Target measurement method and device based on dynamic updating of neural network weights

Abstract

The application provides a target measurement method and device based on neural network weight dynamic updating, and relates to the field of target measurement. The method comprises the following steps: pre-processing an original image to obtain a change feature of a centroid pixel value of the original image changing over time; performing feature extraction on the original image based on the change feature to obtain a dynamic target centroid signal of the original image, wherein the dynamic target centroid signal represents new target features after the original image changes; generating a target image according to the dynamic target centroid signal; inputting the target image into a first neural network for detection to obtain a detection result, wherein the first neural network is a neural network obtained by updating the weight of a second neural network according to the target image, and the second neural network is an initial neural network trained for detecting a target; and measuring the target position according to the detection result.

Description

Technical Field

[0001] This invention relates to the field of target measurement, and more specifically, to a target measurement method and apparatus based on dynamic updating of neural network weights. Background Technology

[0002] Currently, target detection algorithms are widely used in the field of optical sensors, primarily for capturing moving targets such as pedestrians, vehicles, and aircraft. They also have important applications in meteorological fields such as current velocity detection and water depth detection. Current target detection algorithms are mainly divided into traditional image processing methods and neural network methods. Traditional image processing methods for target detection mainly include edge detection, inter-frame differencing, color selection, and HOG (Histogram of Oriented Gridients) methods, but these algorithms have low detection accuracy and are insufficient to meet current needs. Neural network methods, on the other hand, offer higher detection accuracy and can meet various daily target detection requirements.

[0003] However, neural network methods have a significant drawback: they require pre-prepared training samples. Targets that differ significantly from the training samples cannot be recognized by the neural network. While targets in object detection scenarios may not change drastically in the short term, their various attributes evolve over time. Therefore, after this long evolutionary process, the detection accuracy of neural network methods decreases. Summary of the Invention

[0004] (a) Technical problems to be solved

[0005] To address the aforementioned issues, this invention provides a target measurement method and apparatus based on dynamic weight updates of a neural network. By extracting the centroid signal of the moving target as a training sample from the changing part of the target detection scene, the neural network is retrained and its weights are updated, enabling the neural network to change with the target and thus maintain stable detection accuracy.

[0006] (II) Technical Solution

[0007] One aspect of this invention provides a target measurement method based on dynamic weight update of a neural network, comprising: preprocessing an original image to obtain the change characteristics of the centroid pixel value of the original image over time; extracting features from the original image based on the change characteristics to obtain a moving target centroid signal of the original image, wherein the moving target centroid signal represents the new target features after the original image has changed; generating a target image based on the moving target centroid signal; inputting the target image into a first neural network for detection to obtain a detection result, wherein the first neural network is a neural network after weight update of a second neural network based on the target image, and the second neural network is a trained initial neural network for target detection; and measuring the target position based on the detection result.

[0008] In one embodiment of the present invention, preprocessing the original image to obtain the change characteristics of the centroid pixel value of the original image over time includes: performing differential processing on the original image to obtain a differential image; segmenting the differential image to obtain multiple segmented regions; calculating the centroid pixel value of each segmented region; and obtaining the change characteristics of the centroid pixel value of the original image over time based on the centroid pixel value of each segmented region.

[0009] In one embodiment of the present invention, the centroid pixel values ​​of multiple segmented regions are calculated as follows:

[0010]

[0011]

[0012] Among them, X m Represents the centroid position of the image width, Y m The height centroid of the image is represented by I, which represents the difference image. I(x, y) represents the pixel value of the pixel point with x as the horizontal coordinate and y as the vertical coordinate on the difference image.

[0013] In one embodiment of the present invention, preprocessing the original image to obtain the change characteristics of the centroid pixel value of the original image over time further includes: performing bandpass filtering on the change characteristics to obtain filtered change characteristics; and performing sinusoidal correlation processing on the filtered change characteristics to obtain sinusoidal correlation filtered change characteristics.

[0014] In one embodiment of the present invention, the calculation formula for bandpass filtering is as follows:

[0015] A M (t)=A(t)*M

[0016] Where A(t) represents the variation characteristic of the centroid pixel value, A M (t) represents the change characteristics of the centroid pixel value after filtering, where t represents time and M represents the bandpass filter convolution kernel.

[0017] In one embodiment of the present invention, the calculation formula for sinusoidal correlation processing is as follows:

[0018] A S (t)=A M (t)*SIN

[0019] Among them, A S (t) represents the change characteristics of the centroid pixel value after sinusoidal correlation filtering, and SIN represents the sinusoidal signal value within one period.

[0020] In one embodiment of the present invention, feature extraction of the original image based on change features to obtain the centroid signal of the moving target in the original image includes: setting a time threshold based on change features, wherein the change feature portion exceeding the time threshold is extracted as the centroid signal of the moving target.

[0021] In one embodiment of the present invention, a time threshold is set based on the change features. The extraction of the change feature portion exceeding the time threshold as the centroid signal of the moving target includes: determining whether the change feature portion exceeding the time threshold has changed based on the time threshold; if it has changed, the changed feature portion is used as the centroid signal of the moving target to generate a target image and input into a first neural network for detection; if it has not changed, the unchanged feature portion is directly used as the target image and input into a second neural network for detection.

[0022] In one embodiment of the present invention, before the target image is input into a first neural network for detection and the detection result is obtained, the method further includes: adjusting the size and angle of the target image, wherein the adjustment of size and angle includes scaling and rotation.

[0023] Another aspect of this invention provides a target measurement device based on dynamic weight update of a neural network, comprising: a preprocessing module for preprocessing an original image to obtain the change features of the centroid pixel value of the original image over time; an extraction module for extracting features from the original image based on the change features to obtain a moving target centroid signal of the original image, wherein the moving target centroid signal represents the new target features after the original image has changed; a generation module for generating a target image based on the moving target centroid signal; a detection module for inputting the target image into a first neural network for detection to obtain a detection result, wherein the first neural network is a neural network after weight update of a second neural network based on the target image, and the second neural network is a trained initial neural network for target detection; and a measurement module for measuring the target position based on the detection result.

[0024] (III) Beneficial Effects

[0025] The target measurement method and apparatus based on dynamic updating of neural network weights provided in this invention have at least the following beneficial effects:

[0026] The target measurement method and apparatus based on dynamic updating of neural network weights provided in this invention extracts the centroid signal of the moving target as a training sample from the part that changes in the target detection scene, retrains the neural network and updates the weights, so that the neural network can change with the target, thereby maintaining the stability of its own detection accuracy. Moreover, the three-dimensional spatial position of the target can be measured by combining the detected target with the digital model, thus realizing the three-dimensional measurement function of the target. Attached Figure Description

[0027] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0028] Figure 1 The flowchart illustrates a target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0029] Figure 2 The illustration shows a segmentation diagram of the difference image in the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0030] Figure 3 The diagram illustrates the structure of the neural network in the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0031] Figure 4 The schematic diagram illustrates the principle of the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0032] Figure 5 The diagram illustrates the structure of the target measurement device based on dynamic updating of neural network weights provided in an embodiment of the present invention. Detailed Implementation

[0033] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to specific embodiments and the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this invention. All other embodiments obtained by those skilled in the art based on the embodiments of this invention without creative effort are within the scope of protection of this invention. The terminology used herein is merely for describing specific embodiments and is not intended to limit the invention. The terms "comprising," "including," etc., used herein indicate the presence of the stated features, steps, operations, and / or components, but do not exclude the presence or addition of one or more other features, steps, operations, or components.

[0034] In this invention, unless otherwise explicitly specified and limited, the terms "installation," "connection," "linking," and "fixing," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral part; they can refer to a mechanical connection, an electrical connection, or a connection that allows communication between them; they can refer to a direct connection or an indirect connection through an intermediate medium; they can refer to the internal communication of two components or the interaction between two components. Those skilled in the art can understand the specific meaning of the above terms in this invention according to the specific circumstances.

[0035] In the description of this invention, it should be understood that the terms "longitudinal", "length", "circumferential", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the subsystem or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0036] Throughout the accompanying drawings, identical elements are represented by the same or similar reference numerals. Conventional structures or configurations may be omitted where they might cause confusion in understanding the invention. Furthermore, the shapes, sizes, and positional relationships of the components in the drawings do not reflect actual size, scale, or actual positional relationships. Additionally, any reference numerals placed between parentheses in the claims should not be construed as limiting the claims.

[0037] Similarly, to simplify the invention and aid in understanding one or more aspects of the invention, in the above description of exemplary embodiments of the invention, various features of the invention are sometimes grouped together in a single embodiment, figure, or description thereof. The use of terms such as "one embodiment," "some embodiments," "example," "specific example," or "some examples" indicates that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0038] Furthermore, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features indicated. Therefore, a feature defined as "first" or "second" may explicitly or implicitly include one or more of that feature. In the description of this invention, "a plurality of" means at least two, such as two, three, etc., unless otherwise explicitly specified.

[0039] Figure 1 The flowchart illustrates a target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0040] like Figure 1 As shown, the target measurement method based on dynamic updating of neural network weights provided in this embodiment of the invention may include:

[0041] S1, preprocess the original image to obtain the change characteristics of the centroid pixel value of the original image over time.

[0042] The preprocessing of the original image to obtain the characteristics of the centroid pixel value changing over time includes:

[0043] S100: Perform differential processing on the original image to obtain a differential image.

[0044] S101, segment the difference image to obtain multiple segmented regions.

[0045] S102, calculate the centroid pixel value of each segmented region.

[0046] S103, based on the centroid pixel values ​​of each segmented region, obtain the change characteristics of the centroid pixel values ​​of the original image over time.

[0047] Based on the above embodiments, preprocessing the original image to obtain the change characteristics of the centroid pixel value of the original image over time also includes:

[0048] S104 performs bandpass filtering on the changing features to obtain the filtered changing features.

[0049] The calculation formula for bandpass filtering is as follows:

[0050] A M (t)=A(t)*M

[0051] Where A(t) represents the variation characteristic of the centroid pixel value, A M (t) represents the change characteristics of the centroid pixel value after filtering, where t represents time and M represents the bandpass filter convolution kernel. M is usually a 32-element array generated by the MATLAB toolbox fdatool.

[0052] S105, perform sinusoidal correlation processing on the filtered variation characteristics to obtain the variation characteristics after sinusoidal correlation filtering.

[0053] The calculation formula for sinusoidal correlation processing is as follows:

[0054] A s (t)=A M (t)*SIN

[0055] Among them, A s (t) represents the change characteristics of the centroid pixel value after sinusoidal correlation filtering, and SIN represents the sinusoidal signal value within one period.

[0056] S2, based on the change characteristics, feature extraction is performed on the original image to obtain the centroid signal of the moving target in the original image, wherein the centroid signal of the moving target represents the new target features after the original image has changed.

[0057] The process of extracting features from the original image based on change characteristics to obtain the centroid signal of the moving target in the original image includes: setting a time threshold T1 based on the change characteristics, wherein the portion of the change characteristics exceeding the time threshold T1 is extracted as the centroid signal of the moving target. Specifically, it can be: based on the time threshold T1, determining whether the portion of the change characteristics exceeding the time threshold T1 has changed; if it has changed, the changed portion of the change characteristics is used as the centroid signal of the moving target to generate a target image and input into the first neural network for detection; if it has not changed, the unchanged portion of the change characteristics is directly used to generate a target image and input into the second neural network for detection.

[0058] S3 generates a target image based on the centroid signal of the moving target.

[0059] S4, Input the target image into the first neural network for detection and obtain the detection result. The first neural network is a neural network after updating the weights of the second neural network based on the target image. The second neural network is a trained initial neural network used for target detection. Measure the target position based on the detection result.

[0060] Before inputting the target image into the first neural network for detection and obtaining the detection result, the process also includes:

[0061] The target image is resized and its angle is adjusted, including scaling and rotation.

[0062] S5, Measure the target position based on the detection results.

[0063] The measurement of target location based on detection results includes:

[0064] Digital models are used to obtain the three-dimensional spatial information of a target. Specifically, after the digital model undergoes virtual-real fusion, a mapping relationship is established between two-dimensional image coordinates and three-dimensional spatial coordinates, thus allowing the acquisition of the three-dimensional spatial position. In multi-camera scenarios, multiple cameras combined with the digital model can establish spatial straight lines; finding the spatial intersections of these lines also yields the three-dimensional spatial position.

[0065] The present invention provides a target measurement method based on dynamic updating of neural network weights. By extracting the changing part of the target detection scene, namely the centroid signal of the moving target, as a training sample, the neural network is retrained and the weights are updated, so that the neural network can change with the target and thus maintain the stability of its detection accuracy.

[0066] Figure 2 The illustration shows a segmentation diagram of the difference image in the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0067] like Figure 2 In the target measurement method based on dynamic updating of neural network weights provided in the embodiments of the present invention, the differential image is divided into m*n segmentation regions, and the centroid pixel value X of each m*n segmentation region is calculated. The change feature A(t) of the centroid pixel value of each segmentation region is plotted in time sequence. In this embodiment, the change feature can be represented by the centroid pixel value change curve of the centroid pixel value changing with time.

[0068] The centroid pixel value of the segmented region is calculated as follows:

[0069]

[0070]

[0071] Among them, X m Represents the centroid position of the image width, Y m The height centroid of the image is represented by I, which represents the difference image. I(x, y) represents the pixel value of the pixel point with x as the horizontal coordinate and y as the vertical coordinate on the difference image.

[0072] Figure 3 The diagram illustrates the structure of the neural network in the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0073] like Figure 3 As shown, the structure of the neural network in the target measurement method based on dynamic updating of neural network weights provided in this embodiment of the invention may include: an FPN (Feature Pyramid) part, an RPN (Region Candidate Network) part, a ROIalign (Feature Fusion) part, and an FCN (Fully Connected Network) part.

[0074] The FPN part can utilize both the spatial information of the low-level feature map and the semantic information of the high-level feature map. The principle is simple: first, the high-level features with smaller resolution are reduced in dimensionality by 1*1 convolution (to reduce the amount of computation), then upsampled to the same size as the previous feature map, and then added element by element to obtain the fused features.

[0075] The RPN part generates several fixed-size bounding boxes on the features provided by FPN. In this embodiment, all bounding boxes undergo a classification task and a regression task. The classification task is simple: determine whether an object is inside the bounding box, and what the probability of an object being inside is. The regression task generates the coordinates (top left, bottom right) of each bounding box. After obtaining the score and coordinates of each bounding box, their scores are sorted. In this embodiment, a score greater than 0.7 indicates a positive example, and a score less than 0.3 indicates a negative example. Bounding boxes that are too small or too large (outside the feature map range) are also filtered out.

[0076] The ROIalign part pools all ROIs into feature maps of the same size before reshaping them into a one-dimensional vector, thus completing the subsequent classification and regression tasks.

[0077] The FCN part employs a two-branch structure. One branch performs region detection and outputs the masked region structure. The other branch performs classification and regression to obtain the target category and confidence rate.

[0078] The Mask R-CNN neural network provided in this embodiment is a neural network with updated weights based on the target image. The training process of this weight-updated neural network includes:

[0079] Obtain k original images, where 100 < k < 5000.

[0080] The centroid pixel value change characteristics after sinusoidal correlation filtering are obtained using the method in steps S100 to S105 of this embodiment.

[0081] The centroid signal of the moving target is obtained by using the method in step S2 of this embodiment. A target image sample is generated based on the centroid signal of the moving target. The Mask R-CNN neural network is trained using the target image sample to obtain neural network weights that can be used for target detection.

[0082] The next step is to obtain the original image of frame k+1. Repeat steps S1 to S3 of this embodiment to obtain a new training sample of frame k+1, which is named the questionable training sample. However, the questionable training sample cannot be directly added to the training sample of the neural network. It will be processed later.

[0083] Finally, a neural network is used to perform object detection on the (k+1)th frame of the image to obtain the object detection results. Suspected training samples with the same target as those in the neural network's detection results do not need to be added to the neural network's training samples; only different suspected training samples need to be added. If the neural network's training samples have new terms, the neural network is trained based on the existing weights to obtain new weights, thus achieving weight updates. If the neural network's training samples have no new terms, no training is required.

[0084] The neural network in this implementation is not limited to the Mask R-CNN neural network, but can be any kind of target detection neural network, such as the YOLOV1-V5 neural network or the Fast R-CNN neural network.

[0085] One application scenario of the target measurement method based on dynamic updating of neural network weights provided in this invention is flow velocity measurement, where the floating object is the moving target.

[0086] Figure 4 The schematic diagram illustrates the principle of the target measurement method based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0087] like Figure 4 As shown, the principle of the target measurement method based on dynamic updating of neural network weights provided in this embodiment of the invention is as follows:

[0088] The centroid signal of the moving target is extracted from the acquired raw image. A target image sample is generated based on the extracted centroid signal. This target image sample is used to train a neural network and update its weights. The image is then detected to obtain the detection result. The target position is measured using a digital model based on this detection result. If the centroid signal of the moving target is not extracted, the acquired target image is directly fed into the neural network for detection, and the output result is the target position. Combined with the digital model, the three-dimensional information of the target can be measured. The digital model is provided by Funeng Sensor.

[0089] Figure 5 The diagram illustrates the structure of the target measurement device based on dynamic updating of neural network weights provided in an embodiment of the present invention.

[0090] like Figure 5 As shown, the target measurement device based on dynamic updating of neural network weights provided in this embodiment of the invention may include: a preprocessing module 501, an extraction module 502, a generation module 503, a detection module 504, and a measurement module 505.

[0091] The preprocessing module 501 is used to preprocess the original image to obtain the change characteristics of the centroid pixel value of the original image over time.

[0092] The extraction module 502 is used to extract features from the original image based on the change features to obtain the centroid signal of the moving target in the original image, wherein the centroid signal of the moving target represents the new target features after the original image has changed.

[0093] The generation module 503 is used to generate a target image based on the centroid signal of the moving target.

[0094] The detection module 504 is used to input the target image into the first neural network for detection and obtain the detection result. The first neural network is a neural network after updating the weights of the second neural network according to the target image. The second neural network is a trained initial neural network used for target detection.

[0095] The measurement module 505 is used to measure the target position based on the detection results.

[0096] According to embodiments of the present invention, any plurality of modules among the preprocessing module 501, extraction module 502, generation module 503, detection module 504, and measurement module 505 may be combined into one module, or any one of these modules may be split into multiple modules, or at least part of the functionality of one or more of these modules may be combined with at least part of the functionality of other modules and implemented in one module. According to embodiments of the present disclosure, at least one of the preprocessing module 501, extraction module 502, generation module 503, detection module 504, and measurement module 505 may be at least partially implemented as hardware circuitry, such as a field-programmable gate array (FPGA), a programmable logic array (PLA), a system-on-a-chip, a system-on-a-substrate, a system-on-package, an application-specific integrated circuit (ASIC), or implemented in hardware or firmware by any other reasonable means of integrating or packaging the circuitry, or implemented in software, hardware, or firmware, or in any suitable combination of any of these three implementation methods. Alternatively, at least one of the preprocessing module, extraction module, generation module, detection module, and measurement module can be at least partially implemented as a computer program module, which can perform corresponding functions when the computer program module is run.

[0097] It should be noted that the target measurement device based on dynamic updating of neural network weights in the embodiments of the present invention corresponds to the target measurement method based on dynamic updating of neural network weights in the embodiments of the present invention, and their specific implementation details and the resulting technical effects are the same, which will not be repeated here.

[0098] Although the invention has been illustrated and described in detail in the accompanying drawings and the foregoing description, such illustrations and descriptions should be considered illustrative or exemplary rather than restrictive.

[0099] Those skilled in the art will understand that the features described in the various embodiments and / or claims of the present invention can be combined and / or combined in various ways, even if such combinations or combinations are not explicitly described in the present invention. In particular, the features described in the various embodiments and / or claims of the present invention can be combined and / or combined in various ways without departing from the spirit and teachings of the present invention. All such combinations and / or combinations fall within the scope of the present invention.

[0100] Although the invention has been shown and described with reference to specific exemplary embodiments thereof, those skilled in the art will understand that various changes in form and detail may be made to the invention without departing from the spirit and scope of the invention as defined by the appended claims and their equivalents. Therefore, the scope of the invention should not be limited to the above embodiments, but should be determined not only by the appended claims but also by their equivalents.

Claims

1. A target measurement method based on dynamic weight update of a neural network, characterized in that, include: Preprocessing the original image to obtain the change characteristics of the centroid pixel value of the original image over time includes: performing differential processing on the original image to obtain a differential image; segmenting the differential image to obtain multiple segmented regions; calculating the centroid pixel value of each segmented region; and obtaining the change characteristics of the centroid pixel value of the original image over time based on the centroid pixel value of each segmented region. The centroid pixel values ​​of the multiple segmented regions are calculated as follows: in, Represents the centroid position of the image width. I represents the height centroid location of the image, and I represents the difference image. This represents the pixel value corresponding to a pixel point with x-coordinate and y-coordinate on the difference image. Based on the aforementioned change features, feature extraction is performed on the original image to obtain the moving target centroid signal of the original image. This includes: setting a time threshold based on the change features, wherein extracting the change feature portion exceeding the time threshold as the moving target centroid signal includes: determining, based on the time threshold, whether the change feature portion exceeding the time threshold has changed; if it has changed, generating a target image using the changed feature portion as the moving target centroid signal and inputting it into a first neural network for detection; if it has not changed, directly generating a target image using the unchanged feature portion and inputting it into a second neural network for detection, wherein the moving target centroid signal represents the new target features after the original image has changed. Generate a target image based on the moving target centroid signal; The target image is input into a first neural network for detection to obtain a detection result. The first neural network is a neural network with weights updated for a second neural network based on the target image. The second neural network is a trained initial neural network used for target detection. The target location is measured based on the detection results.

2. The target measurement method based on dynamic updating of neural network weights according to claim 1, characterized in that, The preprocessing of the original image to obtain the change characteristics of the centroid pixel value of the original image over time also includes: The changing features are subjected to bandpass filtering to obtain the filtered changing features; The filtered variation features are subjected to sinusoidal correlation processing to obtain the variation features after sinusoidal correlation filtering.

3. The target measurement method based on dynamic updating of neural network weights according to claim 2, characterized in that, The calculation formula for the bandpass filtering process is as follows: in, The characteristic representing the change in centroid pixel value, The value represents the change in centroid pixel value after filtering, where t represents time and M represents the bandpass filter convolution kernel.

4. The target measurement method based on dynamic updating of neural network weights according to claim 3, characterized in that, The calculation formula for the sinusoidal correlation processing is as follows: in, This represents the change characteristics of the centroid pixel value after sinusoidal correlation filtering. It represents the sinusoidal signal value within one cycle.

5. The target measurement method based on dynamic updating of neural network weights according to claim 1, characterized in that, Before inputting the target image into the first neural network for detection and obtaining the detection result, the method further includes: The target image is subjected to size and angle adjustment processing, wherein the size and angle adjustment processing includes scaling processing and rotation processing.

6. A target measurement device based on dynamic weight update of a neural network, characterized in that, include: The preprocessing module is used to preprocess the original image to obtain the change characteristics of the centroid pixel value of the original image over time, including: performing differential processing on the original image to obtain a differential image; segmenting the differential image to obtain multiple segmented regions; calculating the centroid pixel value of each segmented region; and obtaining the change characteristics of the centroid pixel value of the original image over time based on the centroid pixel value of each segmented region. The centroid pixel values ​​of the multiple segmented regions are calculated as follows: in, Represents the centroid position of the image width. I represents the height centroid location of the image, and I represents the difference image. This represents the pixel value corresponding to a pixel point with x-coordinate and y-coordinate on the difference image. An extraction module is used to extract features from the original image based on the change features to obtain the moving target centroid signal of the original image. This includes: setting a time threshold based on the change features; wherein extracting the change feature portion exceeding the time threshold as the moving target centroid signal includes: determining whether the change feature portion exceeding the time threshold has changed based on the time threshold; if it has changed, generating a target image using the changed feature portion as the moving target centroid signal and inputting it into a first neural network for detection; if it has not changed, directly generating a target image using the unchanged feature portion and inputting it into a second neural network for detection, wherein the moving target centroid signal represents the new target features after the original image has changed. The generation module is used to generate a target image based on the centroid signal of the moving target; The detection module is used to input the target image into a first neural network for detection and obtain a detection result. The first neural network is a neural network after updating the weights of a second neural network based on the target image. The second neural network is a trained initial neural network for detecting the target. The measurement module is used to measure the target position based on the detection results.

Images