Three-dimensional target detection system based on two-dimensional perception

Abstract

The invention discloses a three-dimensional target detection system based on two-dimensional perception. The three-dimensional target detection system comprises a two-dimensional perception module, a three-dimensional coordinate acquisition module and a post-processing module, the two-dimensional sensing module is used for performing feature extraction and region detection on an input image based on a deep learning target detection and key point detection algorithm, and outputting a bounding box of a to-be-detected target and two-dimensional coordinates of a plurality of key points; the three-dimensional coordinate acquisition module is used for performing spatial mapping on the two-dimensional coordinates of the plurality of key points and depth data acquired by the depth sensor, and calculating the three-dimensional coordinates of the key points; and the post-processing module is used for performing space consistency optimization on the three-dimensional coordinates of the key points in combination with the structure prior of the to-be-detected target, and calculating the pose of the to-be-detected target. According to the method, the high-precision three-dimensional pose of the to-be-measured target can be directly calculated from the two-dimensional visual information and the depth data without depending on complex three-dimensional modeling, the calculation precision is improved while the calculation amount is greatly reduced, and the calculation process is more stable and reliable.

Description

A 3D target detection system based on 2D perception Technical Field

[0001] This invention relates to the field of target detection, and more specifically to a three-dimensional target detection system based on two-dimensional perception. Background Technology

[0002] 3D object detection and pose estimation are key foundational technologies in fields such as intelligent manufacturing, intelligent logistics, and robotics. 3D object detection aims to accurately identify target objects in complex environments and acquire their position and orientation information in 3D space, thus providing a reliable spatial perception basis for subsequent decision-making, planning, and control. Existing 3D object detection methods can be divided into those based on dedicated 3D sensors and those based on 2D images. Methods based on dedicated 3D sensors use sensors such as LiDAR to acquire 3D point cloud data for target recognition; however, the computational load of point cloud data is large, placing high demands on real-time performance and computing resources. Methods based on 2D images predict the 3D bounding box of the target from RGB images; however, the depth estimation accuracy is often low, resulting in 3D detection results that fail to meet requirements. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention proposes a three-dimensional target detection system based on two-dimensional perception. By recognizing the target and detecting key points on an image, and combining this with a sensor capable of acquiring three-dimensional information, the system obtains the three-dimensional coordinates of the key points and calculates the pose information of the target based on these coordinates.

[0004] The objective of this invention is achieved through the following technical solution: a three-dimensional target detection system based on two-dimensional perception, comprising: a two-dimensional perception module, used for target detection and keypoint detection algorithms based on deep learning, performing feature extraction and region detection on the input image, and outputting the bounding box of the target to be detected and the two-dimensional coordinates of multiple keypoints; a three-dimensional coordinate acquisition module, used for spatially mapping the two-dimensional coordinates of multiple keypoints with the depth data collected by a depth sensor, and calculating the three-dimensional coordinates of the keypoints; and a post-processing module, used for combining the structural prior of the target to be detected, performing spatial consistency optimization on the three-dimensional coordinates of the keypoints, and calculating the pose of the target to be detected.

[0005] Furthermore, the deep learning-based target detection and pose estimation algorithm is the YOLOv11n-pose network.

[0006] Furthermore, the YOLOv11n-pose network includes a backbone feature extraction network and a multi-scale feature fusion network, and incorporates a C2PSA attention mechanism to enhance the key point feature response capability.

[0007] Furthermore, the 3D coordinate acquisition module includes a depth matching unit and an image coordinate transformation unit. The depth matching unit converts the depth information acquired by the depth camera into 3D spatial points, and then maps the 3D spatial points to the RGB camera coordinate system based on the spatial extrinsic relationship between the depth camera and the RGB camera. It also uses the RGB camera intrinsic parameters to complete pixel projection, thereby achieving accurate alignment of the depth information in the RGB camera coordinate system and establishing a correspondence between the 2D keypoint pixel coordinates and the depth data. The image coordinate transformation unit is used to map the 2D pixel coordinates of the key points output by the 2D perception module to spatial rays in the RGB camera coordinate system based on the intrinsic parameter matrix of the RGB camera, and calculates the 3D coordinates of the key points in combination with the acquired depth data.

[0008] Furthermore, the post-processing module includes a spatial consistency correction unit, a geometric constraint optimization unit, a pose solving unit, and a robustness enhancement unit. The spatial consistency correction unit performs spatial filtering and consistency adjustment on the 3D coordinates of key points output by the 3D feature acquisition module. The geometric constraint optimization unit further introduces the geometric constraint model of the target to be detected based on the spatially corrected key point set, and performs global optimization on the 3D coordinates of the key points. The pose solving unit calculates the pose of the target to be detected in the world coordinate system based on the 3D coordinate set of key points of the target to be detected after geometric constraint optimization. The robustness enhancement unit smooths and detects anomalies in the pose results of the target to be detected in consecutive frames by introducing a temporal filtering mechanism and a dynamic weight update strategy. When a sudden pose change is detected, prediction compensation is performed based on historical poses and motion trajectories.

[0009] Furthermore, the pose solving unit adopts a rigid body transformation solution algorithm based on singular value decomposition to spatially register the key point set with the standard model of the target to be detected, and outputs the six-degree-of-freedom pose parameters of the target to be detected.

[0010] Furthermore, the three-dimensional target detection system also includes a PnP algorithm verification module, which is used to verify the accuracy, evaluate the error, and calibrate the system for the estimation results of the pose of the target to be detected.

[0011] Furthermore, the PnP algorithm verification module includes a calibration marker detection unit, a pose calculation unit, a result comparison unit, and an accuracy evaluation unit. The calibration marker detection unit extracts the pixel coordinates of the four corner points of the AprilTag QR code based on an image of the AprilTag QR code captured in real-time by the image acquisition device. The pose calculation unit calculates the pose of the target in the camera coordinate system using the PnP algorithm based on the corner pixel coordinates of the AprilTag and the camera's intrinsic parameter matrix. The result comparison unit compares the pose of the target calculated by the PnP algorithm with the pose calculated based on the three-dimensional coordinates of key points. The accuracy evaluation unit calculates the average error, variance, and maximum deviation of the system under different scenarios based on the error index output by the result comparison unit, and establishes a pose accuracy evaluation model.

[0012] Furthermore, the accuracy evaluation unit is also used to trigger the correction of camera intrinsic and extrinsic parameters or depth alignment parameters when the error of the detection result exceeds a preset threshold.

[0013] Furthermore, the pose of the target to be detected in the camera coordinate system is a six-degree-of-freedom pose, including three-dimensional position coordinates as well as roll angle, pitch angle and yaw angle.

[0014] The beneficial effects of the present invention are as follows: 1. The mechanism of combining two-dimensional perception and three-dimensional coordinate acquisition in the present invention enables the system to directly deduce the high-precision three-dimensional pose of the target under test from two-dimensional visual information and depth data without relying on complex three-dimensional modeling, which greatly reduces the amount of calculation while improving the calculation accuracy.

[0015] 2. This invention utilizes the three-dimensional information from sensors to acquire depth information, and then calculates the three-dimensional position of the target. Compared with predicting the target's 3D bounding box from RGB images, this invention improves accuracy and makes the three-dimensional coordinate calculation process more stable and reliable. Attached Figure Description

[0016] Figure 1 is a schematic diagram of the structure of a three-dimensional target detection system based on two-dimensional perception according to an embodiment of the present invention.

[0017] Figure 2 shows the architecture of the YOLOv11n-pose network.

[0018] Figure 3 is a schematic diagram of the post-processing module according to an embodiment of the present invention.

[0019] Figure 4 is a schematic diagram of the internal structure of a computer device used to implement the present invention.

[0020] Figure 5 is a schematic diagram of 12 key points of the three-dimensional target detection system based on two-dimensional perception applied to the pallet of an unmanned forklift according to an embodiment of the present invention. Detailed Implementation

[0021] The present invention will be described in detail below with reference to the accompanying drawings and preferred embodiments. The purpose and effects of the present invention will become clearer. It should be understood that the specific embodiments described herein are merely for explaining the present invention and are not intended to limit the present invention.

[0022] As shown in Figure 1, the three-dimensional target detection system based on two-dimensional perception provided in this embodiment of the invention includes a two-dimensional perception module, a three-dimensional coordinate acquisition module, a post-processing module, and a PnP algorithm verification module.

[0023] 1. Two-Dimensional Perception Module: The two-dimensional perception module incorporates deep learning-based object detection and keypoint detection algorithms. It performs feature extraction and region detection on the input image, outputting the bounding box of the target and the two-dimensional coordinate information of multiple keypoints. One implementation of the deep learning-based object detection and keypoint detection algorithm is a pre-trained YOLOv11n-pose model. This model employs a backbone feature extraction network and a multi-scale feature fusion network structure, combined with a C2PSA attention mechanism to enhance keypoint feature response capabilities. The model input is a two-dimensional image, and the output is the bounding box and keypoints of the target.

[0024] The YOLOv11n-pose model is a structural optimization and lightweight design based on the traditional YOLO detection framework. The network as a whole adopts a hierarchical structure of a backbone feature extraction network and a multi-scale feature fusion network.

[0025] As shown in Figure 2, the backbone network is primarily responsible for extracting multi-level semantic features from the input image. The backbone network includes a multi-layer convolutional module (Conv), a cross-stage partial residual module (C3k2), a spatial pyramid pooling fast structure (SPPF), and a C2PSA attention mechanism module. The initial layers of the network employ convolutional operations with a stride of 2 to achieve feature downsampling and expand the receptive field. The C3k2 module effectively enhances feature reuse capabilities and reduces computational overhead through a cross-stage connection mechanism. The SPPF module further integrates global contextual information through pooling operations at different scales, improving the network's ability to perceive the complex structure of the tray. Furthermore, the C2PSA attention mechanism module introduced at a deeper level in the backbone network strengthens the feature representation of key point regions through adaptive weight allocation, enabling the model to have a higher response capability when detecting detailed features of the tray (such as fork edges and corners).

[0026] The multi-scale feature fusion network (Head) employs a multi-scale upsampling and feature concatenation strategy. By upsampling layer by layer and concatenating the feature maps from different levels of the backbone network, it achieves full fusion of features at different spatial resolutions. Subsequently, a multi-layer C3k2 structure is used to further refine the fused features, thus balancing detection accuracy and computational efficiency. Finally, the detection head outputs the tray's 2D bounding box information and the coordinates of 12 key points simultaneously through the Pose detection layer (Pose Head), achieving multi-task joint prediction.

[0027] To enhance the model's adaptability, this embodiment of the YOLOv11n-pose network offers multiple scale configurations (n, s, m, l, x), which can be flexibly selected according to actual computing resources and real-time requirements. Among them, the lightweight n-type network contains approximately 2.9 million parameters and is suitable for resource-constrained embedded platforms; while the l-type and x-type networks have better performance in detection accuracy and are suitable for server-side offline training and high-precision recognition tasks.

[0028] During the training of the YOLOv11n-pose network, a custom dataset is used for supervised learning. This dataset contains image samples of different types or materials of targets to be detected under various angles, lighting conditions, and occlusion conditions. The annotations include the 2D bounding boxes of the targets and the locations of several keypoints. The loss function consists of both target detection loss and keypoint regression loss. A joint optimization strategy balances the weights of these two losses to ensure the model achieves optimal performance between detection accuracy and keypoint localization accuracy.

[0029] 2. 3D Coordinate Acquisition Module The 3D coordinate acquisition module spatially maps the key point information detected by the 2D sensing module with the depth data collected by the depth sensor, calculating the 3D coordinate information of the key points. The depth sensor here can be, for example, a TOF (Time of Flight) camera or a LiDAR.

[0030] The 3D coordinate acquisition module includes a depth matching unit and an image coordinate transformation unit: The depth matching unit aligns the TOF camera depth data with the RGB image, thereby establishing the correspondence between the pixel coordinates of 2D keypoints and the depth data. It converts the depth information acquired by the TOF camera into 3D spatial points, and then, based on the spatial extrinsic relationship between the TOF camera and the RGB camera, maps the 3D spatial points to the RGB camera coordinate system. Finally, it uses the RGB camera intrinsic parameters to complete pixel projection, thus achieving accurate alignment of the TOF depth information in the RGB camera coordinate system.

[0031] The image coordinate transformation unit is used to transform the key point pixel coordinates (u) output by the 2D perception module into their corresponding values ​​based on the camera's intrinsic parameter matrix K. i ,v iThe spatial rays are mapped to the camera coordinate system, and the 3D coordinates of the key points are calculated by combining them with the acquired depth data.

[0032] Wherein, D(u) i ,v i (X) represents the depth value corresponding to a pixel. i ,Y i Z i () represents the 3D coordinates of the key point in the camera coordinate system.

[0033] This yields a high-precision set of three-dimensional coordinates for key points, providing reliable data support for the pose calculation of the target under test.

[0034] 3. Post-processing module: The post-processing module combines the geometric structural attributes of the target to be detected with the 3D coordinate data of key points to perform spatial consistency correction and optimization on the 3D coordinate data of the detected key points, and finally calculates the 3D pose parameters of the tray, including position coordinates and attitude angle information.

[0035] As shown in Figure 3, the post-processing module specifically includes a spatial consistency correction unit, a geometric constraint optimization unit, a pose solving unit, and a robustness enhancement unit. The spatial consistency correction unit performs spatial filtering and consistency adjustment on the 3D coordinates of key points output by the 3D feature acquisition module. Since some key points may exhibit local deviations due to lighting changes, occlusion, or depth noise during actual detection, this unit first calculates the spatial geometric relationships between all key points and then filters and corrects outliers based on the prior structural model of the target object (e.g., a pallet, where the prior model includes length, width, and height ratios, fork spacing, and pallet corner topology). For key points with large detection errors, neighborhood mean interpolation or spatial fitting based on the RANSAC algorithm can be used for compensation to ensure the overall spatial consistency of the key point set.

[0036] The geometric constraint optimization unit is used to further introduce the geometric constraint model of the target object to be detected based on the spatially corrected keypoint set, and to globally optimize the 3D coordinates of the keypoints. This unit constructs a constraint matrix between keypoints based on the theoretical structural model of the target object, such as a CAD model or design parameters, and performs nonlinear least-squares fitting on the coordinates of the detected keypoints to ensure that the spatial distance and angular relationships between keypoints are as consistent as possible with the theoretical values. The optimization objective function can be expressed as:

[0037] Among them, P i Q represents the three-dimensional coordinates of the key points obtained through measurement. irepresents the coordinates of the corresponding point in the theoretical structural model of the target to be detected; R and t are the rotation matrix and translation vector, respectively.

[0038] By iteratively optimizing the solution, the matching error of the set of three-dimensional coordinates of the key points of the detected target within the structural space is minimized.

[0039] The pose calculation unit is used to calculate the pose information of the target in the world coordinate system based on the 3D coordinate set of key points of the target after geometric constraint optimization. This unit employs a rigid body transformation algorithm based on singular value decomposition (SVD) to spatially register the key point set with the standard model of the target, outputting the six-degree-of-freedom pose parameters of the target, including 3D position coordinates (x, y, z) and attitude angle information (roll, pitch, yaw). To improve real-time performance and numerical stability, this unit introduces a quaternion attitude representation, effectively avoiding the gimbal lock problem in the Euler angle calculation process.

[0040] The robustness enhancement unit improves the system's stability in complex environments. This unit introduces a temporal filtering mechanism and a dynamic weight update strategy to smooth the pose results of consecutive frames and detect anomalies. When abrupt pose changes are detected (such as those caused by occlusion or false detections), the robustness enhancement unit performs predictive compensation based on historical poses and motion trajectories, thereby ensuring the continuity and stability of the pose output. Furthermore, this unit can combine the estimation results from the PnP algorithm verification module to perform cross-comparison and adaptive correction of the pose output, improving the overall system's accuracy and reliability.

[0041] The post-processing module achieves high-precision conversion from the three-dimensional coordinates of key points to the complete spatial pose of the target object through key point spatial consistency optimization, geometric constraint fitting, and robust pose solving. The post-processing module enables the system to maintain high detection accuracy and pose stability even in industrial environments with varying lighting, complex backgrounds, and partial occlusion.

[0042] 4. PnP Algorithm Verification Module The PnP algorithm verification module is an optional module used to verify the accuracy, evaluate errors, and calibrate the system based on the pose estimation results of the target to be detected, thereby ensuring the reliability and stability of the overall detection system in practical application environments. The PnP algorithm verification module includes a calibration and identification detection unit, a pose calculation unit, a result comparison unit, and an accuracy evaluation unit.

[0043] The calibration and identification detection unit is used to detect the pre-set AprilTag QR code on the tray. AprilTag is a robust two-dimensional visual identifier that can achieve fast and accurate corner location under complex lighting and angle conditions. This unit uses an image of the pre-set AprilTag QR code on the target object, captured in real-time by an image acquisition device, and employs a detection algorithm based on grayscale gradient and edge fitting to accurately extract the pixel coordinates of the four corner points of the QR code. The detection results are used not only for pose calculation but also as benchmark data for system calibration and accuracy verification.

[0044] The pose calculation unit uses the AprilTag corner pixel coordinates and the intrinsic parameter matrix of the RGB camera to calculate the pose parameters of the target in the RGB camera coordinate system using the Perspective-n-Point (PnP) algorithm. The PnP algorithm obtains the rotation matrix R by solving the spatial transformation relationship between a known 3D reference point and its corresponding 2D projection point. PnP With translation vector t PnP The specific calculation formula is as follows:

[0045] Where (u,v) are the image coordinates of the AprilTag corner points, (X,Y,Z) are the three-dimensional coordinates of the known marker points in the world coordinate system, and K is the camera intrinsic parameter matrix. The optimal pose calculation result is obtained by minimizing the reprojection error.

[0046] The result comparison unit is used to compare the pose (R) of the target to be detected calculated by the PnP algorithm. PnP ,t PnP The pose of the target to be detected (R) is calculated based on the 3D coordinates of key points. 3D ,t 3D This unit performs a difference comparison analysis. It calculates the deviations between the two sets of pose results in the translation and rotation dimensions, defined as follows:

[0047]

[0048] in, Indicates translation error. The expression represents the attitude angle error, and `trace()` represents the trace of the matrix.

[0049] Comparative analysis allows for a direct assessment of the accuracy and consistency of the system's pose calculation.

[0050] The accuracy evaluation unit is used to compare the error metrics output by the results unit, statistically analyze the system's average error, variance, and maximum deviation under different scenarios, and establish a pose accuracy evaluation model. If the error of the detection result exceeds a preset threshold, the system can automatically trigger a recalibration process to correct the camera's intrinsic and extrinsic parameters or depth alignment parameters, thereby achieving an automated self-calibration and performance maintenance mechanism. This unit also supports time-series-based error trend analysis, which can be used to evaluate the system's long-term operational stability and reliability.

[0051] Each module in the aforementioned 3D target detection device based on 2D perception can be implemented entirely or partially through software, hardware, or a combination thereof. These modules can be embedded in or independent of the processor in a computer device, or stored in the memory of a computer device as software, so that the processor can call and execute the corresponding operations of each module.

[0052] Another embodiment of the present invention provides a computer device, which may be a server, and its internal structure is shown in Figure 4. The computer device includes a processor, memory, and a network interface connected via a system bus. The processor provides computing and control capabilities. The memory includes a non-volatile storage medium and internal memory. The non-volatile storage medium stores an operating system, computer programs, and a database. The internal memory provides an environment for the operation of the operating system and computer programs in the non-volatile storage medium. The database stores recorded two-dimensional images, three-dimensional sensor data, and network model data. The network interface communicates with external terminals via a network connection. When the computer program is executed by the processor, it implements a three-dimensional target detection method based on an attention mechanism.

[0053] The following section will further explain and illustrate the three-dimensional target detection system based on two-dimensional perception of the present invention using a specific application scenario of an unmanned forklift pallet.

[0054] In this scenario, combining 2D perception and 3D coordinate acquisition enables 3D detection of pallets by unmanned forklifts. The 2D perception module performs target detection and key point localization on the pallet, extracting the pallet's 2D bounding box and the coordinates of 12 key points. A TOF camera or LiDAR is then used to acquire the 3D coordinates of the pallet's key points, and a post-processing module performs spatial consistency optimization to calculate the pallet's 3D pose. Finally, the PnP algorithm is used for pose estimation verification. By comparing the pose of the AprilTag QR code pre-attached to the pallet with the calculated 3D coordinates of the key points, the accuracy and stability of the pose calculation are ensured.

[0055] Figure 5 shows the 12 key points of the tray. A YOLOv11n-pose-based network structure was pre-trained to jointly detect the tray's 2D bounding box and key points, enabling the simultaneous output of the tray's 2D bounding box recognition and the 2D coordinates of the 12 key points. The training dataset included image samples of trays made of different materials (plastic, iron, and wood) under different angles, lighting conditions, and occlusion conditions. The annotations included the tray's 2D bounding box and the positions of the 12 key points. In actual operation, the pre-trained YOLOv11n-pose network structure receives real-time image data from the camera and quickly outputs the tray's 2D bounding box and the 2D coordinates of the key points, providing the basic input for subsequent 3D coordinate calculations.

[0056] In addition, in this embodiment, the post-processing module provides a reliable visual perception foundation for the automated docking and handling tasks of the unmanned forklift. The post-processing module also interacts with the motion control system of the unmanned forklift, transmitting the calculated pallet pose parameters to the control terminal in real time to guide the forklift in performing precise path planning and fork docking operations.

[0057] It will be understood by those skilled in the art that the above descriptions are merely preferred examples of the invention and are not intended to limit the invention. Although the invention has been described in detail with reference to the foregoing examples, those skilled in the art can still modify the technical solutions described in the foregoing examples or make equivalent substitutions for some of the technical features. All modifications and equivalent substitutions made within the spirit and principles of the invention should be included within the scope of protection of the invention.

Claims

1. A three-dimensional target detection system based on two-dimensional perception, characterized in that, The system includes: a 2D perception module, which uses deep learning-based target detection and keypoint detection algorithms to extract features and detect regions from the input image, and outputs the bounding box of the target to be detected and the 2D coordinates of multiple keypoints; a 3D coordinate acquisition module, which spatially maps the 2D coordinates of multiple keypoints with the depth data collected by the depth sensor, and calculates the 3D coordinates of the keypoints; and a post-processing module, which combines the structural prior of the target to be detected to optimize the spatial consistency of the 3D coordinates of the keypoints and calculate the pose of the target to be detected.

2. The three-dimensional target detection system based on two-dimensional perception according to claim 1, characterized in that, The deep learning-based target detection and pose estimation algorithm is the YOLOv11n-pose network.

3. The three-dimensional target detection system based on two-dimensional perception according to claim 2, characterized in that, The YOLOv11n-pose network includes a backbone feature extraction network and a multi-scale feature fusion network, and incorporates a C2PSA attention mechanism to enhance the key point feature response capability.

4. The three-dimensional target detection system based on two-dimensional perception according to claim 1, characterized in that, The three-dimensional coordinate acquisition module includes a depth matching unit and an image coordinate transformation unit. The depth matching unit converts the depth information acquired by the depth camera into three-dimensional spatial points, and then maps the three-dimensional spatial points to the RGB camera coordinate system based on the spatial extrinsic relationship between the depth camera and the RGB camera. It also uses the RGB camera intrinsic parameters to complete pixel projection, thereby achieving accurate alignment of depth information in the RGB camera coordinate system and establishing the correspondence between the two-dimensional key point pixel coordinates and the depth data. The image coordinate transformation unit is used to map the two-dimensional pixel coordinates of the key points output by the two-dimensional perception module to the spatial ray in the RGB camera coordinate system according to the intrinsic parameter matrix of the RGB camera, and calculate the three-dimensional coordinates of the key points in combination with the acquired depth data.

5. The three-dimensional target detection system based on two-dimensional perception according to claim 1, characterized in that, The post-processing module includes a spatial consistency correction unit, a geometric constraint optimization unit, a pose solving unit, and a robustness enhancement unit. The spatial consistency correction unit performs spatial filtering and consistency adjustment on the 3D coordinates of key points output by the 3D feature acquisition module. The geometric constraint optimization unit further introduces the geometric constraint model of the target to be detected based on the spatially corrected key point set, and performs global optimization on the 3D coordinates of the key points. The pose solving unit calculates the pose of the target to be detected in the world coordinate system based on the 3D coordinate set of key points of the target to be detected after geometric constraint optimization. The robustness enhancement unit smooths and detects anomalies in the pose results of the target to be detected in consecutive frames by introducing a temporal filtering mechanism and a dynamic weight update strategy. When a sudden pose change is detected, prediction compensation is performed based on historical poses and motion trajectories.

6. The three-dimensional target detection system based on two-dimensional perception according to claim 5, characterized in that, The pose solving unit uses a rigid body transformation algorithm based on singular value decomposition to spatially register the set of key points with the standard model of the target to be detected, and outputs the six-degree-of-freedom pose parameters of the target to be detected.

7. The three-dimensional target detection system based on two-dimensional perception according to claim 1, characterized in that, The three-dimensional target detection system also includes a PnP algorithm verification module, which is used to verify the accuracy, evaluate the error, and calibrate the system for the estimation results of the pose of the target to be detected.

8. The three-dimensional target detection system based on two-dimensional perception according to claim 7, characterized in that, The PnP algorithm verification module includes a calibration marker detection unit, a pose calculation unit, a result comparison unit, and an accuracy evaluation unit. The calibration marker detection unit is used to extract the pixel coordinates of the four corner points of the AprilTag QR code based on the image of the AprilTag QR code on the target to be detected captured in real time by the image acquisition device, using a detection algorithm based on gray-level gradient and edge fitting. The pose calculation unit is used to calculate the pose of the target to be detected in the camera coordinate system based on the corner pixel coordinates of the AprilTag and the intrinsic parameter matrix of the camera, using the PnP algorithm. The result comparison unit is used to compare and analyze the differences between the pose of the target to be detected calculated by the PnP algorithm and the pose of the target to be detected calculated based on the three-dimensional coordinates of key points; the accuracy evaluation unit is used to statistically analyze the average error, variance and maximum deviation of the system under different scenarios based on the error index output by the result comparison unit, and to establish a pose accuracy evaluation model.

9. The three-dimensional target detection system based on two-dimensional perception according to claim 8, characterized in that, The accuracy evaluation unit is also used to trigger the correction of camera intrinsic and extrinsic parameters or depth alignment parameters when the error of the detection result exceeds a preset threshold.

10. The three-dimensional target detection system based on two-dimensional perception according to claim 8, characterized in that, The pose of the target to be detected in the camera coordinate system is a six-degree-of-freedom pose, including three-dimensional position coordinates as well as roll angle, pitch angle and yaw angle.

Images