Flying target detection method based on multi-directional filtering
Through multi-directional filtering and clustering analysis technology, the problems of weak brightness and noise in infrared target detection are solved, and the accurate identification and position determination of flight targets are achieved, which improves detection efficiency and accuracy.
Patent Information
- Application Number
- CN202410347118.5
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2024-03-26
- Publication Date
- 2024-11-05
- Estimated Expiration
- 2044-03-26
AI Technical Summary
Infrared target detection is affected by atmospheric refraction and scattering in long distances and complex scenarios. The target brightness is weak and the internal noise of the imaging system is large, making it difficult to distinguish small targets, especially in occlusion.
Using a flight object detection method based on multi-directional filtering, filtering processing, clustering analysis and motion region determination by acquiring infrared image sequence data, combining Sobel edge detection operator and DBSCAN clustering algorithm, clutter is filtered and target features are extracted to reduce The impact of noise improves the accuracy of object detection.
It effectively reduces high-brightness noise in infrared images, enhances target brightness value, improves the accuracy and efficiency of small-object detection, reduces background noise interference, and can accurately identify the flight target position in complex scenarios.
Smart Images

Figure CN118072224B_ABST
Abstract
Description
Technical Field
[0001] The invention relates to the technical field of infrared target detection, monitoring and data processing, and more specifically to a flying target detection method based on multi-directional filtering. Background Art
[0002] In recent years, with the continuous improvement of thermal imaging technology and the gradual development of aerospace remote sensing and other fields, thermal infrared imaging has been more closely integrated with target detection and monitoring, and infrared target detection has therefore played an important role in monitoring, security and other fields. However, infrared imaging is generally far away, and in the system output image, the target often only occupies a few pixels, so it is impossible to obtain information such as the texture, structure and color of the target. At the same time, during the imaging process, the target brightness is also relatively weak due to the influence of atmospheric refraction and scattering. In addition, it may also cause random electrical noise during operation due to defects in the internal components of the imaging system, and ultimately present high-brightness noise points on the infrared image, making small targets even more difficult to distinguish. Finally, in various real complex scenes, there may be problems such as occlusion, which adds considerable difficulty to infrared weak target detection and monitoring. Summary of the invention
[0003] Based on the above technical problems, the main purpose of the present invention is to propose a flying target detection method based on multi-directional filtering to solve at least one of the above technical problems.
[0004] The present invention provides a flying target detection method based on multi-directional filtering, comprising: obtaining N frames of infrared image data of a flying target to be detected, wherein the infrared image data includes the flying target to be detected; filtering each frame of infrared image data to obtain N frames of filtered images, wherein each frame of the infrared image data obtains a corresponding filtered image, and the filtering processing is used to filter the clutter of the infrared image data; two-dimensionally superimposing the N frames of filtered images to obtain a frame of feature image; clustering analysis of the feature image based on a clustering algorithm to obtain a clustering result, wherein the clustering result includes feature points of the flying target in the feature image; determining a motion area of the flying target in the feature image based on the clustering result; and detecting the N frames of filtered images based on the motion area to obtain the position of the flying target in the N frames of the infrared image sequence, wherein N is an integer greater than or equal to 2.
[0005] In some embodiments, N frames of infrared image data are sorted based on a time series, and filtering is performed on each frame of infrared image data, including: performing multi-directional filtering on each frame of infrared image data based on a Sobel edge detection operator to obtain a multi-directional filtering feature map; and generating a filtered image based on the multi-directional filtering feature map.
[0006] In some embodiments, the Sobel edge detection operator includes filter kernels in 8 directions, and the size of the filter kernels in the 8 directions is 5×5. Multi-directional filtering of each frame of infrared image data based on the Sobel edge detection operator includes: filtering each frame of infrared image data in 8 directions to obtain 8 filter feature maps; filtering and fusing the 8 filter feature maps to obtain a fused image; and threshold segmenting the fused image to obtain a filtered image.
[0007] In some embodiments, filtering and fusing 8 filter feature maps includes: generating a minimum filter map based on the 8 filter feature maps, wherein the size of the minimum filter map is the same as the size of the 8 filter feature maps, and the pixel value of each pixel coordinate of the minimum filter map is the minimum pixel value of the corresponding pixel coordinates in the 8 filter feature maps; superimposing the pixel values of the same pixel coordinates of the 8 filter feature maps to obtain a filter feature map; and performing dot multiplication and normalization on the minimum filter map and the filter feature map to obtain a fused image.
[0008] In some embodiments, performing cluster analysis on the feature image based on a clustering algorithm includes: performing cluster analysis on the feature image based on a DBSCAN clustering algorithm to obtain a clustering result.
[0009] In some embodiments, determining the motion area of the flying target in the feature image based on the clustering result includes: calculating the motion distance of the flying target in the feature image; and determining the motion area of the flying target in the feature image based on the flight distance.
[0010] In some embodiments, calculating the movement distance of the flying target in the feature image includes: determining the coordinate values of each category in the feature image based on the clustering results; determining the flight distance of the flying target based on the coordinate values of each category; and determining the movement area of the flying target based on the flight distance.
[0011] In some embodiments, determining the flight distance of a flying target based on coordinate values of each category includes: sorting all coordinate values in each category based on size, and obtaining the two farthest points in each category based on the sorting result, obtaining the farthest distance of each category; and taking the farthest distance with the largest length in all categories as the flight distance of the flying target.
[0012] In some embodiments, before two-dimensionally superimposing N frames of filtered images to obtain a frame of feature image, it also includes sorting the N frames of filtered images based on the time series of N frames of sequential infrared images corresponding to the N frames of filtered images; and stacking the sorted N frames of filtered images into a third-order tensor, and establishing a layer index of the third-order tensor.
[0013] In some embodiments, a third-order tensor is traversed and searched layer by layer to query the position of the flying target in each frame of the filtered image, and the traversed and searched area is the moving area of the flying target; and the correspondence between the filtered image and the coordinates of the flying target is obtained based on the layer index. BRIEF DESCRIPTION OF THE DRAWINGS
[0014] Figure 1 A flow chart of a flying target detection method based on multi-directional filtering according to an embodiment of the present invention is schematically shown. DETAILED DESCRIPTION
[0015] In order to make the objectives, technical solutions and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with specific embodiments and with reference to the accompanying drawings.
[0016] Figure 1 A flowchart of a flying target detection method based on multi-directional filtering according to an embodiment of the present invention is schematically shown. Figure 1 As shown, the method includes operations S110 to S160.
[0017] In operation S110, N frames of sequence infrared image data of a flying target to be detected are acquired, where the infrared image data includes the flying target to be detected.
[0018] According to an embodiment of the present invention, N frames of sequential infrared image data are sorted based on the order of time.
[0019] In operation S120, filtering is performed on each frame of infrared image data to obtain N frames of filtered images.
[0020] In some embodiments, each frame of sequential infrared image data obtains a corresponding filtered image.
[0021] In some embodiments, the filtering process includes performing multi-directional filtering on each frame of infrared image data based on a Sobel edge detection operator to obtain a multi-directional filtering feature map; and generating a filtered image based on the multi-directional filtering feature map.
[0022] In some embodiments, the Sobel operator is simple and efficient in detection. Specifically, small infrared targets will not be filtered out by edge detection because of their point-like features, and large areas of images can be quickly and efficiently filtered out through edge detection algorithms, laying a good foundation for subsequent clustering processing.
[0023] In some embodiments, the Sobel operator includes filter kernels in 8 directions, and the size of the filter kernels in the 8 directions is 5×5. For example, they can be divided into 8 directions of 0°, 22.5°, 45°, 67.5°, 90°, 112.50°, 135°, and 157.50° in a clockwise direction.
[0024] In some embodiments, increasing the filtering angle can filter out as much background clutter as possible, and the center pixel weight of the filter kernel 5×5 is set to be greater than 0, and correction factors are established in each direction so that the sum of the filter kernel coefficients is zero, so that the filter kernel can make good use of the point-like characteristics of small infrared targets, increase the output value of the target, thereby reducing the difficulty of subsequent threshold segmentation and improving the efficiency and accuracy of image filtering.
[0025] In some embodiments, filtering is performed in eight directions on each frame of infrared image data according to filtering kernels in eight directions to obtain eight filtering feature maps.
[0026] In some embodiments, 8 filter feature maps are filtered and fused to obtain a fused image, including generating a minimum filter map based on the 8 filter feature maps, wherein the size of the minimum filter map is the same as the size of the 8 filter feature maps, and the pixel value of each pixel coordinate of the minimum filter map is the minimum pixel value of the corresponding pixel coordinates in the 8 filter feature maps.
[0027] And the pixel values of the same pixel coordinates of the 8 filter feature maps are superimposed to obtain the filter feature map, and the negative pixel values are set to 0.
[0028] In some embodiments, the minimum filter map and the filter feature map are dot-multiplied and normalized to achieve filter fusion of the eight filter feature maps to obtain a fused image.
[0029] In some embodiments, filter fusion can enhance the brightness value of the target without increasing background clutter, thereby suppressing background noise.
[0030] In some embodiments, the fused image is segmented by threshold value to obtain a filtered image. Because the filter fusion has increased the brightness value of the target, a lower threshold value can be selected for segmentation to obtain a better segmentation result.
[0031] In operation S130, N frames of filtered images are two-dimensionally superimposed to obtain a frame of feature image.
[0032] In some embodiments, because each frame of infrared image corresponds to a filtered image, the N frames of filtered images can be sorted based on the time sequence of the N frames of sequential infrared images corresponding to the N frames of filtered images.
[0033] In some embodiments, after sorting, an index relationship is established between each frame of filtered image and the time series corresponding to the filtered image, that is, N frames of filtered images are stacked into a third-order tensor in time sequence, and a layer index is established to associate the relationship between the infrared image and the target position in subsequent target detection.
[0034] In some embodiments, because the flying target is in a scene with a fixed background or slight shaking, the flying area will form a nearly linear area in the feature image, while the background noise will form a dot-shaped area.
[0035] In operation S140, cluster analysis is performed on the feature image based on a clustering algorithm to obtain a clustering result.
[0036] In some embodiments, a DBSCAN clustering algorithm is used to perform clustering analysis on the feature image to obtain a clustering result. The DBSCAN clustering algorithm has the advantages of not requiring the number of clusters to be known in advance, being able to find clusters of any shape, and being insensitive to noise.
[0037] In operation S150, a motion region of the flying object in the feature image is determined based on the clustering result.
[0038] In some embodiments, it is necessary to calculate the movement distance of the flying target in the feature image; and determine the movement area of the flying target in the feature image based on the flight distance.
[0039] In some embodiments, the coordinate values of each category in the feature image are first determined based on the clustering results, and all coordinate values in each category are sorted based on size, and based on the sorting results, the two points farthest apart in each category are obtained, and the distance between the first and last points of each category is obtained.
[0040] In some embodiments, because the flying area will form a nearly linear area in the feature image, and the background noise will form a point-like area, the longest distance with the largest length in all categories is taken as the flight distance of the flying target.
[0041] In some embodiments, in the sorting algorithm, the theoretical longest distance of the target trajectory cluster is L Tg_theory , the longest distance of the target trajectory cluster under this sorting algorithm is L Tg_sort , the theoretical maximum distance of the longest clutter in all clutter clusters is L Clu_theory , the longest distance of the longest clutter under this sorting algorithm is L Clu_sort , there is an inequality:
[0042]
[0043] In some embodiments, the two farthest points of the flying target are determined as diagonals to delineate a rectangular area to obtain the moving area of the flying target, and the distance between the two farthest points is the flight distance of the target.
[0044] In some embodiments, in the process of extracting the target flight trajectory, the three-dimensional full-space-time search or frame-by-frame differential calculation method is abandoned, and the two-dimensional plane clustering algorithm and geometric algorithm are used instead, which greatly reduces the amount of calculation and memory space in the detection process, improves the efficiency of target detection, and reduces costs.
[0045] In some embodiments, clustering and farthest distance screening strategies are adopted, multiple images are merged into one, and unified calculations are performed to directly extract the target area, eliminating a large number of repeated redundant calculations such as noise point identification in non-target areas of each frame, thereby further improving the efficiency of the target detection algorithm.
[0046] In operation S160, the N frames of filtered images are detected based on the motion region to obtain the position of the flying target in the N frames of sequential infrared images.
[0047] In some embodiments, a third-order tensor is traversed and searched layer by layer to query the position of the flying target in each frame of the filtered image, and the traversed and searched area is the moving area of the flying target; and the correspondence between the coordinates of the filtered image and the flying target is obtained based on the layer index, thereby obtaining a correspondence table between the filtered image and the target position.
[0048] The above specific embodiments further illustrate the purpose, technical solutions and beneficial effects of the present invention in detail. It should be understood that the above are only specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc. made within the spirit and principles of the present invention should be included in the protection scope of the present invention.
Claims
1. A flying target detection method based on multi-directional filtering, characterized in that: include: Acquire N frames of sequence infrared image data of the flying target to be detected, wherein the infrared image data includes the flying target to be detected; Performing filtering processing on each frame of the infrared image data to obtain N frames of filtered images, wherein each frame of the sequence of infrared image data obtains a corresponding frame of the filtered image, and the filtering processing is used to filter the clutter of the infrared image data; the filtering processing includes: performing multi-directional filtering on each frame of the infrared image data based on the Sobel edge detection operator to obtain filtering feature maps corresponding to multiple directions respectively; performing filtering fusion on multiple filtering feature maps to obtain a fused image, and performing threshold segmentation on the fused image to obtain the filtered image; Performing two-dimensional superposition on the N frames of filtered images to obtain a frame of feature image; Performing cluster analysis on the feature image based on a clustering algorithm to obtain a clustering result, wherein the clustering result includes feature points of the flying target in the feature image; determining a motion region of the flying target in the feature image based on the clustering result; and The N frames of filtered images are detected based on the motion area to obtain the position of the flying target in the N frames of sequential infrared images, where N is an integer greater than or equal to 2.
2. The method according to claim 1, characterized in that The N frames of sequential infrared image data are sorted based on a time series.
3. The method according to claim 2, characterized in that The Sobel edge detection operator includes filter kernels in 8 directions, the size of the filter kernels in the 8 directions is 5×5, and the multi-directional filtering of each frame of the infrared image data based on the Sobel edge detection operator includes: Perform filtering processing in 8 directions on each frame of infrared image data to obtain 8 filtering feature maps; The eight filtering feature maps are filtered and fused to obtain a fused image; and the fused image is threshold segmented to obtain the filtered image.
4. The method according to claim 3, characterized in that The filtering and fusing of the eight filtering feature maps comprises: Generate a minimum filter map based on the eight filter feature maps, wherein the size of the minimum filter map is the same as the size of the eight filter feature maps, and the pixel value of each pixel coordinate of the minimum filter map is the minimum pixel value of the corresponding pixel coordinate in the eight filter feature maps; Superimposing the pixel values of the same pixel coordinates of the eight filtering feature maps to obtain a filtering feature map; and The minimum filter image and the sum filter feature image are point-multiplied and normalized to obtain the fused image.
5. The method according to claim 1, characterized in that The clustering analysis of the feature image based on the clustering algorithm includes: The characteristic image is clustered and analyzed based on the DBSCAN clustering algorithm to obtain the clustering result.
6. The method according to claim 1, characterized in that The determining the motion area of the flying target in the feature image based on the clustering result comprises: Calculating the flight distance of the flying target in the feature image; and A motion region of the flying object in the feature image is determined based on the flying distance.
7. The method according to claim 6, characterized in that The calculating the flight distance of the flying target in the feature image comprises: Determine the coordinate values of each category in the feature image based on the clustering result; determining the flight distance of the flying target based on the coordinate values of each category; and A movement area of the flying object is determined based on the flying distance.
8. The method according to claim 7, characterized in that The determining the flight distance of the flying target based on the coordinate values of each category includes: Sorting all coordinate values in each category based on size, and obtaining the two points in each category that are the farthest apart based on the sorting result, thereby obtaining the farthest distance of each category; and The longest distance among all categories is taken as the flight distance of the flight target.
9. The method according to claim 1, before performing two-dimensional superposition on the N frames of filtered images to obtain a frame of feature image, further comprises: Sorting the N frames of filtered images based on a time sequence of N frames of sequential infrared images corresponding to the N frames of filtered images; and The sorted N frames of filtered images are stacked into a third-order tensor, and a layer index of the third-order tensor is established.
10. The method according to claim 9, characterized in that The detecting the N frames of filtered images based on the motion region to obtain the position of the flying target in the N frames of sequential infrared images comprises: Performing a traversal search on the third-order tensor layer by layer to query the position of the flying target in each frame of the filtered image, wherein the traversal search area is the moving area of the flying target; and The corresponding relationship between the filtered image and the coordinates of the flying target is obtained based on the layer index.
Citation Information
Patent Citations
Image processing device and method
CN107123129A
Texture feature extraction method, texture feature extraction device and terminal equipment
CN110473242A
Infrared image enhancement method based on 3D filtering
CN111899200A
Infrared small target detection method based on multidirectional fusion
CN115147613A
Infrared moving target detection method and device
CN116228819A
