Three-dimensional object detection method based on two-dimensional background enhancement

Abstract

The application discloses a three-dimensional target detection method based on two-dimensional background enhancement, which comprises the following steps: acquiring multiple two-dimensional images collected at different time points of each fixed spatial position of a target region, and grouping two-dimensional images with the same spatial position into a group; inputting the images into a pre-trained two-dimensional target detection model to obtain a two-dimensional boundary box of each detected dynamic target in the images; performing pixel-level elimination on the detected dynamic targets based on semantic segmentation to generate a static background image of a group of eliminated dynamic targets of each spatial position, and then performing background image synthesis; training a three-dimensional target detector comprising a two-dimensional image feature extraction network, a two-dimensional visual depth estimation module, a three-dimensional voxel backbone network and a three-dimensional detection head based on the synthesized background image; collecting target region images to be detected and the synthesized background image, inputting the images into the trained three-dimensional target detector, and outputting a three-dimensional target detection box. The three-dimensional target detection method has high detection precision.

Description

Technical Field

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

[0002] In intelligent monitoring and autonomous driving applications in closed or semi-closed environments (such as industrial parks, ports, and mining areas), vision-based 3D target detection technology is crucial. However, existing methods typically perform detection directly on monocular images or video sequences, which faces significant challenges in practical applications. These dynamic scenes not only contain dynamic targets such as vehicles and pedestrians of interest but also often include complex static background interference such as numerous parked fences. These factors make it difficult for detection models to accurately distinguish between true dynamic targets and background interference, leading to decreased robustness in 3D detection and a high likelihood of false detections. Therefore, how to effectively utilize prior scene knowledge to suppress dynamic interference and improve 3D detection performance has become a pressing technical challenge. Summary of the Invention

[0003] To address the shortcomings of existing technologies, this invention provides a three-dimensional target detection method based on two-dimensional background enhancement, the specific technical solution of which is as follows:

[0004] A three-dimensional target detection method based on two-dimensional background enhancement, the method comprising the following steps:

[0005] S1: Acquire multiple two-dimensional images at different time points at each fixed spatial location in the target area, and group all two-dimensional images with the same spatial coordinates into a group to form the image sample set corresponding to that spatial location;

[0006] S2: Input each image in the image sample set into the pre-trained two-dimensional target detection model to obtain the two-dimensional bounding box of each detected dynamic target in the image;

[0007] S3: Based on semantic segmentation, the detected dynamic targets are removed at the pixel level, generating a set of static background image samples with removed dynamic targets at each spatial location;

[0008] S4: Perform background image synthesis on a set of static background image samples with dynamic targets removed from each spatial location obtained in S3;

[0009] S5: Based on the background image synthesized in S4, train a vision-based 3D target detector; the 3D target detector includes a 2D image feature extraction network, a 2D visual depth estimation module, a 3D voxel backbone network, and a 3D detection head; the input of the 3D target detector is the currently acquired image and the background image at the corresponding location, and the output is the detection box of the 3D target;

[0010] S6: Acquire the image of the target region to be detected and the synthesized background image, input them into the trained 3D target detector, and output the 3D target detection box.

[0011] Furthermore, the pre-trained two-dimensional object detection model is an open vocabulary detection model based on the Transformer architecture.

[0012] Further, step S3 includes the following sub-steps:

[0013] S3.1: For each image in the image sample set corresponding to each fixed spatial location, input the two-dimensional bounding box of each dynamic target in the image output by the pre-trained two-dimensional target detection model into the pre-trained semantic segmentation model to obtain the pixel-level segmentation mask of the dynamic target within the bounding box.

[0014] S3.2: Based on the segmentation mask, remove dynamic targets from the original two-dimensional image to obtain a set of static background image samples with dynamic targets removed corresponding to each fixed spatial location.

[0015] Further, step S4 includes the following sub-steps:

[0016] S4.1: Perform a pixel-by-pixel summation operation on a set of static background image samples with dynamic targets removed at each fixed spatial location to obtain the cumulative RGB value of each pixel position in all samples;

[0017] S4.2: Calculate the average of the accumulated results to obtain the final background image for each fixed spatial location.

[0018] Furthermore, the two-dimensional image feature extraction network is a two-dimensional spatial backbone network that receives six-channel input. The input of the two-dimensional spatial backbone network is the stitched six-channel currently acquired image and the background image at the corresponding position. The output of the two-dimensional spatial backbone network is two-dimensional visual features.

[0019] Furthermore, the two-dimensional image feature extraction network is a combination of two identical two-dimensional backbone networks and a feature fusion module; the inputs of the two two-dimensional backbone networks are the currently acquired image and the background image at the corresponding location, respectively, and the output is two-dimensional visual features; the feature fusion module is a convolutional layer or Transformer structure, used to perform deep fusion of the two two-dimensional visual features, and the fused features contain both the dynamic information of the current scene and the stable features of the static background.

[0020] Furthermore, the loss function used to train the 3D target detector includes a focus loss function for predicting 3D coordinates, a Smooth L1 loss for predicting the size and orientation angle of the 3D target, and a cross-union ratio loss for predicting the degree of overlap between the 3D bounding box and the ground truth box.

[0021] A three-dimensional target detection device based on two-dimensional background enhancement includes one or more processors for implementing a three-dimensional target detection method based on two-dimensional background enhancement.

[0022] An electronic device, comprising:

[0023] One or more processors;

[0024] A storage device for storing one or more programs that, when executed by the electronic device, enable the electronic device to implement a three-dimensional target detection method based on two-dimensional background enhancement.

[0025] A computer-readable storage medium having a program stored thereon that, when executed by a processor, implements a three-dimensional target detection method based on two-dimensional background enhancement.

[0026] The beneficial effects of this invention are as follows:

[0027] This invention fundamentally changes the traditional passive processing mode of complex scenes by systematically constructing a clean two-dimensional background image library of the target scene and using it as key prior information to deeply fuse with real-time images. Compared with existing technologies, its core advantages are reflected in two aspects:

[0028] First, at the data level, through the innovative process of "multiple acquisitions - precise positioning - dynamic target detection and pixel-level removal - multi-image averaging synthesis", high-quality, static background images without dynamic interference can be generated, providing a reliable scene background reference for the detector;

[0029] Secondly, at the model level, the detector can actively compare the differences between the current scene and the static background, thereby significantly enhancing its ability to distinguish real dynamic targets and effectively suppressing false detections and missed detections caused by complex backgrounds and residual interference. Attached Figure Description

[0030] Figure 1 This is a flowchart of the three-dimensional target detection method based on two-dimensional background enhancement according to the present invention.

[0031] Figure 2 This is a flowchart illustrating one implementation method for obtaining multiple two-dimensional images of each fixed spatial location in the target area in step one.

[0032] Figure 3 This is a flowchart illustrating the execution of an embodiment of the present invention when a channel merging strategy is employed.

[0033] Figure 4 This is a flowchart illustrating the execution of an embodiment of the present invention when employing a twin network fusion strategy.

[0034] Figure 5 This is a schematic diagram of a three-dimensional target detection device based on two-dimensional background enhancement, according to another embodiment of the present invention. Detailed Implementation

[0035] 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.

[0036] The three-dimensional target detection method based on two-dimensional background enhancement of this invention is particularly suitable for dynamic target detection tasks such as vehicles and pedestrians in fixed-range areas (such as industrial parks, ports, and mining areas), mainly to maintain the static background unchanged in the short term. This method forms a complete technical process from background data acquisition and processing, two-dimensional target detection and semantic segmentation, background image synthesis, to vision-based three-dimensional target detector design and training.

[0037] like Figure 1 As shown, the three-dimensional target detection method based on two-dimensional background enhancement in this embodiment includes the following steps one through five.

[0038] Step 1: Acquire multiple two-dimensional images of each fixed spatial location in the target area, and group all two-dimensional images with the same spatial coordinates into a set to form the image sample set corresponding to that spatial location.

[0039] like Figure 2 As shown, when acquiring two-dimensional images at each fixed spatial location within the target area, a data acquisition system needs to be deployed. This data acquisition system should include:

[0040] 1) A mobile platform that can travel automatically or manually along a fixed route in a loop;

[0041] 2) At least one image acquisition device disposed on the mobile platform for acquiring a two-dimensional image of the target area;

[0042] 3) A positioning and attitude determination module installed on the mobile platform is used to acquire the spatial position and attitude of the mobile platform in the target area in real time.

[0043] Simultaneously, a fixed acquisition route covering key areas needs to be planned and set for the data acquisition system within the target area. This fixed route must cover all key areas within the target area where 3D target detection needs to be performed, including but not limited to roads, intersections, parking areas, and sidewalks—areas where 3D dynamic targets frequently appear. Then, the mobile platform is controlled to circulate along the fixed acquisition route at a constant low speed. During the movement of the mobile platform, based on preset spatial or time interval trigger conditions, a single 2D image is acquired through its onboard image acquisition device, and the spatial coordinates of the mobile platform at the time of image acquisition are simultaneously recorded by its positioning and attitude determination module.

[0044] The spatial interval trigger condition is to trigger a data acquisition once every predetermined distance (e.g., 0.5 meters); the time interval trigger condition is to trigger a data acquisition once every fixed time interval (e.g., 0.1 seconds).

[0045] The mobile platform is run multiple times along a fixed acquisition route to obtain multiple 2D images at different time points for each fixed spatial location along the route. Ultimately, for each fixed spatial location, at least 50 images are acquired, preferably 100 or more. Based on data recorded by the positioning and orientation module on the mobile platform, each acquired 2D image is assigned precise spatial coordinates. All 2D images with the same spatial coordinates are then grouped together to form an image sample set corresponding to each spatial location.

[0046] Step 2: Input each image in the image sample set into the pre-trained two-dimensional object detection model to perform two-dimensional object detection, and obtain the two-dimensional bounding box of each detected dynamic object in the image.

[0047] The pre-trained 2D object detection model is an open vocabulary detection model based on the Transformer architecture, such as Grounding DINO.

[0048] Step 3: Based on semantic segmentation, pixel-level removal of detected dynamic targets is performed to generate multiple static background image samples with removed dynamic targets for each spatial location.

[0049] Step three includes the following sub-steps:

[0050] S3.1: For each image in the image sample set corresponding to each fixed spatial location, input the two-dimensional bounding box of each dynamic target in the image output by the pre-trained two-dimensional target detection model into the pre-trained semantic segmentation model to obtain the pixel-level segmentation mask of the dynamic target within the bounding box.

[0051] Pre-trained semantic segmentation models are zero-shot segmentation models, such as the Segment Anything Model (SAM).

[0052] S3.2: Based on the segmentation mask, remove dynamic targets from the original two-dimensional image to obtain a set of static background image samples with dynamic targets removed corresponding to each fixed spatial location.

[0053] Here, based on the segmentation mask, the RGB values ​​of the pixel regions belonging to dynamic targets in the original 2D image are modified to predetermined values, thereby generating an image that retains only the static background. The predetermined value is preferably set to 0, i.e., black.

[0054] After performing the above processing on each image in the image sample set corresponding to the same fixed spatial location, a set of static background image samples corresponding to that fixed spatial location, with dynamic targets removed, is obtained.

[0055] Step 4: Perform background image synthesis on a set of static background image samples with dynamic targets removed from each spatial location obtained in Step 3.

[0056] Step four includes the following sub-steps:

[0057] S4.1: Perform a pixel-by-pixel summation operation on a set of static background image samples (e.g., 50 images) at each fixed spatial location after removing dynamic targets, to obtain the cumulative RGB value of each pixel location across all samples.

[0058] S4.2: Calculate the average of the accumulated results, that is, divide the accumulated RGB values ​​by the number of images to obtain the final background image for each fixed spatial location.

[0059] This background image represents the most common static scene appearance at this spatial location under conditions of no dynamic target interference (such as the inherent texture of roads, the fixed color of buildings, the normal distribution of vegetation, etc.).

[0060] By iteratively processing all collection points (e.g., 4000 locations) within a fixed range, a set of background images corresponding one-to-one with each collection point is finally generated. For example, 4000 background images can be prioritized by region, and key regions such as road intersections and high-frequency target occurrence points can be selected to synthesize background images.

[0061] Step 5: Based on the background image synthesized in Step 4, train a vision-based 3D object detector.

[0062] The 3D object detector includes a 2D image feature extraction network, a 2D visual depth estimation module, a 3D voxel backbone network, and a 3D detection head. The input of the 3D object detector is the currently acquired image and the background image at the corresponding location, and the output is the detection box of the 3D object.

[0063] This step is crucial for improving the accuracy of 3D target detection. There are two ways to implement the 3D target detector here:

[0064] One approach is a channel merging strategy, where the 3D target detector comprises a 2D spatial backbone network receiving six channels of input, a 2D visual depth estimation module, a 3D voxel backbone network, and a 3D detection head. For example... Figure 3 As shown, the original acquired image is a three-channel RGB image (3 channels), and the background image is also a three-channel RGB image (3 channels). The currently acquired image and the background image at the corresponding location (matched according to the positioning information) are concatenated along the channel dimension to form a six-channel input image (3+3=6 channels). The input layer of the two-dimensional backbone network (such as ResNet, VGG, or a lightweight CNN optimized for BEV tasks) is adjusted accordingly to receive six-channel input, while other network structure parameters such as kernel size, stride, and pooling method remain unchanged.

[0065] A two-dimensional backbone network is used for feature extraction from six-channel two-dimensional images;

[0066] Two-dimensional visual depth estimation modules (such as depth map generation modules based on IPM inverse perspective transformation or deep learning estimation) are used to convert two-dimensional features into depth information in three-dimensional voxel space.

[0067] A three-dimensional voxel backbone network (such as 3D ResNet, SparseConvNet, etc.) further processes the voxel features and finally connects to a three-dimensional detection head (such as an anchor-based or anchor-free detection head) to output a three-dimensional target detection box, including the center coordinates, size and orientation of the detection box.

[0068] The advantage of this type of 3D target detector is its simplicity. It directly utilizes multi-channel input to fuse background and current image information without requiring an additional feature fusion module.

[0069] Another type of 3D target detector, such as Figure 4As shown, a strategy of using a Siamese network and a feature fusion module (diff module) is adopted to improve the utilization rate of background information through explicit feature interaction. This 3D object detector includes two identical 2D backbone networks (i.e., "Siamese networks"), a feature fusion module (diff module), a 2D visual depth estimation module, a 3D voxel backbone network, and a 3D detection head.

[0070] The two identical two-dimensional backbone networks are input to the currently acquired image and the background image at the corresponding location, respectively. The two two-dimensional backbone networks independently extract the two-dimensional visual features of their respective images (the feature map size and the number of channels are kept consistent).

[0071] The feature fusion module performs deep fusion of two feature maps using convolutional layers (conv) or Transformer structures. For example, it can employ channel-wise differencing (calculating the difference between current image features and background features to highlight the changing parts of dynamic targets), concatenated convolution (merging the two features and extracting interaction information through multiple convolutional layers), or attention-based weighted fusion (calculating attention weights between the two features using a Transformer encoder to dynamically adjust the fusion ratio). The fused features contain both dynamic information of the current scene and stable features of the static background, thereby enhancing the network's ability to distinguish dynamic targets.

[0072] The fused features are then fed into a 2D visual depth estimation module, a 3D voxel backbone network, and a 3D detection head to output the final 3D target detection bounding box.

[0073] Both of the aforementioned 3D object detectors use 2D visual depth estimation modules (such as the depth estimation module in the BEVDepth model) to project 2D features into a 3D voxel space. BEVDepth maps feature points in the 2D image to a voxel grid in the 3D world coordinate system through camera intrinsic and extrinsic parameter calibration and a depth prediction network, forming a voxel feature representation with depth information. Subsequently, a 3D voxel backbone network (such as a 3D CNN based on sparse convolution) models the spatial context of the voxel features and extracts multi-scale features. Finally, a 3D detection head (such as a detection head based on CenterPoint or an improved PV-RCNN) predicts the 3D bounding box parameters of the target based on the feature map, including the center point coordinates (x, y, z), length, width, and height dimensions (l, w, h), and the driving direction angle (yaw, i.e., the angle between the target's orientation and the coordinate axes).

[0074] When training a vision-based 3D object detector, a multi-task joint optimization strategy is adopted for the loss function to address the specific characteristics of the 3D detection task:

[0075] The prediction of 3D coordinates (x, y, z) adopts Focal Loss, and the depth estimation accuracy is improved by using the supervision signals of different levels of the feature pyramid.

[0076] The prediction of length, width, and height dimensions (l, w, h) and driving direction angle (yaw) uses Smooth L1 Loss (to smooth the regression error and avoid gradient explosion).

[0077] The degree of overlap between the 3D bounding box and the ground truth box (i.e., detection accuracy) is determined using IoU Loss (Intersection over Union Loss, which directly optimizes the matching degree of the detection boxes).

[0078] By combining the above loss functions, we can ensure that the model achieves high accuracy in localization, size estimation, and orientation prediction in three-dimensional space.

[0079] During training, network parameters are iteratively updated through backpropagation and optimizers (such as AdamW or SGD).

[0080] Step 6: Acquire images of the target region to be detected and the synthesized background image, input them into the trained 3D target detector, and output the 3D target detection bounding box.

[0081] like Figure 5 As shown, on the other hand, the present invention also provides an embodiment of a three-dimensional target detection device based on two-dimensional background enhancement, the device including one or more processors for implementing the three-dimensional target detection method based on two-dimensional background enhancement in the above embodiment.

[0082] The embodiments of the three-dimensional target detection device based on two-dimensional background enhancement of the present invention can be applied to any device with data processing capabilities, such as a computer. The device embodiments can be implemented in software, hardware, or a combination of both. Taking software implementation as an example, as a logical device, it is formed by the processor of the device with data processing capabilities loading the corresponding computer program instructions from non-volatile memory into memory for execution. From a hardware perspective, in addition to the processor, memory, network interface, and non-volatile memory, the device with data processing capabilities in the embodiments may also include other hardware depending on its actual functions, which will not be elaborated further.

[0083] The specific implementation process of the functions and roles of each unit in the above device can be found in the implementation process of the corresponding steps in the above method, and will not be repeated here.

[0084] For the device embodiments, since they basically correspond to the method embodiments, the relevant parts can be referred to in the description of the method embodiments. The device embodiments described above are merely illustrative. The units described as separate components may or may not be physically separate, and the components shown as units may or may not be physical units, that is, they may be located in one place or distributed across multiple network units. Some or all of the modules can be selected to achieve the purpose of the present invention according to actual needs. Those skilled in the art can understand and implement this without creative effort.

[0085] This invention also provides a computer-readable storage medium storing a program thereon, which, when executed by a processor, implements the three-dimensional target detection method based on two-dimensional background enhancement described in the above embodiments.

[0086] The computer-readable storage medium can be an internal storage unit of any data processing device described in any of the foregoing embodiments, such as a hard disk or memory. The computer-readable storage medium can also be an external storage device, such as a plug-in hard disk, smart media card (SMC), SD card, flash card, etc., equipped on the device. Furthermore, the computer-readable storage medium can include both internal storage units of any data processing device and external storage devices. The computer-readable storage medium is used to store the computer program and other programs and data required by the data processing device, and can also be used to temporarily store data that has been output or will be output.

[0087] The effectiveness of the method of the present invention will be verified through a specific embodiment below.

[0088] Within a company's closed campus, which spans approximately 2 kilometers, data collection points were set up along the campus roads at 0.5-meter intervals, totaling about 4,000 potential background locations. A specially designed data collection vehicle circulated along a fixed route within the target closed area, collecting image data. For each location, the vehicle circulated and collected images up to 50 times, ultimately obtaining approximately 4,000 × 50 = 200,000 raw image data points. From these 4,000 locations, key regions were selected, resulting in 4,000 high-quality background images. These images served as the core input for background enhancement in the subsequent 3D object detector. Combining the original training dataset (containing dynamic object annotations and 3D ground truth), the BEVDepth model is used as the basic framework. (BEVDepth is a 3D object detector based on bird's-eye view (BEV) representation. Its core is to convert 2D image features into 3D voxel features through depth estimation and achieve efficient detection through convolutional networks in BEV space. In this embodiment, the front end of the 3D object detector uses ResNet-50 as a 2D backbone network to extract image features; a depth map is generated through IPM or a depth prediction module, which combines 2D features with depth information and maps them to 3D voxel space; the back end uses a sparse convolutional network to process voxel features, and finally outputs 3D bounding box parameters through a multi-scale feature fusion detection head.)

[0089] This embodiment employs two 3D object detectors in its experiments. The results show that, compared to traditional 3D object detection methods without background enhancement (i.e., directly inputting the currently acquired image for detection), the first 3D object detector based on a channel merging strategy improves the 3D mAP (3D average accuracy) by 6.9 percentage points, while the second, a 3D object detector using a Siamese network + diff module, improves the 3D mAP by 8.2 percentage points. This significant improvement validates the effective suppression of dynamic interference and the full utilization of static background information by 2D background enhancement technology, thereby greatly improving the accuracy and robustness of 3D object detection.

[0090] 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 object detection method based on two-dimensional background enhancement, characterized in that, The method comprises the following steps: S1: acquiring a plurality of two-dimensional images collected at different time points for each fixed spatial position of a target region, and grouping all two-dimensional images with the same spatial position coordinates into a set to form an image sample set corresponding to the spatial position; S2: inputting each image in the image sample set into a pre-trained two-dimensional target detection model to obtain a two-dimensional bounding box of each detected dynamic target in the image; S3: performing pixel-level elimination on the detected dynamic target based on semantic segmentation to generate a set of static background image samples of the eliminated dynamic target for each spatial position; S4: synthesizing the set of static background image samples of the eliminated dynamic target for each spatial position obtained in S3; S5: training a three-dimensional target detector based on vision based on the synthesized background image in S4; the three-dimensional target detector comprises a two-dimensional image feature extraction network, a two-dimensional visual depth estimation module, a three-dimensional voxel backbone network, and a three-dimensional detection head; The input of the three-dimensional target detector is a current acquisition image and a background image at a corresponding position, and the output is a detection box of a three-dimensional target; S6: acquiring a target region image to be detected and a synthesized background image, and inputting them into the trained three-dimensional target detector to output a three-dimensional target detection box.

2. The method of claim 1, wherein, The pre-trained two-dimensional target detection model is an open vocabulary detection model based on a Transformer architecture. 3.The 3D object detection method based on 2D background enhancement of claim 1, wherein, S3 comprises the following sub-steps: S3.1: for each image in the image sample set corresponding to each fixed spatial position, inputting the two-dimensional bounding box of each dynamic target in the image output by the pre-trained two-dimensional target detection model into a pre-trained semantic segmentation model to obtain a pixel-level segmentation mask of the dynamic target in the bounding box; S3.2: based on the segmentation mask, eliminating the dynamic target from the original two-dimensional image to obtain a set of static background image samples of the eliminated dynamic target for each fixed spatial position.

4. The method of claim 1, wherein, S4 comprises the following sub-steps: S4.1: performing pixel-by-pixel summation operation on the set of static background image samples of the eliminated dynamic target for each fixed spatial position to obtain the RGB value accumulation result of each pixel position in all samples; S4.2: performing average calculation on the accumulation result to obtain the final background image of each fixed spatial position.

5. The method of claim 1, wherein, The two-dimensional image feature extraction network is a two-dimensional spatial backbone network receiving six-channel input, the input of the two-dimensional spatial backbone network is the six-channel current acquisition image and the background image at the corresponding position after splicing, and the output of the two-dimensional spatial backbone network is a two-dimensional visual feature.

6. The method of claim 1, wherein, The two-dimensional image feature extraction network is a combination of two two-dimensional backbone networks with the same structure and a feature fusion module; the inputs of the two two-dimensional backbone networks are the current acquisition image and the background image at the corresponding position, and the outputs are two-dimensional visual features; the feature fusion module is a convolution layer or a Transformer structure, which is used for deep fusion of the two two-dimensional visual features, and the fused features contain both dynamic information of the current scene and stable features of the static background.

7. The method of claim 1, wherein the three-dimensional object detection based on two-dimensional background enhancement is characterized by, The loss function in training the three-dimensional target detector includes a focal loss function for predicting three-dimensional coordinates, a Smooth L1 loss for predicting three-dimensional target sizes and direction angles, and an intersection over union loss for predicting the degree of overlap between a predicted three-dimensional bounding box and a real box.

8. A three-dimensional object detection device based on two-dimensional background enhancement, characterized by, The electronic device comprises one or more processors for implementing the two-dimensional background enhanced three-dimensional target detection method in any one of claims 1-7.

9. An electronic device, comprising: The electronic device comprises: one or more processors; a storage device for storing one or more programs, which, when executed by the electronic device, cause the electronic device to implement the two-dimensional background enhanced three-dimensional target detection method in any one of claims 1-7.

10. A computer-readable storage medium, characterized in that, A storage device having stored thereon a program, which, when executed by a processor, implements the two-dimensional background enhanced three-dimensional target detection method in any one of claims 1-7.

Images