Target detection method and apparatus

Abstract

The application relates to a target detection method and device, and relates to the technical field of computer vision. The method is executed by a computer device. The computer device acquires detection frame data of a detection target, a camera image, and camera extrinsic parameter data. The detection frame data comprises target confidence and detection frame position coordinates, and the camera image is an image containing the detection target. According to the detection frame data, the camera image, and the camera extrinsic parameter data, modified detection frame data subjected to geometric prior correction is acquired. An enhancement operation is performed on a local image corresponding to the modified detection frame data in the camera image. The enhancement operation comprises brightness entropy screening and background suppression enhancement. According to the modified detection frame data, a deep abnormality mask compensation operation is performed on the local image subjected to the enhancement operation, so as to generate target detection frame data, thereby improving the accuracy of target detection.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, and in particular to a target detection method and device. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, mobile robots (such as service robots, unmanned delivery vehicles, and cleaning robots) are increasingly being used in commercial and home scenarios. In practical applications, it is particularly important for mobile robots to achieve safe and reliable autonomous navigation and near-field obstacle avoidance based on target detection results.

[0003] In related technologies, object detection primarily relies on deep learning-based 2D (2D) image detection frameworks. These methods typically involve computer devices directly extracting features from camera images using convolutional neural networks to generate detection boxes containing category and location coordinates. To improve detection performance, existing techniques often focus on optimizing the depth and width of the network structure, or training on large-scale datasets, striving to achieve object and background classification and regression at the image pixel level.

[0004] However, the above scheme treats object detection as a pure image classification problem, ignoring the spatial geometric relationship between the object and the ground plane in the physical world, as well as the ability of depth information to represent the object's structure, thus reducing the accuracy of the robot's object detection. Summary of the Invention

[0005] This invention provides a target detection method and device that can improve the accuracy of target detection. The technical solution is as follows: On the one hand, a target detection method is provided, the method comprising: Acquire detection bounding box data of the target, camera image, and camera extrinsic data. The detection bounding box data includes the target confidence score and the detection bounding box position coordinates. The camera image is an image containing the target. Based on the detection box data, the camera image, and the camera extrinsic data, obtain the corrected detection box data after geometric prior correction; An enhancement operation is performed on the local image corresponding to the corrected detection box data in the camera image. The enhancement operation includes brightness entropy screening and background suppression enhancement. Based on the corrected detection box data, a depth anomaly mask compensation operation is performed on the local image after the enhancement operation to generate target detection box data.

[0006] On the other hand, a target detection device is provided, the device comprising: The basic data acquisition module is used to acquire the detection box data of the detected target, the camera image, and the camera extrinsic data. The detection box data includes the target confidence score and the detection box position coordinates, and the camera image is an image containing the detected target. The modified detection box data acquisition module is used to acquire modified detection box data after geometric prior correction based on the detection box data, the camera image, and the camera extrinsic data. An enhancement operation execution module is used to perform enhancement operations on the local image corresponding to the corrected detection box data in the camera image. The enhancement operations include brightness entropy screening and background suppression enhancement. The mask compensation execution module is used to perform a depth anomaly mask compensation operation on the local image after the enhancement operation is performed, based on the corrected detection box data, to generate target detection box data.

[0007] In another aspect, a computer device is provided, the computer device comprising a processor and a memory, the memory storing at least one instruction, at least one program, code set or instruction set, the at least one instruction, the at least one program, the code set or instruction set being loaded and executed by the processor to implement the target detection method as described above.

[0008] In another aspect, a computer-readable storage medium is provided, wherein at least one instruction, at least one program, code set, or instruction set is stored therein, wherein the at least one instruction, the at least one program, the code set, or the instruction set is loaded and executed by a processor to implement the target detection method described above.

[0009] In another aspect, a computer program product is provided, comprising a computer program stored in a computer-readable storage medium. A processor of a computer device reads the computer program from the computer-readable storage medium and executes the computer program, causing the computer device to perform the target detection method provided in the various optional implementations described above.

[0010] The technical solution provided by this invention may include the following beneficial effects: The computer equipment uses camera extrinsic data and geometric priors to correct the detection bounding box data of the detected target. This effectively eliminates unreasonable bounding boxes caused by perspective distortion or positioning deviation, improving positioning accuracy. For the corrected local image, brightness entropy screening and background suppression enhancement are introduced. On the one hand, this increases image information entropy by adjusting uneven illumination; on the other hand, it suppresses interference from complex backgrounds, highlighting target details. Finally, combined with a depth anomaly mask compensation mechanism, contour refinement and occlusion compensation are performed on the enhanced local image to generate the final detection bounding box data. This process achieves closed-loop optimization from coarse detection to fine-grained correction. Through multi-stage collaborative optimization, it significantly reduces the false detection rate and false negative rate, significantly improving the accuracy of target detection.

[0011] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and are not intended to limit the invention. Attached Figure Description

[0012] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with the invention and, together with the description, serve to explain the principles of the invention.

[0013] Figure 1 This is a system configuration diagram of a target detection method according to an embodiment of the present invention; Figure 2 This is a flowchart of a target detection method provided in one embodiment of the present invention; Figure 3 This is a flowchart of a target detection method provided in one embodiment of the present invention; Figure 4 This is a flowchart of a target detection method provided in one embodiment of the present invention; Figure 5 This is a flowchart of a target detection method provided in one embodiment of the present invention; Figure 6 This is a flowchart of a target detection method provided in one embodiment of the present invention; Figure 7 This is a block diagram of a target detection device provided in an exemplary embodiment of the present invention. Detailed Implementation

[0014] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numerals in different drawings denote the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present invention. Rather, they are merely examples of apparatuses and methods consistent with some aspects of the invention as detailed in the appended claims.

[0015] Figure 1This is a system configuration diagram of a target detection method according to an embodiment of the present invention. Figure 1 As shown, the system includes a requester device 120 and a servicer device 130 for target detection.

[0016] The requesting device 120 can be a personal computer device such as a desktop computer, a laptop computer, or a personal workstation.

[0017] The service provider equipment 130 can be a single server, or a combination of several servers, or a virtualization platform, or a cloud computing service center.

[0018] The requesting device 120 and the serving device 130 can be connected via a communication network. Optionally, this communication network can be a wired network or a wireless network.

[0019] Optionally, the aforementioned wireless or wired networks use standard communication technologies and / or protocols. The network is typically the Internet, but can also be any network, including but not limited to Local Area Networks (LANs), Metropolitan Area Networks (MANs), Wide Area Networks (WANs), mobile, wired or wireless networks, private networks, or any combination of virtual private networks. In some embodiments, technologies and / or formats including Hyper Text Markup Language (HTML), Extensible Markup Language (XML), etc., are used to represent data exchanged over the network. Furthermore, conventional encryption technologies such as Secure Socket Layer (SSL), Transport Layer Security (TLS), Virtual Private Networks (VPNs), and Internet Protocol Security (IPsec) can be used to encrypt all or some links. In other embodiments, custom and / or dedicated data communication technologies can be used to replace or supplement the aforementioned data communication technologies.

[0020] Please refer to Figure 2 , Figure 2 This is a flowchart of a target detection method provided in one embodiment of the present invention. The target detection method can be executed by a computer device, for example, the computer device described above. Figure 1 The service provider equipment 130 shown is shown.

[0021] The above target detection method may include the following steps: Step 210: Obtain the detection bounding box data of the detected target, the camera image, and the camera extrinsic data. The detection bounding box data includes the target confidence score and the detection bounding box position coordinates. The camera image is an image containing the detected target.

[0022] The aforementioned camera image refers to a two-dimensional grating pattern containing the target to be detected, captured by an imaging sensor.

[0023] In some embodiments, the computer device may acquire camera images via a corresponding external camera.

[0024] The aforementioned detection target refers to the target object detected by the computer device from the camera image. In the autonomous driving scenario, the detection target can be a static obstacle (e.g., roadblock, wall), a dynamic traffic participant (e.g., pedestrian, vehicle, cyclist), or traffic infrastructure (e.g., traffic lights, speed limit signs, lane lines). In the intelligent security scenario, the target object can be a specific person or a leftover package.

[0025] In some embodiments, a computer device may detect a target from a camera image based on a target detection algorithm, such as YOLO (You Only LookOnce, a target detection network).

[0026] The aforementioned detection box data consists of two-dimensional rectangular bounding boxes obtained by a computer device performing a target detection algorithm on a camera image.

[0027] The target confidence score mentioned above is the probability value that the computer device determines that the detection box contains the target. The target confidence score usually ranges from 0 to 1. The higher the target confidence score, the greater the computer device's confidence that the target object exists in the prediction box and the category prediction is correct, that is, it believes that the detection result is more likely to be real and effective. The lower the target confidence score, the less confident the computer device is that the target object exists in the prediction box, that is, it believes that the area is more likely to be background, noise or false detection.

[0028] The coordinates of the detection box are usually defined as the pixel coordinates of the top left and bottom right corners in the image coordinate system.

[0029] In some embodiments, the computer device may first perform preprocessing on the camera image (e.g., normalization, scaling to 640 x 640 pixels), and then input the preprocessed camera image into a pre-trained deep learning model for forward inference. The deep learning model extracts features from the camera image through the backbone network and performs multi-scale feature fusion on the camera image using the neck network. Finally, the detection head outputs the detection result, which is detection box data containing the target confidence and the location coordinates of the detection box.

[0030] Step 220: Based on the detection box data, camera image, and camera extrinsic data, obtain the corrected detection box data after geometric prior correction.

[0031] The aforementioned corrected detection box data refers to the precise bounding box data adjusted by computer equipment based on the detection box data and combined with the principles of perspective projection geometry.

[0032] Step 230: Perform enhancement operations on the local image corresponding to the corrected detection box data in the camera image. The enhancement operations include brightness entropy screening and background suppression enhancement.

[0033] The aforementioned partial image specifically refers to a sub-image cropped from a camera image by a computer device based on the modified detection box data.

[0034] The aforementioned brightness entropy screening refers to the computer device calculating the grayscale entropy value of a local image to assess the information richness of the local image. In some embodiments, if the entropy value of the image is lower than a threshold, the image is determined to be overexposed or underexposed, triggering brightness adjustment.

[0035] The aforementioned background suppression enhancement aims to highlight the edges and details of the detected target by computer devices suppressing background texture interference unrelated to the detected target in local images through frequency domain filtering (e.g., high-pass filtering) or spatial domain algorithms (e.g., guided filtering).

[0036] Optionally, the computer device can first perform brightness entropy screening on the local image corresponding to the corrected detection box data in the camera image, and then perform background suppression enhancement on the local image after brightness entropy screening.

[0037] Optionally, the computer device can first perform background suppression enhancement on the local image corresponding to the corrected detection box data in the camera image, and then perform brightness entropy screening on the local image after background suppression enhancement.

[0038] Optionally, the computer device can simultaneously perform brightness entropy screening and background suppression enhancement on the local images corresponding to the corrected detection box data in the camera image, and then fuse the images that have undergone brightness entropy screening and background suppression enhancement to obtain a single image.

[0039] For example, background suppression enhancement is performed on a local image to obtain an enhanced local image. Then, background suppression enhancement is performed on a copy of the local image (a copy image that is exactly the same as the local image) to obtain another enhanced local image. These two enhanced local images are then fused to obtain a single image, which is the image after the enhancement operation. In this example, the fusion operation can be implemented through an image fusion model, which is a machine learning model that has the ability to fuse two images into one image.

[0040] Step 240: Based on the corrected detection box data, perform a depth anomaly mask compensation operation on the local image after the enhancement operation to generate target detection box data.

[0041] The aforementioned depth anomaly mask compensation operation refers to an anomaly mask generated by a computer device using a deep learning segmentation network (e.g., U-Net or Mask R-CNN). This anomaly mask is used to identify the precise pixel-level contours of the detected target in the local image after the enhancement operation and to compensate for and fill in any occluded areas or missing parts that may exist in the previous steps.

[0042] In this embodiment, the computer device uses camera extrinsic data and geometric priors to correct the detection bounding box data of the detected target. This effectively eliminates unreasonable bounding boxes caused by perspective distortion or positioning deviation, improving positioning accuracy. For the corrected local image, brightness entropy screening and background suppression enhancement are introduced. On the one hand, this increases image information entropy by adjusting uneven illumination; on the other hand, it suppresses interference from complex backgrounds, highlighting target details. Finally, combined with a depth anomaly mask compensation mechanism, contour refinement and occlusion compensation are performed on the enhanced local image to generate the final detection bounding box data. This process achieves closed-loop optimization from coarse detection to fine-grained correction. Through multi-stage collaborative optimization, it significantly reduces the false detection rate and false negative rate, significantly improving the accuracy of target detection.

[0043] based on Figure 2 For any one or more embodiments, please refer to Figure 3 , Figure 3 This is a flowchart of a target detection method provided in one embodiment of the present invention.

[0044] Please refer to Figure 3 The above step 220 can be implemented as steps 220a, 220b, 220c, and 220d, as detailed below: Step 220a: Calculate the projection area of ​​the ground plane in the camera image based on the camera extrinsic data to obtain the ground plane mask data corresponding to the projection area.

[0045] The aforementioned projection area of ​​the ground plane in the camera image refers to the area projected onto the physical ground plane (usually set as a plane) in the three-dimensional world coordinate system by the computer device based on the camera's extrinsic parameters (e.g., rotation matrix and translation vector) and intrinsic parameters, using the principle of perspective projection. The specific pixel region formed after mapping to the two-dimensional image coordinate system, the projection area of ​​the ground plane in the camera image is usually presented in geometric shape as a trapezoid with the vanishing point as the vertex, representing the observable ground range within the camera's field of view.

[0046] The aforementioned ground plane mask data is a binary or categorized image data structure (usually represented as a single-channel grayscale image or a Boolean matrix), and the size of the ground plane mask data is consistent with that of the camera image.

[0047] In the ground plane masked data, pixels belonging to the projection area of ​​the ground plane in the camera image are assigned specific values ​​(e.g., 255 or 1), while pixels belonging to non-ground planes (e.g., sky, buildings, vehicles themselves, distant obstacles) are assigned background values ​​(e.g., ...). Ground plane mask data is essentially a pixel-level spatial filter used to precisely separate ground information from camera images.

[0048] In some embodiments, the computer device constructs a projection matrix using camera intrinsic and extrinsic parameters, maps the corner points of a preset ground plane region in the three-dimensional world coordinate system to a two-dimensional image plane, thereby determining the polygonal projection contour of the ground plane in the image, and then uses a graphic filling algorithm to mark the region inside the contour as the foreground and the remaining region as the background, generating binary mask data to distinguish between ground and non-ground pixels, which serves as the ground plane band mask data corresponding to the projection region.

[0049] Step 220b: Perform spatial overlap calculation on the detection box data and the ground plane mask data to obtain the positional relationship data of each detection box relative to the ground plane.

[0050] In this embodiment, the computer device can construct a corresponding binarized region mask in a blank matrix based on the pixel coordinates of the detection box. Then, a bitwise AND operation is performed between the binarized region mask and the ground plane mask to extract the overlapping pixel set of the two in the image space. The overlapping area and crossover-union ratio are accurately calculated by statistically analyzing the number of non-zero pixels in the intersection and union regions. At the same time, the vertical pixel distance between the bottom edge of the detection box and the edge of the ground plane mask is analyzed to finally quantify and generate positional relationship data including overlap rate, contact state and spatial deviation dimensions.

[0051] The positional relationship data of each detection box relative to the ground plane refers to a set of quantitative indicators describing the geometric topological relationship between the target detection box and the ground plane projection area (i.e., the ground plane mask) in the two-dimensional image space. The positional relationship data includes, but is not limited to, overlap area, overlap rate, containment degree, and vertical distance. The positional relationship data is used to characterize whether the detected target is on the ground, whether it is suspended in the air, or whether it is located in a non-driving area (e.g., building walls or the sky).

[0052] Step 220c: Perform an adaptive confidence weighting operation on the detection box data based on the positional relationship data.

[0053] The aforementioned confidence adaptive weighting operation refers to multiplicatively or additively correcting the target confidence in the detection box data based on the geometric reasonableness score in the positional relationship data.

[0054] In some embodiments, the computer device can obtain weighting coefficients by looking up a table or generating a model, and then perform an adaptive confidence weighting operation on the detection box data based on the obtained weighting coefficients.

[0055] Optionally, the computer device obtains weighting coefficients by querying a confidence weight mapping table based on the location relationship data, and performs an adaptive confidence weighting operation on the detection box data based on the obtained weighting coefficients.

[0056] Specifically, the computer device is equipped with a confidence weight mapping table, which stores the mapping relationship between positional relationship data and weighting coefficients. A set of positional relationship data corresponds to a weighting coefficient. After obtaining the positional relationship data, the computer device traverses the confidence weight mapping table to obtain the weighting coefficient corresponding to the positional relationship data in the confidence weight mapping table, and performs a confidence adaptive weighting operation on the detection box data based on the obtained weighting coefficient.

[0057] Optionally, the computer device inputs the location relationship data and the detection box data into the weighted coefficient generation model, obtains the weighted coefficients output by the weighted coefficient generation model, and then performs an adaptive confidence weighting operation on the detection box data based on the obtained weighted coefficients. The weighted coefficient generation model is a machine learning model with the ability to generate weighted coefficients based on the location relationship data and the detection box data.

[0058] Step 220d: Obtain the detection box data after performing the confidence adaptive weighting operation as the corrected detection box data.

[0059] In this embodiment, the computer device accurately calculates the projection area of ​​the ground plane in the image through camera extrinsic parameters, generates corresponding ground plane mask data, and accurately determines the position of each detection box relative to the ground plane by calculating the spatial overlap relationship between the detection box and the mask. Based on this positional relationship, the confidence of the detection box is adaptively weighted. The corrected detection box data after weighting can effectively filter out false detections and false alarms above or below the ground, significantly improving the accuracy and reliability of target detection, and is especially suitable for visual perception scenarios based on ground movement.

[0060] Based on any one or more of the above embodiments, please refer to Figure 4 , Figure 4 This is a flowchart of a target detection method provided in an embodiment of the present invention. Step 220c above can be implemented as steps 220c1 and 220c2, as detailed below: Step 220c1: Generate geometric weight coefficients based on location relationship data.

[0061] The aforementioned geometric weighting coefficient is a scalar value between 0 and 1, used to characterize the degree to which the detected target conforms to the physical prior of being grounded.

[0062] In some embodiments, a computer device may generate geometric weighting coefficients based on location relationship data using a function.

[0063] Specifically, the computer device contains a piecewise function or a sigmoid function, which processes positional relationship data (e.g., the vertical deviation of the bottom point). Mapped to weights :when When it approaches 0 (i.e., the bottom of the target perfectly fits the ground plane), Approaching 1; when When the target is too large (suspended in the air or sunk underground), It decays rapidly, approaching 0.

[0064] For example, if the target to be detected is a pedestrian, and the bottom center point of the detection box containing the pedestrian is exactly inside the ground plane mask, the resulting geometric weight coefficient is 0.95.

[0065] Step 220c2: Perform a correction operation on the detection box data according to the geometric weight coefficients. The correction operation is one of weighted amplification and suppression reduction.

[0066] The aforementioned weighted amplification refers to multiplying the confidence level of the detection box data by the geometric weight coefficient, which amplifies the confidence level of the detection box data.

[0067] The aforementioned suppression reduction refers to significantly reducing the confidence level of the detection box data or even setting it to zero.

[0068] In this embodiment, the computer device acquires the geometric weight coefficients corresponding to the detection box data. And perform confidence update operation. This means performing a correction operation on the detection frame data.

[0069] For example, a real car driving on the road (i.e., the detection target) has an original detection confidence of 0.7, a geometric weight coefficient of 0.95, and a corrected confidence of 0.665. However, a false detection box floating in the air (e.g., a car on a billboard) has a geometric weight coefficient of 0.1, and its corrected confidence drops sharply from 0.7 to 0.07. Thus, it is effectively filtered out in the non-maximum suppression stage, achieving accurate purification of the detection results.

[0070] In this embodiment, the adaptive weighting mechanism based on spatial geometric relationships effectively suppresses background interference and model misjudgment, making the final output corrected detection box data more consistent with the actual layout of the physical world, thereby improving the accuracy and robustness of the detection results without increasing hardware costs.

[0071] Based on any one or more of the above embodiments, in some embodiments, the positional relationship data includes the overlap ratio, the position of the bottom edge of the frame, the height of the center point of the frame, and the coverage of the frame in the ground zone.

[0072] The aforementioned overlap ratio refers to the cross-union ratio of the detection box and the ground plane mask at the pixel level.

[0073] The bottom edge position of the above-mentioned box refers to the Y-axis coordinate of the midpoint of the bottom edge of the detection box in the image coordinate system.

[0074] The height of the center point of the detection frame mentioned above refers to the Y-axis coordinate value of the center point of the detection frame.

[0075] The coverage of the aforementioned frame in the ground zone refers to the percentage of the length of the continuous area at the bottom of the detection frame that is covered by the ground plane mask.

[0076] Step 220c1 above can be implemented as follows: calculate the correlation between the detection box corresponding to the detection target and the ground area based on the overlap ratio, the bottom edge position of the box, the height of the center point of the box, and the coverage of the box in the ground area; when the correlation is higher than the first preset threshold, generate the first geometric weight coefficient, or when the correlation is lower than the second preset threshold, generate the second geometric weight coefficient, wherein the first preset threshold is higher than the second preset threshold and the first geometric weight coefficient is greater than the second geometric weight coefficient.

[0077] In some embodiments, the computer device can construct a multidimensional feature vector, as follows: ,in, It's the overlap ratio. It refers to the coverage of the frame within the ground zone. Yes, the bottom edge of the frame. It is the center height of the frame, and then, through a preset scoring function... It can be a weighted linear combination or a simple rule engine to calculate a comprehensive relevance score, which is used to measure the authenticity of the detected target's landing.

[0078] The aforementioned first preset threshold is a high threshold value close to 1.0 (e.g., 0.95) used to filter out detection results that are geometrically near perfect.

[0079] In some embodiments, when the correlation score exceeds a first preset threshold, it is determined that the detected target conforms to physical laws, not only landing but also having a standard posture. At this time, the generated first geometric weight coefficient is set to a value greater than 1 (e.g., 1.1 or 1.2).

[0080] The aforementioned second preset threshold is a relatively low threshold value (e.g., 0.3), and the second preset threshold is lower than the first preset threshold (i.e., ...). ).

[0081] In some embodiments, when the correlation score is lower than a second preset threshold, it indicates that the detected target seriously violates physical laws (e.g., the target is suspended in the air, its bottom is seriously deviated from the ground, or it is occluded, resulting in a misaligned position). In this case, the second geometric weight coefficient generated by the computer device is set to a value less than 1 (e.g., 0.5 or 0.1).

[0082] In this embodiment, the computer device performs comprehensive correlation analysis by constructing a multi-dimensional geometric feature vector (overlap ratio, bottom edge position, center point height, bottom coverage) to generate geometric weight coefficients. This makes the generated geometric weight coefficients more closely resemble the actual situation of the detection target and the on-site environment, further improving the accuracy of target detection.

[0083] Based on any one or more of the above embodiments, please refer to Figure 5 , Figure 5 This is a flowchart of a target detection method provided in an embodiment of the present invention. Step 230 above can be implemented as steps 230a and 230b, as detailed below: Step 230a: Perform brightness entropy screening on the corresponding local image in the camera image based on the corrected detection box data. In some embodiments, the computer device can crop out the corresponding local region from the original camera image based on the coordinate information of the modified detection box, randomly perform grayscale processing on the local region and count the pixel grayscale histogram, and then calculate the brightness entropy value of the region based on the normalized grayscale probability distribution using the Shannon entropy formula to quantify its texture complexity and information richness. Finally, the calculated brightness entropy value is compared with a preset threshold. If the entropy value is lower than the threshold, the region is determined to lack effective features and is removed. If the entropy value is higher than the threshold, it is determined to be a valid target and is retained, thereby completing the brightness entropy screening of the detection results. Step 230b: Perform background suppression enhancement on the local image after performing brightness entropy screening.

[0084] In this embodiment, luminance entropy effectively reflects the complexity and sharpness of an image's brightness distribution. Luminance entropy screening can assess the imaging quality of local images, identifying areas that are too dark, overexposed, or blurry. Subsequently, a background suppression and enhancement operation is adaptively performed on the image after luminance entropy screening. This significantly enhances the texture and edge details of the target area while suppressing complex background interference. This process not only improves the visibility and sharpness of the target but also reduces the interference of background noise on subsequent recognition. Ultimately, this enhances the accuracy of target recognition, maintaining excellent visual perception performance, especially in complex environments such as low light and high noise.

[0085] Based on any one or more of the above embodiments, in some embodiments, step 230a can be implemented as follows: performing grayscale processing on a local image to obtain a grayscale image; calculating the mean brightness, variance brightness, and entropy value of the local image based on the grayscale image; obtaining an image region in the local image whose mean brightness is lower than a brightness threshold, whose variance brightness is lower than a variance threshold, and whose entropy value is lower than a grayscale entropy value threshold, as a first image region, wherein the first image region is an image region that simultaneously has low brightness and low texture features.

[0086] The grayscale processing mentioned above refers to converting the color value of each pixel, which was originally represented by three components (red (R), green (G), and blue (B), into a grayscale value represented by only one component. This value is usually between 0 (black) and 255 (white) to characterize the brightness of the pixel, thereby generating a grayscale image.

[0087] In some embodiments, the computer device may use a weighted average method to calculate the gray value corresponding to a local image based on the differences in human eye sensitivity to different colors, according to the formula $Gray = 0.299R + 0.587G + 0.114B$.

[0088] The aforementioned average brightness value refers to the arithmetic mean of the grayscale values ​​of all pixels in a grayscale image, used to measure the overall brightness level of the image.

[0089] The aforementioned brightness variance is the average of the squares of the differences between the gray values ​​of each pixel and the mean, used to measure the dispersion of the image's gray-level distribution, i.e., the strength of the contrast.

[0090] The grayscale entropy value mentioned above is used to measure the complexity or randomness of image texture.

[0091] In some embodiments, a computer device can generate a grayscale histogram and then use a numerical computing library (e.g., NumPy) to directly call the `mean()`, `var()` functions and a custom entropy calculation function to complete the index extraction.

[0092] In this embodiment, when the average brightness of a local image is lower than a brightness threshold (dark), the brightness variance is lower than a variance threshold (low contrast), and the grayscale entropy value is lower than a grayscale entropy value threshold (simple texture), the computer device determines that the region is a first image region. The first image region typically corresponds to a deep shadow, a severely occluded target, or blurred background clutter, and is characterized by having both low brightness and low texture.

[0093] In this embodiment, the solution simplifies the data dimensions and retains brightness information by performing grayscale processing on local images, thereby accurately calculating the mean, variance, and grayscale entropy value of brightness, comprehensively evaluating the light intensity, contrast, and texture complexity of the image. By setting multiple thresholds for joint judgment, the solution effectively identifies the first image region that simultaneously has low brightness and low texture characteristics, i.e., the potential target region with poor imaging quality and missing details. The screening process has strong discriminative power, avoids misjudgment based on a single indicator, and improves the accuracy of region positioning.

[0094] Based on any one or more of the above embodiments, in some embodiments, step 230b can be implemented as follows: analyzing the brightness difference, texture difference and gradient difference between the first image region and the background region, wherein the background region is the surrounding background region of the first image region in the local image; based on the analysis results, removing the image region in the first image region that is consistent with the background region, and enhancing the foreground region in the first image region that is inconsistent with the background region.

[0095] The aforementioned first image region refers to the low-brightness, low-texture region determined after brightness entropy screening, which usually corresponds to the image region under shadow or occlusion.

[0096] The aforementioned background region is the surrounding ring region adjacent to the first image region in the local image, used to represent the environmental context in which the target is located.

[0097] The brightness difference mentioned above is the absolute difference between the average gray values ​​of the two regions, representing the contrast between light and dark.

[0098] The aforementioned texture difference can be calculated by the difference in contrast or energy value between the gray-level co-occurrence matrices of the two regions, representing the difference in structural complexity.

[0099] The gradient difference mentioned above is the Euclidean or chiral distance between the histograms of pixel gradient magnitudes in two regions, representing the difference in edge sharpness.

[0100] In this embodiment, the above-mentioned elimination refers to setting pixels with similar features to the background area to black or transparent in order to eliminate interference, and the above-mentioned enhancement refers to increasing the contrast or brightness of pixels with significantly different features from the background area in order to highlight the target.

[0101] In some embodiments, the computer device sets difference thresholds (including brightness difference threshold, texture difference threshold, and gradient difference threshold). If the brightness difference, texture difference, and gradient difference of a pixel are all less than the threshold, the pixel is determined to be consistent with the background area, and a culling operation is performed (e.g., the pixel value is set to 0). Conversely, if the difference is significant (e.g., exceeding the threshold), the pixel is determined to be a foreground area inconsistent with the background area, and an enhancement operation is performed, such as improving the brightness and contrast of the pixel through histogram stretching or non-linear transformation.

[0102] In this embodiment, the above-described scheme achieves effective separation of the foreground and background by meticulously analyzing the differences in brightness, texture, and gradient between the first image region and its surrounding background region. Based on the analysis results, regions consistent with background features are removed, which can significantly suppress background noise and interference, reducing the false detection rate. Simultaneously, targeted enhancement of foreground regions inconsistent with the background can significantly highlight the key details and features of the target, not only optimizing the signal-to-noise ratio of the image but also making the target features more distinct, thereby greatly improving the accuracy of subsequent target recognition and analysis, especially suitable for detecting small targets in complex backgrounds.

[0103] Based on any one or more of the above embodiments, please refer to Figure 6 , Figure 6 This is a flowchart of a target detection method provided in an embodiment of the present invention. Step 240 above can be implemented as steps 240a to 240e, as detailed below: Step 240a: Obtain the depth map and appearance anomaly features corresponding to the local image after the enhancement operation.

[0104] The enhanced local image refers to the target region image block that has undergone background suppression and brightness adjustment processing.

[0105] The depth map corresponding to the local image after the above enhancement operation is a single-channel image. The gray value or numerical value of each pixel in the local image after the enhancement operation directly encodes the physical distance (usually in meters) of that pixel from the camera in the three-dimensional scene. The larger the value, the farther the distance. This is the key data for recovering the three-dimensional structure from the two-dimensional image through stereo matching algorithm or deep neural network.

[0106] The aforementioned abnormal appearance features refer to feature vectors extracted through traditional image processing or shallow neural networks that characterize the surface texture of the target surface in a way that does not conform to expectations. Examples include unnatural noise, color banding, or patches that do not conform to the material's reflection rules in local areas.

[0107] In some embodiments, a computer device may invoke a pre-trained monocular depth estimation model (e.g., MiDaS or LeRes) to infer local images and generate depth maps corresponding to the augmented local images.

[0108] In some embodiments, a computer device may use a lightweight anomaly detection network to extract appearance features corresponding to a local image after enhancement.

[0109] Step 240b: Generate a depth anomaly mask based on the depth map.

[0110] The aforementioned depth anomaly mask is a binary image of the same size as the local image, used to mark the locations of pixels with unreasonable depth distribution within the target region.

[0111] In some embodiments, the computer device may first perform preprocessing on the depth map, namely, calculating the average depth within a local image. and standard deviation Then, a decision threshold based on statistical distribution is set (e.g., Alternatively, morphological operations (e.g., top-hat transformation) can be used to extract local extrema in the depth map. For pixels that deviate from the normal depth range (too close or too far), the pixel is marked as 1 (abnormal) in the mask, and the remaining pixels are marked as 0 (normal).

[0112] For example, if the target is a standing pedestrian, but its depth map shows that the pedestrian's "head" area is extremely far from the camera (close to the background distance), the computer device determines that there is a "hole" or "penetration" anomaly in the head area, and the generated depth anomaly mask will appear white in the chest area (1), thereby accurately delineating the suspicious parts in the target's three-dimensional structure.

[0113] Step 240c: Calculate the percentage of abnormal pixels and the distribution of abnormal locations based on the depth anomaly mask.

[0114] The aforementioned percentage of anomalous pixels refers to the percentage of pixels with a value of 1 in the depth anomaly mask out of the total number of pixels in the local image, used to measure the overall degree of anomalousness of the target.

[0115] The above-mentioned abnormal location distribution refers to the spatial distribution characteristics of abnormal pixels in the mask. It is usually described by calculating the centroid coordinates and distribution moments of the abnormal region or by dividing it into grids and then counting the abnormal density of each grid. It is used to determine whether the abnormality is concentrated in the key parts of the target (e.g., the central region or edge contour).

[0116] In some embodiments, the computer device first performs connected component analysis on the depth anomaly mask, counts the total number of anomaly pixels and divides it by the total pixel area to obtain the percentage of anomaly pixels; then it uses an image moment function (e.g., cv2.moments) to calculate the geometric center of the anomaly region, or uses histogram projection analysis to analyze its horizontal and vertical distribution to obtain the distribution of anomaly locations.

[0117] Step 240d: Obtain the compensation signal based on the percentage of abnormal pixels, the distribution of abnormal locations, and the abnormal appearance features.

[0118] The aforementioned compensation signal is a control parameter, typically represented as a floating-point number within the range [0,1] or a classification instruction, used to indicate the direction and magnitude of the subsequent correction module's adjustment of the detection box confidence.

[0119] In some embodiments, the computer device may employ a rule engine or a lightweight classifier to obtain a compensation signal based on the proportion of abnormal pixels, the distribution of abnormal locations, and appearance abnormalities.

[0120] Optionally, the computer device is equipped with a rule engine that obtains a compensation signal based on the percentage of abnormal pixels, the distribution of abnormal locations, and the appearance abnormalities according to set rules.

[0121] Specifically, the rules are as follows: if the proportion of abnormal pixels is less than 5% and the appearance features are normal, a positive compensation signal is generated (e.g., 1.0, indicating maintenance or slight increase); if the proportion of abnormal pixels is greater than 30% or the abnormal position is distributed in the core area of ​​the target (e.g., the central 1 / 3 area) and the appearance is significantly abnormal, a negative compensation signal is generated (e.g., 0.5, indicating significant suppression).

[0122] Step 240e: Perform a confidence-based secondary correction operation on the corrected detection box data based on the compensation signal to generate target detection box data.

[0123] The aforementioned corrected detection box data refers to the detection boxes after geometric prior correction, which include location coordinates and initial confidence scores. .

[0124] The aforementioned confidence level secondary correction operation refers to using a compensation signal to nonlinearly adjust the initial confidence level.

[0125] In some embodiments, the computer device can perform a second confidence correction operation on the corrected detection box data based on the compensation signal using a formula, as follows: or ,in It is determined by the compensation signal.

[0126] Among them, the above C initial It is the initial confidence level, W comp It is the compensation weight, Δ is the compensation increment, and C final It is the final confidence level (i.e., the confidence level obtained after performing a second confidence level correction operation).

[0127] In this embodiment, the scheme introduces a joint analysis mechanism of depth map and appearance anomaly features. First, a depth anomaly mask is generated, and then the proportion and location distribution of anomalous pixels are quantified to accurately characterize the depth reliability of the target. By fusing appearance anomaly features, a comprehensive compensation signal is generated, effectively compensating for the limitations of single-modality judgment. The compensation signal is used to perform secondary confidence correction on the corrected detection box, realizing dual verification from image enhancement to geometric consistency. This not only suppresses false detections caused by depth loss or anomalies but also improves the accuracy of target localization and confidence scoring, ultimately outputting more reliable target detection box data and significantly enhancing detection stability in complex scenes.

[0128] Based on any one or more of the above embodiments, in some embodiments, step 240d can be implemented as follows: generating a compensation score based on the proportion of abnormal pixels, the distribution of abnormal positions, and the abnormal appearance features; and converting the compensation score into a compensation signal.

[0129] The aforementioned compensation score is a numerical index calculated through weighted fusion, used to comprehensively characterize the degree of confidence decay of the current detection box due to geometric and appearance anomalies.

[0130] In some embodiments, the computer device may perform normalization processing on three parameters: abnormal pixel ratio, abnormal location distribution, and abnormal appearance features, and calculate the compensation score using a weighted summation strategy.

[0131] The specific formula can be as follows: S_comp=w_1⋅r_ap+w_2⋅D_loc+w_3⋅F_app; Where r_ap represents the proportion of abnormal pixels, D_loc represents the normalized score of the distribution of abnormal locations (e.g., the score is higher for abnormal locations in the core region), F_app represents the intensity value of the appearance abnormal features, w_1, w_2, w_3 are trainable weight coefficients, and S_comp is the compensation score.

[0132] In some embodiments, the computer device may input the calculated compensation score S_comp into a preset mapping function (e.g., a piecewise linear function or a sigmoid function in the lower-level description) to convert the compensation score into a compensation signal Δc.

[0133] The compensation signal is an incremental value or scaling factor used to correct the original confidence level. The numerical range of the compensation signal is usually limited to the interval [-1, 0], representing the amount of deduction from the original confidence level.

[0134] In this embodiment, the compensation score comprehensively reflects both depth and appearance anomaly information in the image, offering greater accuracy and robustness than a single threshold judgment. Converting the compensation score into a compensation signal provides a fine and continuous adjustment basis for subsequent confidence level correction, effectively avoiding information loss and misjudgment caused by simple threshold truncation. This score-based compensation mechanism significantly improves the detection system's adaptability and discrimination accuracy in complex anomaly scenes, enhancing the reliability of target detection.

[0135] For example, based on Figures 2 to 6 For any one or more embodiments, the present invention proposes a small-body obstacle recognition enhancement mechanism based on spatial layering and geometric constraints.

[0136] In robot vision navigation and obstacle avoidance systems, traditional target detection models (such as the YOLO series) struggle to reliably detect small, near-ground obstacles, targets with weak features, and transparent / reflective objects. The core of this application example is to introduce a series of multi-stage post-processing enhancement pipelines after the YOLO model outputs the original detection boxes and before performing non-maximum suppression and final decision-making. This pipeline sequentially performs geometric prior correction, appearance weak feature enhancement, and depth anomaly compensation on the original detection results, thereby systematically improving the detection robustness and recall rate in complex scenarios.

[0137] Specifically, the pipeline first utilizes the ground geometry prior module to analyze the original detection boxes (corresponding to the above). Figures 2 to 6 The module performs spatial weighted correction on the detection bounding box data of the detected targets. Based on camera extrinsic parameters, such as installation height and pitch angle, it calculates the projection area of ​​the ground plane onto the image in the current field of view (i.e., the ground plane zone). Then, it calculates the relative position of each candidate detection box to this zone (e.g., overlap ratio, bottom edge position of the box). Based on this, detection boxes that fall into or are close to the ground zone (where near-ground obstacles such as cables and low steps are more likely to appear) are given higher geometric weights to amplify their confidence. Conversely, detection boxes far from the ground maintain or reduce their weights. This transforms the 3D spatial prior knowledge that small obstacles are more likely to appear near the ground into confidence correction for 2D detection results, effectively improving the recall rate of near-ground targets and suppressing aerial artifacts.

[0138] Next, the system further performs appearance weak feature compensation based on brightness entropy screening and background suppression enhancement on the geometrically corrected detection box data. This step starts from the original image (corresponding to the above). Figures 2 to 6The algorithm extracts local image patches corresponding to each candidate bounding box from the camera image and calculates features such as the mean brightness, variance, and gray-level entropy of these patches. If a region simultaneously meets the conditions of low brightness, low variance, and low entropy, it is identified as a weak feature region with low brightness and low texture. This is typical of many small, dark-textured near-ground obstacles (e.g., dark pet toys, objects in shadows). Subsequently, by analyzing the differences in brightness, texture, and gradient between this weak feature region and its surrounding background, the algorithm actively suppresses image portions consistent with the background while enhancing foreground regions that differ from the background. This visually "highlights" the weak feature obstacles, reducing missed detections caused by background interference.

[0139] Finally, for special material targets such as transparent glass and reflective objects, which have weak RGB appearance features and large areas of voids or abrupt changes in depth information (e.g., from an RGB-D camera), the solution introduces a compensation mechanism based on depth anomaly masks. This mechanism comprehensively analyzes the appearance anomalies (e.g., highlights, missing internal textures) and depth anomalies (e.g., abnormal pixel percentage, distribution) of the target region to generate a compensation signal. The compensation signal is used to further correct the confidence of the corresponding detection box, thereby improving the system's probability of retaining objects that are easily missed due to sensor limitations. Simultaneously, this compensation signal can also be passed to subsequent box-depth fusion steps to trigger a more conservative distance estimation strategy, avoiding incorrect identification of unobstructed objects due to depth data distortion.

[0140] In summary, this solution utilizes a sequential post-processing pipeline of "geometric prior correction → brightness entropy screening and background suppression → depth anomaly mask compensation" to organically integrate multi-source cues such as spatial geometry, appearance texture, and depth information without altering the YOLO backbone network, thereby providing targeted enhancement and compensation for the original detection results. This design is particularly suitable for applications such as home service robots and autonomous driving platforms that require accurate perception of various obstacles in complex ground environments, significantly improving the overall detection rate and system robustness for small, ground-hugging, weak-feature, and special-material obstacles.

[0141] Please refer to Figure 7 The diagram illustrates a block diagram of a target detection device provided in an exemplary embodiment of the present invention. This target detection device can be implemented as all or part of a robot system through hardware or a combination of hardware and software to achieve the above-described... Figures 2 to 6 All or part of the steps in the illustrated embodiments. For example... Figure 7 As shown, the target detection device includes: The basic data acquisition module 701 is used to acquire the detection box data of the detected target, the camera image, and the camera extrinsic data. The detection box data includes the target confidence score and the detection box position coordinates, and the camera image is an image containing the detected target. The modified detection box data acquisition module 702 is used to acquire modified detection box data after geometric prior correction based on the detection box data, camera image, and camera extrinsic data. The modified detection box data is the detection box data after geometric prior correction. The enhancement operation execution module 703 is used to perform enhancement operations on the local image corresponding to the corrected detection box data in the camera image. The enhancement operations include brightness entropy screening and background suppression enhancement. The mask compensation execution module 704 is used to perform depth anomaly mask compensation operation on the local image after the enhancement operation based on the corrected detection box data, so as to generate target detection box data.

[0142] In some embodiments, the correction detection box data acquisition module 702 is used to calculate the projection area of ​​the ground plane in the camera image based on the camera extrinsic data, so as to obtain the ground plane mask data corresponding to the projection area. Spatial overlap calculation is performed on the detection box data and the ground plane mask data to obtain the positional relationship data of each detection box relative to the ground plane. Perform an adaptive confidence weighting operation on the detection box data based on the positional relationship data; The detection box data after performing the confidence-adaptive weighting operation is obtained as the corrected detection box data.

[0143] In some embodiments, the modified detection box data acquisition module 702 is used to generate geometric weight coefficients based on positional relationship data; The detection box data is corrected based on the geometric weight coefficients. The correction operation is one of weighted amplification and suppression reduction.

[0144] In some embodiments, the positional relationship data includes the overlap ratio, the position of the bottom edge of the frame, the height of the center point of the frame, and the degree of coverage of the frame in the ground zone; The detection box data acquisition module 702 is used to calculate the correlation between the detection box corresponding to the detection target and the ground zone area based on the overlap ratio, the bottom edge position of the box, the height of the center point of the box, and the coverage of the box in the ground zone. When the correlation is higher than the first preset threshold, a first geometric weight coefficient is generated; or when the correlation is lower than the second preset threshold, a second geometric weight coefficient is generated. The first preset threshold is higher than the second preset threshold, and the first geometric weight coefficient is greater than the second geometric weight coefficient.

[0145] In some embodiments, the enhanced operation execution module 703 is used to perform brightness entropy screening on the corresponding local image in the camera image based on the corrected detection box data; Background suppression enhancement is performed on the local image after brightness entropy screening.

[0146] In some embodiments, the enhanced operation execution module 703 is used to perform grayscale processing on a local image to obtain a grayscale image; Based on the grayscale image, calculate the mean brightness, variance brightness, and grayscale entropy value of the local image. Image regions in a local image whose mean brightness is lower than a brightness threshold, whose brightness variance is lower than a variance threshold, and whose grayscale entropy value is lower than a grayscale entropy value threshold are obtained as the first image region. The first image region is an image region that simultaneously has low brightness and low texture features.

[0147] In some embodiments, the enhanced operation execution module 703 is used to analyze the brightness difference, texture difference and gradient difference between the first image region and the background region, wherein the background region is the surrounding background region of the first image region in the local image. Based on the analysis results, image regions in the first image region that are consistent with the background region are removed, and foreground regions in the first image region that are inconsistent with the background region are enhanced.

[0148] In some embodiments, the mask compensation execution module 704 is used to obtain the depth map corresponding to the local image after the enhancement operation, as well as the appearance anomaly features; Generate a depth anomaly mask from the depth map; Calculate the percentage of abnormal pixels and the distribution of abnormal locations based on the depth anomaly mask; The compensation signal is obtained based on the proportion of abnormal pixels, the distribution of abnormal locations, and the abnormal appearance features; A second confidence correction operation is performed on the corrected detection box data based on the compensation signal to generate target detection box data.

[0149] In some embodiments, the mask compensation execution module 704 is used to generate a compensation score based on the proportion of abnormal pixels, the distribution of abnormal positions, and the appearance abnormal features. The compensation score is converted into a compensation signal.

[0150] In an exemplary embodiment, a computer device is also provided, the computer device including a processor and a memory, the memory storing instructions which are executed by the processor.

[0151] In an exemplary embodiment, a computer program product is also provided, comprising computer instructions stored in a computer-readable storage medium. A processor of a computer device reads the computer instructions from the computer-readable storage medium and executes the computer instructions to implement all or part of the steps of the methods shown in the above embodiments of the present invention.

[0152] In an exemplary embodiment, a computer-readable storage medium is also provided, which stores a computer program that is loaded and executed by a processor to implement all or part of the steps of the methods shown in the above embodiments of the present invention.

[0153] Those skilled in the art will understand that all or part of the steps of the above embodiments can be implemented by hardware, or by a program instructing related hardware. The program can be stored in a computer-readable storage medium, such as a read-only memory, a disk, or an optical disk.

[0154] Although the operations are described in a specific order in the accompanying drawings, this should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all of the operations shown to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0155] Any step, operation, or procedure described herein may be performed or implemented using one or more hardware or software modules, either alone or in combination with other devices. In one embodiment, the software module is implemented using a computer program product comprising a computer-readable medium containing computer program code, which is executable by a computer processor to perform any or all of the described steps, operations, or procedures.

[0156] The foregoing description of embodiments of the invention has been provided for purposes of illustration and description. The foregoing description is not exhaustive and is not intended to limit the invention to the exact forms disclosed; various modifications and variations may be made in accordance with the foregoing teachings, or may be derived from the practice of the invention. These embodiments were chosen and described to illustrate the principles of the invention and its practical application, enabling those skilled in the art to utilize the invention in various embodiments and with various modifications to suit the particular purpose of the concept.

[0157] It can be further understood that, unless otherwise specified, "connection" includes both direct connections where no other components exist between the two parties and indirect connections where other components exist between them.

[0158] It is further understood that although the operations are described in a specific order in the accompanying drawings in the embodiments of the present invention, this should not be construed as requiring these operations to be performed in the specific order or serial order shown, or requiring all the operations shown to obtain the desired result. In certain environments, multitasking and parallel processing may be advantageous.

[0159] Other embodiments of the invention will readily occur to those skilled in the art upon consideration of the specification and practice of the invention disclosed herein. This invention is intended to cover any variations, uses, or adaptations of the invention that follow the general principles of the invention and include common knowledge or customary techniques in the field of the invention not disclosed herein. The specification and examples are to be considered exemplary only, and the true scope and spirit of the invention are indicated by the following claims.

[0160] It should be understood that the present invention is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of the invention is limited only by the appended claims.

[0161] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A target detection method, characterized in that, The method includes: Acquire detection bounding box data of the target, camera image, and camera extrinsic data. The detection bounding box data includes the target confidence score and the detection bounding box position coordinates. The camera image is an image containing the target. Based on the detection box data, the camera image, and the camera extrinsic data, obtain the corrected detection box data after geometric prior correction; An enhancement operation is performed on the local image corresponding to the corrected detection box data in the camera image. The enhancement operation includes brightness entropy screening and background suppression enhancement. Based on the corrected detection box data, a depth anomaly mask compensation operation is performed on the local image after the enhancement operation to generate target detection box data.

2. The method according to claim 1, characterized in that, The step of obtaining geometrically prior-corrected modified detection box data based on the detection box data, the camera image, and the camera extrinsic data includes: Calculate the projection area of ​​the ground plane in the camera image based on the camera extrinsic data to obtain the ground plane mask data corresponding to the projection area; Spatial overlap calculation is performed on the detection frame data and the ground plane mask data to obtain the positional relationship data of each detection frame relative to the ground plane. Perform a confidence-adaptive weighted operation on the detection box data based on the positional relationship data; The detection box data after performing the confidence adaptive weighting operation is obtained as the corrected detection box data.

3. The method according to claim 2, characterized in that, The step of performing a confidence-adaptive weighting operation on the detection box data based on the positional relationship data includes: Geometric weight coefficients are generated based on the location relationship data; The detection box data is corrected according to the geometric weight coefficients, and the correction operation is one of weighted amplification and suppression reduction.

4. The method according to claim 3, characterized in that, The positional relationship data includes the overlap ratio, the position of the bottom edge of the frame, the height of the center point of the frame, and the degree of coverage of the frame in the ground zone. The generation of geometric weight coefficients based on the positional relationship data includes: Based on the overlap ratio, the position of the bottom edge of the frame, the height of the center point of the frame, and the coverage of the frame in the ground zone, the correlation between the detection frame corresponding to the detection target and the ground zone area is calculated. When the correlation is higher than a first preset threshold, a first geometric weight coefficient is generated; or when the correlation is lower than a second preset threshold, a second geometric weight coefficient is generated. The first preset threshold is higher than the second preset threshold, and the first geometric weight coefficient is greater than the second geometric weight coefficient.

5. The method according to claim 1, characterized in that, The enhancement operation performed on the local image corresponding to the corrected detection box data in the camera image includes: Based on the corrected detection box data, a brightness entropy screening is performed on the corresponding local image in the camera image; For the local image after performing the brightness entropy screening, background suppression enhancement is performed.

6. The method according to claim 5, characterized in that, The step of performing brightness entropy screening on the corresponding local image in the camera image based on the corrected detection box data includes: Perform grayscale conversion on a local image to obtain a grayscale image; Based on the grayscale image, calculate the mean brightness, variance brightness, and grayscale entropy value of the local image; Image regions in a local image where the mean brightness is lower than a brightness threshold, the brightness variance is lower than a variance threshold, and the grayscale entropy value is lower than a grayscale entropy value threshold are obtained as a first image region. The first image region is an image region that simultaneously has low brightness and low texture features.

7. The method according to claim 6, characterized in that, The background suppression enhancement is performed on the local image after the brightness entropy screening, including: Analyze the brightness difference, texture difference, and gradient difference between the first image region and the background region, wherein the background region is the surrounding background region of the first image region in the local image; Based on the analysis results, image regions in the first image region that are consistent with the background region are removed, and foreground regions in the first image region that are inconsistent with the background region are enhanced.

8. The method according to claim 1, characterized in that, The step of performing depth anomaly mask compensation operation on the local image after the enhancement operation based on the corrected detection box data includes: Obtain the depth map and appearance anomaly features of the local image after the enhancement operation; Generate a depth anomaly mask based on the depth map; Calculate the percentage of abnormal pixels and the distribution of abnormal locations based on the depth anomaly mask; A compensation signal is obtained based on the percentage of abnormal pixels, the distribution of abnormal locations, and the abnormal appearance features; Based on the compensation signal, a second confidence correction operation is performed on the corrected detection box data to generate target detection box data.

9. The method according to claim 8, characterized in that, The step of obtaining a compensation signal based on the percentage of abnormal pixels, the distribution of abnormal locations, and the abnormal appearance features includes: A compensation score is generated based on the percentage of abnormal pixels, the distribution of abnormal locations, and the abnormal appearance features. The compensation score is converted into a compensation signal.

10. A target detection device, characterized in that, The device includes: The basic data acquisition module is used to acquire the detection box data of the detected target, the camera image, and the camera extrinsic data. The detection box data includes the target confidence score and the detection box position coordinates, and the camera image is an image containing the detected target. The modified detection box data acquisition module is used to acquire modified detection box data after geometric prior correction based on the detection box data, the camera image, and the camera extrinsic data. An enhancement operation execution module is used to perform enhancement operations on the local image corresponding to the corrected detection box data in the camera image. The enhancement operations include brightness entropy screening and background suppression enhancement. The mask compensation execution module is used to perform a depth anomaly mask compensation operation on the local image after the enhancement operation is performed, based on the corrected detection box data, to generate target detection box data.

Images