A method for efficiently training a 3D target detection task model for service lighthouse cognition

Abstract

This application provides an efficient training method for a 3D object detection task model serving lighthouse cognition, relating to the field of autonomous driving technology. The method includes: processing an RGB image using a preset number of pre-trained first 3D object detection task models from previous times to obtain object detection results; superimposing the pre-defined object category annotations of the RGB image with the object detection results to obtain training data; training a second 3D object detection task model at the current time using the training data to obtain a trained second 3D object detection task model at the current time; and updating the first 3D object detection task model from the previous time using the trained second 3D object detection task model to obtain a trained first 3D object detection task model at the current time. This method solves the catastrophic forgetting problem in model training through continuous learning.

Description

Technical Field

[0001] This application relates to the field of autonomous driving technology, and in particular to an efficient training method for a 3D object detection task model for service lighthouse cognition. Background Technology

[0002] Currently, roadside visual 3D object detection models are typically trained using data from different tasks in a certain order, and then incrementally updated based on the task execution results. However, when learning new tasks, AI agents are prone to catastrophic forgetting, meaning they forget the knowledge learned from older tasks. For example, a model trained sequentially on roadside car data and truck data may fail to accurately detect cars in subsequent roadside data. Summary of the Invention

[0003] In view of this, this application provides an efficient training method for a 3D object detection task model that serves lighthouse cognition, in order to solve the technical problem of catastrophic forgetting in the training of roadside vision 3D object detection task models.

[0004] In a first aspect, embodiments of this application provide an efficient training method for a 3D object detection task model for lighthouse cognition, the method comprising:

[0005] Acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment;

[0006] The first 3D object detection task model, trained to a predetermined number of times prior to the current moment, is used to process the RGB image to obtain the object detection result;

[0007] The training data is obtained by overlaying the predefined target category annotations and target detection results of the RGB images;

[0008] The training data is used to train the second 3D object detection task model at the current moment, resulting in the trained second 3D object detection task model at the current moment; wherein the second 3D object detection task model at the current moment is obtained by copying the first 3D object detection task model at the previous moment after training.

[0009] The first 3D object detection task model, which was trained at the previous time step, is updated using the second 3D object detection task model trained at the current time step to obtain the first 3D object detection task model trained at the current time step.

[0010] Furthermore, the target detection results include the categories of multiple targets and the position and size of the detection boxes.

[0011] Furthermore, the preset target categories include at least one of the following: cars, trucks, vans, buses, motorcycles, bicycles, and pedestrians; the annotation results of the preset target categories of the RGB images include the position and size of the detection boxes of the preset target categories.

[0012] Furthermore, the RGB image is processed using a pre-defined number of trained 3D object detection task models prior to the current time step to obtain object detection results, including:

[0013] Obtain the first 3D object detection task model at different times after multiple training completions;

[0014] The RGB image is processed using the first 3D object detection task model at each time step after training, resulting in M ​​detection boxes;

[0015] Cluster the M detection boxes to obtain C clusters, where each cluster represents the same target detected by multiple first 3D object detection task models from RGB images;

[0016] Calculate the positional accuracy of multiple detection boxes for each cluster, and take the detection box corresponding to the maximum positional accuracy as the detection box of the cluster.

[0017] Furthermore, obtain multiple first 3D object detection task models at different times after training completion; including:

[0018] If the current time sequence number is less than N, then obtain all the first 3D object detection task models before the current time; otherwise, randomly select N first 3D object detection task models from all the first 3D object detection task models before the current time.

[0019] Furthermore, the M detection boxes are clustered to obtain C clusters; including:

[0020] Step S1: Place the M detection boxes into set B;

[0021] Step S2: Select the detection box a with the highest confidence from set B, delete the detection box a from set B, and construct a cluster c based on the detection box a;

[0022] Step S3: For any detection box b in set B, determine whether IoU3D(a,b) is greater than a preset threshold. If yes, put detection box b into cluster c and delete detection box b from set B; where IoU3D(a,b) is the three-dimensional intersection-union ratio of detection box a and detection box b.

[0023] Step S4: Determine if set B is empty. If not, proceed to step S2.

[0024] Furthermore, the positional accuracy of multiple detection boxes for each cluster is calculated; including:

[0025] Calculate the i-th detection box b in cluster c i Position accuracy DIoU(b i ,b m ):

[0026]

[0027] In the formula, b m IoU3D(b) represents the detection box with the highest confidence in cluster C, 1≤i≤K, where K represents the number of detection boxes in cluster C. i ,b m ) represents the detection box b i and detection box b m The 3D intersection-union ratio, ρ represents the detection box b. i Center point and detection box b m The Euclidean distance between the center points of and d represents the Euclidean distance between the detection boxes b. i and detection box b m The diagonal distance of the minimum envelope rectangle.

[0028] Further, the training data is used to train the second 3D object detection task model at the current time, resulting in a completed second 3D object detection task model at the current time; including:

[0029] The RGB image is processed using the second 3D object detection task model at the current moment to obtain prediction results for multiple objects;

[0030] Based on the labeled results of the preset target categories and the prediction results of the corresponding targets in the training data, the ground truth label supervision loss is calculated.

[0031] Based on the target detection results and corresponding target prediction results in the training data, the pseudo-label supervision loss is calculated.

[0032] Calculate total loss

[0033]

[0034] Where λ is a hyperparameter that balances the proportion of the true label supervision loss and the false label supervision loss;

[0035] Utilizing total loss Update the parameters of the second 3D object detection task model at the current moment.

[0036] Thirdly, embodiments of this application provide an efficient training device for a 3D object detection task model that serves lighthouse cognition, the device comprising:

[0037] The acquisition unit is used to acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment;

[0038] The object detection unit is used to process the RGB image using a preset number of first 3D object detection task models trained before the current time step to obtain object detection results;

[0039] The overlay unit is used to overlay the annotation results of the preset target categories and the target detection results of the RGB image to obtain training data;

[0040] The training unit is used to train the second 3D object detection task model at the current moment using the training data, so as to obtain the trained second 3D object detection task model at the current moment; wherein the second 3D object detection task model at the current moment is obtained by copying the first 3D object detection task model at the previous moment that has been trained.

[0041] The update unit is used to update the first 3D object detection task model that was trained at the previous time step using the second 3D object detection task model that has been trained at the current time step, so as to obtain the first 3D object detection task model that has been trained at the current time step.

[0042] Thirdly, embodiments of this application provide an electronic device, including: a memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor executes the computer program to implement the method of embodiments of this application.

[0043] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program, which, when executed by a processor, implements the methods of embodiments of this application.

[0044] The method in this application solves the catastrophic forgetting problem that occurs during model training through continuous learning. Attached Figure Description

[0045] To more clearly illustrate the technical solutions in the specific embodiments of this application or the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0046] Figure 1A schematic diagram illustrating catastrophic forgetting during model training provided in an embodiment of this application;

[0047] Figure 2 A schematic diagram illustrating model training based on continuous learning, provided for an embodiment of this application;

[0048] Figure 3 A flowchart illustrating an efficient training method for a 3D object detection task model for service beacon cognition provided in this application embodiment;

[0049] Figure 4 Functional structure diagram of the efficient training device for the 3D target detection task model of service lighthouse cognition provided in the embodiments of this application;

[0050] Figure 5 A functional structure diagram of an electronic device provided in an embodiment of this application. Detailed Implementation

[0051] To make the objectives, technical solutions, and advantages of the embodiments of this application clearer, the technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. The components of the embodiments of this application described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0052] Therefore, the following detailed description of the embodiments of this application provided in the accompanying drawings is not intended to limit the scope of the claimed application, but merely to illustrate selected embodiments of the application. All other embodiments obtained by those skilled in the art based on the embodiments of this application without inventive effort are within the scope of protection of this application.

[0053] First, a brief introduction to the design concept of the embodiments of this application will be given.

[0054] Roadside sensing devices, also known as roadside beacons, are infrastructure architectures with a "five-knowledge" functional system. These five knowledge functions are: **Knowledge Layer:** The roadside beacon considers static maps of intersections, intersection models, and sensor information. **Cognitive Layer:** The roadside beacon includes modules for dynamic intersection maps and localization, traffic participant detection, and traffic participant intent recognition, enabling it to sense and identify vehicles, pedestrians, and bicycles within the intersection and infer their behavioral intentions. **Predictive Layer:** Modules for traffic scenario prediction, traffic flow anomaly prediction, and traffic violation prediction can predict future traffic conditions, flow anomalies, and potential traffic violations using historical data and real-time sensor information. **Other-Knowledge Layer:** The end-to-end sensing model and sensor calibration modules are responsible for integrating information from different sensors and data sources to construct a comprehensive model for perceiving and understanding the traffic environment. These models are one source of the roadside beacon system's sensing capabilities. **Common-Knowledge Layer:** The adaptive fusion model module fuses and processes information from different functional layers to generate comprehensive sensing information. This is one of the keys to collaborative sensing. Roadside beacon sensing is an important component of the cognitive layer.

[0055] 3D target detection on the roadside is an important task in the field of environmental perception for autonomous driving. Based on roadside images, it can estimate the category of surrounding targets and regress 3D bounding boxes. Its low cost advantage makes it have broad application prospects and commercial value in the fields of autonomous driving and robotics.

[0056] With the development of artificial intelligence, deep neural networks have shown promising results in 3D object detection tasks. However, considering that data captured by roadside cameras in the real world is collected from cameras several meters above the ground, with a large perception range, covering many objects and exhibiting significant distribution differences, the data annotation cost is high, making it difficult to obtain a highly accurate 3D object detection algorithm with comprehensive coverage in one go. A typical setup for roadside 3D object detection involves learning a series of data one after another, including cars, trucks, buses, motorcycles, or pedestrians. This significantly improves accuracy and reduces annotation costs. As roadside 3D object detection technology becomes more widespread, even more data will be added in the future. Therefore, how to continuously acquire, update, accumulate, and utilize roadside data and enhance the adaptability of the algorithm is of great research value.

[0057] Currently, models learn from data from different tasks in a specific order, and then incrementally update the model based on the task execution results. However, when learning new tasks, AI agents are prone to severe forgetting of knowledge learned from old tasks, a phenomenon known as catastrophic forgetting. This explains why models trained sequentially on roadside vehicle data and truck data cannot accurately detect vehicles in the roadside data. Figure 1 As shown.

[0058] The effectiveness of deep neural networks is contingent upon the similarity between the distribution of test and training data; they cannot continuously and adaptively learn in dynamic environments. Therefore, continuous learning has been introduced to address this issue. The significance of continuous learning lies in its ability to efficiently transform and utilize previously learned knowledge to complete new tasks, and it can significantly reduce the problems caused by forgetting, thereby enabling deep neural networks to meet the requirements of on-line 3D object detection, such as... Figure 2 As shown.

[0059] Continuous learning helps algorithms accumulate knowledge from multiple training datasets by comprehensively applying data from both new and old tasks, thereby fully utilizing existing labeled data and continuously improving algorithm capabilities. In recent years, numerous continuous learning methods have been successfully applied to classification and 2D object detection. In the field of continuous learning, a common approach has been to add a storage module to preserve previous data and periodically replay the learned knowledge with cross-referenced data from new samples to mitigate the problem of catastrophic forgetting. However, a common drawback of storage-module-based continuous learning methods is their significant memory requirements.

[0060] To address this, this application proposes an efficient training method for a 3D object detection task model serving lighthouse cognition, comprising: acquiring RGB images of the target scene collected by the roadside perception subject at the current moment; processing the RGB images using a preset number of first 3D object detection task models trained before the current moment to obtain object detection results; superimposing the preset target category annotation results of the RGB images with the object detection results to obtain training data; training the second 3D object detection task model at the current moment using the training data to obtain the trained second 3D object detection task model at the current moment; wherein the second 3D object detection task model at the current moment is obtained by copying the first 3D object detection task model trained at the previous moment; updating the first 3D object detection task model trained at the previous moment using the trained second 3D object detection task model at the current moment to obtain the trained first 3D object detection task model at the current moment. The method of this application can effectively utilize data from both new and old tasks, efficiently transform and utilize previously learned knowledge to complete the learning of new tasks, and can greatly reduce model detection problems caused by forgetting.

[0061] After introducing the application scenarios and design concepts of the embodiments of this application, the technical solutions provided by the embodiments of this application will be described below.

[0062] like Figure 3 As shown, this application provides an efficient training method for a 3D object detection task model that serves lighthouse cognition, including the following steps:

[0063] Step 101: Acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment;

[0064] Preferably, the sensing entity at the roadside is the roadside lighthouse.

[0065] In this embodiment, different tasks are assigned at different sampling times. These tasks refer to target detection tasks, such as truck detection, pedestrian detection, or car detection. Figure 2 As shown, the detection target for task t-1 is a car, the detection target for task t is a truck, and the detection target for task t+1 is a bicycle.

[0066] Step 102: Use the first 3D object detection task model, which has been trained to a predetermined number of times prior to the current moment, to process the RGB image and obtain the object detection result;

[0067] The target detection results include the categories of multiple targets and the position and size of the detection boxes.

[0068] Step 103: Overlay the annotation results of the preset target categories and the target detection results of the RGB images to obtain training data;

[0069] The preset target categories include at least one of the following: cars, trucks, vans, motorcycles, bicycles, and pedestrians. The annotation results of the preset target categories in the RGB image include the position and size of the detection boxes for the preset target categories.

[0070] In this embodiment, the target category is determined by the task. Since the annotation results of the preset target category of the RGB image only include partial target annotation results (the workload of annotating all targets is huge), if only the annotation results of the preset target category of the RGB image are used to train the 3D target detection task model, the resulting model will have poor recognition results for targets that are not in the preset target category. To solve this problem, this application uses the 3D target detection task model obtained at multiple time points to detect RGB images and obtain extended annotation data. The annotation results of the preset target category and the extended annotation data are combined as training data.

[0071] Step 104: Train the second 3D object detection task model at the current moment using the training data to obtain the trained second 3D object detection task model at the current moment; wherein, the second 3D object detection task model at the current moment is obtained by copying the trained first 3D object detection task model at the previous moment.

[0072] In this model, the first 3D object detection task model at each time step is the final model, similar to a teacher model; while the second 3D object detection task model at each time step is an intermediate model, similar to a student model.

[0073] Step 105: Use the second 3D object detection task model trained at the current time step to update the first 3D object detection task model trained at the previous time step, so as to obtain the first 3D object detection task model trained at the current time step.

[0074] As one possible implementation, the RGB image is processed using a first 3D object detection task model that has been trained to a predetermined number of times prior to the current time, to obtain object detection results; including:

[0075] Obtain the first 3D object detection task model at different times after multiple training completions;

[0076] The RGB image is processed using the first 3D object detection task model at each time step after training, resulting in M ​​detection boxes;

[0077] Cluster the M detection boxes to obtain C clusters, where each cluster represents the same target detected by multiple first 3D object detection task models from RGB images;

[0078] Calculate the positional accuracy of multiple detection boxes for each cluster, and take the detection box corresponding to the maximum positional accuracy as the detection box of the cluster.

[0079] As one possible implementation, a first 3D object detection task model is obtained at multiple training completion times; including:

[0080] If the current time sequence number is less than N, then obtain all the first 3D object detection task models before the current time; otherwise, randomly select N first 3D object detection task models from all the first 3D object detection task models before the current time.

[0081] As one possible implementation, the M detection boxes are clustered to obtain C clusters; including:

[0082] Step S1: Place the M detection boxes into set B;

[0083] Step S2: Select the detection box a with the highest confidence from set B, delete the detection box a from set B, and construct a cluster c based on the detection box a;

[0084] Step S3: For any detection box b in set B, determine whether IoU3D(a,b) is greater than a preset threshold. If yes, put detection box b into cluster c and delete detection box b from set B; where IoU3D(a,b) is the three-dimensional intersection-union ratio of detection box a and detection box b.

[0085] Step S4: Determine if set B is empty. If not, proceed to step S2.

[0086] As one possible implementation, the positional accuracy of multiple detection boxes for each cluster is calculated; including:

[0087] Calculate the i-th detection box b in cluster c i Position accuracy DIoU(b i ,b m ):

[0088]

[0089] In the formula, b m IoU3D(b) represents the detection box with the highest confidence in cluster C, 1≤i≤K, where K represents the number of detection boxes in cluster C. i ,b m ) represents the detection box b i and detection box b m The 3D intersection-union ratio, ρ represents the detection box b. i Center point and detection box b m The Euclidean distance between the center points of and d represents the Euclidean distance between the detection boxes b. i and detection box b m The diagonal distance of the minimum envelope rectangle.

[0090] As one possible implementation, the training data is used to train the second 3D object detection task model at the current time, resulting in a trained second 3D object detection task model at the current time; including:

[0091] The RGB image is processed using the second 3D object detection task model at the current moment to obtain prediction results for multiple objects;

[0092] Based on the labeled results of the preset target categories and the prediction results of the corresponding targets in the training data, the ground truth label supervision loss is calculated.

[0093] Based on the target detection results and corresponding target prediction results in the training data, the pseudo-label supervision loss is calculated.

[0094] Calculate total loss

[0095]

[0096] Where λ is a hyperparameter that balances the proportion of the true label supervision loss and the false label supervision loss;

[0097] Utilizing total loss Update the parameters of the second 3D object detection task model at the current moment.

[0098] Based on the above embodiments, this application provides an efficient training device for a 3D object detection task model that serves lighthouse cognition. (See attached document.) Figure 4 As shown in the embodiments of this application, the efficient training device 200 for the 3D object detection task model of service lighthouse cognition at least includes:

[0099] The acquisition unit 201 is used to acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment;

[0100] The target detection unit 202 is used to process the RGB image using a first 3D target detection task model that has been trained to a preset number of times prior to the current time, and to obtain the target detection result;

[0101] The overlay unit 203 is used to overlay the annotation results of the preset target category and the target detection results of the RGB image to obtain training data;

[0102] Training unit 204 is used to train the second 3D object detection task model at the current time using the training data to obtain the trained second 3D object detection task model at the current time; wherein the second 3D object detection task model at the current time is obtained by copying the first 3D object detection task model at the previous time after training.

[0103] The update unit 205 is used to update the first 3D object detection task model that was trained at the previous time using the second 3D object detection task model that has been trained at the current time, so as to obtain the first 3D object detection task model that has been trained at the current time.

[0104] It should be noted that the principle of the efficient training device 200 for the 3D target detection task model of service lighthouse cognition provided in this application embodiment to solve the technical problem is similar to the method provided in this application embodiment. Therefore, the implementation of the efficient training device 200 for the 3D target detection task model of service lighthouse cognition provided in this application embodiment can refer to the implementation of the method provided in this application embodiment, and the repeated parts will not be described again.

[0105] Based on the above embodiments, this application also provides an electronic device, see below. Figure 5 As shown, the electronic device 300 provided in this application embodiment includes at least: a processor 301, a memory 302, and a computer program stored on the memory 302 and executable on the processor 301. When the processor 301 executes the computer program, it implements the efficient training method for the 3D target detection task model of service beacon cognition provided in this application embodiment.

[0106] The electronic device 300 provided in this application embodiment may further include a bus 303 connecting different components (including processor 301 and memory 302). The bus 303 represents one or more types of bus structures, including memory bus, peripheral bus, local area bus, etc.

[0107] The memory 302 may include a readable medium in the form of volatile memory, such as random access memory (RAM) 3021 and / or cache memory 3022, and may further include read-only memory (ROM) 3023.

[0108] The memory 302 may also include a program tool 3025 having a set (at least one) of program modules 3024, including but not limited to: an operating subsystem, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0109] Electronic device 300 can also communicate with one or more external devices 304 (e.g., keyboard, remote control, etc.), and with one or more devices that enable a user to interact with electronic device 300 (e.g., mobile phone, computer, etc.), and / or with any device that enables electronic device 300 to communicate with one or more other electronic devices 300 (e.g., router, modem, etc.). This communication can be performed through input / output (I / O) interface 305. Furthermore, electronic device 300 can also communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) through network adapter 306. Figure 5 As shown, network adapter 306 communicates with other modules of electronic device 300 via bus 303. It should be understood that, although... Figure 5 As not shown, other hardware and / or software modules may be used in conjunction with electronic device 300, including but not limited to: microcode, device drivers, redundant processors, external disk drive arrays, Redundant Arrays of Independent Disks (RAID) subsystems, tape drives, and data backup storage subsystems.

[0110] It should be noted that, Figure 5 The electronic device 300 shown is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of this application.

[0111] This application also provides a computer-readable storage medium storing computer instructions. When executed by a processor, these instructions implement the efficient training method for the 3D target detection task model based on roadside perception provided in this application. Specifically, the executable program can be built into or installed in an electronic device 300, allowing the electronic device 300 to implement the efficient training method for the 3D target detection task model based on roadside perception provided in this application by executing the built-in or installed executable program.

[0112] The efficient training method for the 3D target detection task model of service lighthouse cognition provided in this application embodiment can also be implemented as a program product. The program product includes program code. When the program product can run on the electronic device 300, the program code is used to enable the electronic device 300 to execute the efficient training method for the 3D target detection task model of service lighthouse cognition provided in this application embodiment.

[0113] The program product provided in this application embodiment can be any combination of one or more readable media, wherein the readable media can be a readable signal medium or a readable storage medium, and the readable storage medium can be, but is not limited to, an electrical, magnetic, optical, electromagnetic, infrared, or semiconductor system, device or apparatus, or any combination thereof. Specifically, more specific examples of readable storage media (a non-exhaustive list) include: electrical connections with one or more wires, portable disks, hard disks, RAM, ROM, erasable programmable read-only memory (EPROM), optical fibers, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0114] The program product provided in this application embodiment can be a CD-ROM and include program code, and can also run on a computing device. However, the program product provided in this application embodiment is not limited thereto. In this application embodiment, the readable storage medium can be any tangible medium that contains or stores a program, which can be used by or in conjunction with an instruction execution system, apparatus, or device.

[0115] It should be noted that although several units or sub-units of the device have been mentioned in the detailed description above, this division is merely exemplary and not mandatory. In fact, according to embodiments of this application, the features and functions of two or more units described above can be embodied in one unit. Conversely, the features and functions of one unit described above can be further divided and embodied by multiple units.

[0116] Furthermore, although the operations of the method of this application are described in a specific order in the accompanying drawings, this does not require or imply that these operations must be performed in that specific order, or that all the operations shown must be performed to achieve the desired result. Additionally or alternatively, certain steps may be omitted, multiple steps may be combined into one step, and / or one step may be broken down into multiple steps.

[0117] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application and are not intended to limit it. Although this application has been described in detail with reference to the embodiments, those skilled in the art should understand that modifications or equivalent substitutions to the technical solutions of this application do not depart from the spirit and scope of the technical solutions of this application, and should all be covered within the scope of the claims of this application.

Claims

1. An efficient training method for a 3D object detection task model serving lighthouse cognition, characterized in that, The method includes: Acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment; The first 3D object detection task model, trained to a predetermined number of times prior to the current moment, is used to process the RGB image to obtain the object detection result; The training data is obtained by overlaying the predefined target category annotations and target detection results of the RGB images; The training data is used to train the second 3D object detection task model at the current moment, resulting in the trained second 3D object detection task model at the current moment; wherein the second 3D object detection task model at the current moment is obtained by copying the first 3D object detection task model at the previous moment after training. Using the second 3D object detection task model that has been trained at the current time, the first 3D object detection task model that has been trained at the previous time is updated to obtain the first 3D object detection task model that has been trained at the current time. The first 3D object detection task model, trained to a predetermined number of units prior to the current time, is used to process the RGB image to obtain object detection results, including: Obtain the first 3D object detection task model at different times after multiple training completions; The RGB images are processed using the first 3D object detection task model trained at each time step to obtain... One detection box; right Clustering the detection boxes yields C clusters, where each cluster represents the same target detected by multiple first 3D object detection task models from the RGB image; Calculate the positional accuracy of multiple detection boxes for each cluster, and take the detection box corresponding to the maximum positional accuracy as the detection box of the cluster. Obtain multiple first 3D object detection task models at different time points after training completion; including: If the current time sequence number is less than N, then obtain all the first 3D object detection task models before the current time; otherwise, randomly select N first 3D object detection task models from all the first 3D object detection task models before the current time.

2. The method according to claim 1, characterized in that, The target detection results include the categories of multiple targets and the position and size of the detection boxes.

3. The method according to claim 2, characterized in that, The preset target categories include at least one of the following: cars, trucks, vans, buses, motorcycles, bicycles, and pedestrians; the annotation results of the preset target categories of the RGB images include the position and size of the detection boxes of the preset target categories.

4. The method according to claim 1, characterized in that, right The detection boxes are clustered to obtain C clusters; including: Step S1: Place the M detection boxes into set B; Step S2: Select the detection box with the highest confidence from set B. , detect the box Remove from set B, based on the detection box. Construct a cluster ; Step S3: For any detection box in set B ,judge If the value exceeds a preset threshold, then the detection box is... Add to cluster Remove the detection box from set B. ;in, For detection box and detection box 3D intersection-union ratio; Step S4: Determine if set B is empty. If not, proceed to step S2.

5. The method according to claim 4, characterized in that, Calculate the positional accuracy of multiple bounding boxes for each cluster; include: Computational clustering The Detection box Position accuracy : In the formula, Clustering The detection bounding box with the highest confidence. , Clustering The number of detection frames, Represents the detection box and detection box The three-dimensional intersection-union ratio, Represents the detection box and detection box The Euclidean distance between them Indicates that it contains a detection box. and detection box The diagonal distance of the minimum envelope rectangle.

6. The method according to claim 1, characterized in that, The training data is used to train the second 3D object detection task model at the current time, resulting in a completed second 3D object detection task model at the current time; including: The RGB image is processed using the second 3D object detection task model at the current moment to obtain prediction results for multiple objects; Based on the labeled results of the preset target categories and the prediction results of the corresponding targets in the training data, the ground truth label supervision loss is calculated. ; Based on the target detection results and corresponding target prediction results in the training data, the pseudo-label supervision loss is calculated. ; Calculate total loss : in, A hyperparameter used to balance the proportions of truth label supervision loss and pseudo label supervision loss; Utilizing total loss Update the parameters of the second 3D object detection task model at the current moment.

7. A high-efficiency training device for a 3D object detection task model serving lighthouse cognition, characterized in that, The device includes: The acquisition unit is used to acquire the RGB image of the target scene collected by the roadside sensing subject at the current moment; The object detection unit is used to process the RGB image using a preset number of first 3D object detection task models trained before the current time step to obtain object detection results; The overlay unit is used to overlay the annotation results of the preset target categories and the target detection results of the RGB image to obtain training data; The training unit is used to train the second 3D object detection task model at the current moment using the training data, so as to obtain the trained second 3D object detection task model at the current moment; wherein the second 3D object detection task model at the current moment is obtained by copying the first 3D object detection task model at the previous moment that has been trained. The update unit is used to update the first 3D object detection task model that was trained at the previous time step using the second 3D object detection task model that has been trained at the current time step, so as to obtain the first 3D object detection task model that has been trained at the current time step. The target detection unit is specifically used for: Obtain the first 3D object detection task model at different times after multiple training completions; The RGB images are processed using the first 3D object detection task model trained at each time step to obtain... One detection box; right Clustering the detection boxes yields C clusters, where each cluster represents the same target detected by multiple first 3D object detection task models from the RGB image; Calculate the positional accuracy of multiple detection boxes for each cluster, and take the detection box corresponding to the maximum positional accuracy as the detection box of the cluster. Obtain multiple first 3D object detection task models at different time points after training completion; including: If the current time sequence number is less than N, then obtain all the first 3D object detection task models before the current time; otherwise, randomly select N first 3D object detection task models from all the first 3D object detection task models before the current time.

8. An electronic device, characterized in that, include: A memory, a processor, and a computer program stored in the memory and executable on the processor, wherein the processor, when executing the computer program, implements the method as claimed in any one of claims 1-6.

Images