enter a next round of filter update.
8. The kalman filter based target detection method according to any one of claims 1-7, wherein S6 specifically comprises the following operations: S61, acquiring target point tracks in the binary image, wherein the target point tracks comprise moving objects and pseudo-moving objects, and establishing an empty track list for the target point tracks; S62, regarding all the points detected by the first frame binary image as new points, associating the new points with the points of the second frame binary image based on a nearest neighbor association algorithm, establishing a new track, and adding the new track into the track list, wherein each track comprises a prediction state of a corresponding target, and the prediction state comprises a prediction position and a prediction speed; S63, based on the prediction state of each track in the track list, constructing an association threshold area in the binary image of the current frame by taking the prediction position as the center, and taking the target point track falling into the association threshold area as a candidate point track with a target mark, wherein the association threshold area is a space area; s64, calculating the statistical distance between the candidate point track and the corresponding track through a nearest neighbor correlation algorithm, and distributing a correlation point track for each track; S65, for candidate points with target marks which are not successfully completed in a continuous set number of frames, the candidate marks are withdrawn, the track is terminated in a track list, the corresponding prediction state is updated by using the candidate points with target marks which are successfully completed in the continuous set number of frames, and the target marks are maintained, so that target point track information is obtained, wherein the target point track information comprises coordinates, areas and circumscribed rectangles.
9. The kalman filter based object detection method according to claim 8, wherein the step of looping S62-S65 takes candidate points with object marks of the current frame not associated to the existing tracks as new track start points.
10. The kalman filter-based object detection method according to claim 8, wherein in S64, the statistical distance is a mahalanobis distance, and the nearest neighbor correlation algorithm is replaced with an overall optimal correlation algorithm according to need.

Description

Target detection method based on Kalman filtering Technical Field The invention belongs to the technical field of computer vision detection, and particularly relates to an infrared target detection method. Background The wide area Zhou Saogong external detection platform is generally used for detecting moving objects in a ground complex scene, wherein the detected objects comprise people, vehicles, small animals and the like. Because infrared imaging is jointly influenced by target radiation characteristics, ground surface background and environmental factors, the detection of infrared moving targets under the condition of complex ground background faces great difficulty. On one hand, the ground moving object usually shows low signal-to-noise ratio and low contrast in an infrared image, the gray level difference between the object and the background is not obvious, and on the other hand, the ground background has the characteristics of complex structure, strong time variability, non-stability and the like, and the size, shape and moving speed of the object in an actual scene are greatly changed. In addition, environmental factors such as environmental temperature change and weather conditions can further aggravate the existence of noise and interference in the infrared image, and the stable detection of the moving target is seriously influenced. In view of the above, the core idea of detecting infrared motion in existing ground backgrounds is generally to use motion information of a target, radiation or texture difference between the target and the background, or a combination of the two to achieve separation of the target and the background. The method for detecting the target mainly comprises the following steps of (1) establishing a background model of a scene based on a background modeling method, and differentiating a current frame image with the background model to extract a foreground moving target. The method has clear structure and wide application, and typical algorithms comprise Gaussian mixture models and the like. However, in a scene with rapid background change or large noise, the background model is easy to fail, so that false detection and omission are caused, and (2) an optical flow-based method is used for detecting a target by analyzing the motion information of pixels in an image sequence, so that the method can adapt to a camera motion scene to a certain extent. The method is sensitive to noise, has higher computational complexity and higher requirements on real-time processing capacity and hardware resources, and is widely researched and used because the method is generally used for directly detecting targets in an end-to-end mode or is combined with the traditional method, can learn complex background and target characteristics and has higher detection performance. However, the deep learning detection method has larger dependence on large-scale labeling data and a high-performance computing platform, is limited in application in an infrared detection scene with limited resources or higher real-time requirements, and still has the problems of poor feature extraction effect, larger noise interference, larger detection error caused by complex ground and still failure in detection accuracy to meet market demands. Therefore, under the condition of complex ground background, how to effectively inhibit noise interference and realize stable detection of a moving target on the premise of considering detection precision and calculation efficiency is still a technical problem which needs to be solved in the field of infrared moving target detection at present. Disclosure of Invention In view of the above, the invention provides a target detection method based on Kalman filtering, which aims to solve the problems of how to effectively inhibit noise interference and realize stable detection of a moving target under the premise of considering detection precision and calculation efficiency under the complex background condition of the ground. The invention provides a target detection method based on Kalman filtering, which comprises the following steps: A target detection method based on Kalman filtering comprises the following steps: s1, acquiring images continuously acquired by infrared imaging equipment, performing differential operation on two adjacent frames of images, and taking absolute values to obtain a plurality of frames of differential images; S2, establishing an independent Kalman filter for each pixel point of the differential image, and initializing each Kalman filter by taking the differential gray value of the corresponding pixel point as an initial value; s3, carrying out frame-by-frame filtering updating on the differential gray values of all pixel points in the differential image by using the Kalman filter; S4, constructing a threshold image according to a filtering updating result, establishing an available threshold image to be 60-600 frames, and resetting the Kalman filter to