A post-fusion incremental training method and system for roadside target detection

Abstract

The application provides a roadside target detection post-fusion incremental training method and system, which comprises the following steps: S1: analyzing an existing encapsulation format weight file of an edge device, and initializing a first target detection model based on the file; S2: using the first model to infer an input image to obtain a first detection result set, using a second model to detect the same image to obtain a second detection result set, and performing post-fusion on the two result sets to generate a supervision signal; and S3: using the supervision signal as a target, and training the second model through a loss function combined with a non-maximum suppression post-processing logic, wherein the training comprises: performing first training by using an initial data set, and performing incremental training by using a hierarchical partial unfreezing strategy when a new incremental data set is obtained. The application can directly use the edge-end encapsulation weight file, unify the training and inference logic, and support efficient and continuous optimization of the model on the edge side.

Description

Technical Field

[0001] This invention relates to the field of model training technology, and in particular to a post-fusion incremental training method and system for roadside target detection. Background Technology

[0002] In existing technologies, model training typically relies on a complete initial weight file (usually in .pt format) and a large-scale, fully labeled dataset. This process requires training to begin with a pre-trained model possessing complete structural and parameter information, and is completed on a centralized server or high-performance computing platform. However, as intelligent computing extends to the edge, model inference on edge devices such as roadside mobile edge computing (MEC) units has become a common requirement. For reasons of inference performance optimization and data security, edge devices typically only retain and load weight files that are specifically optimized and packaged for inference (e.g., engine files using lightweight storage formats) when deploying models, rather than retaining the complete network structure and training state information that can be used for retraining.

[0003] When edge devices accumulate large amounts of real-time data (such as images and videos continuously collected by roadside cameras) during operation and generate a need for model updates, existing technical approaches exhibit significant bottlenecks. First, the locally accumulated data must be collected and manually or semi-automatically labeled to form an incremental training set. Then, since the edge device only possesses encapsulated inference weights, which cannot be directly used for initial training, the complete original model file must be retrieved from a remote server and combined with the newly labeled data to perform full or incremental training. This process not only wastes valuable data resources on the edge side (data is not used for model optimization in a timely manner) but also leads to significant update delays and communication overhead due to remote dependencies and the transmission of the complete model.

[0004] Furthermore, the loss function of conventional object detection model training typically relies solely on the direct matching of predicted bounding boxes, failing to effectively integrate the post-processing logic used in the actual inference stage. This disconnect between training and inference leads to a discrepancy between the model's optimization objectives during training and the evaluation criteria for final inference applications, thus affecting the model's performance in real-world scenarios. Meanwhile, edge devices themselves possess considerable available computing resources. In scenarios with continuously generated data, deploying efficient incremental learning mechanisms at the edge, enabling the model to continuously learn and update itself using local data and computing power, will significantly improve the system's adaptability and response efficiency.

[0005] Therefore, the following core problems in the existing technology urgently need to be solved: 1. Edge devices cannot directly use their existing, encapsulated inference weight files for incremental model training, resulting in low data utilization efficiency and delayed updates.

[0006] 2. The training process is disconnected from the post-inference processing logic, which limits the accuracy and robustness of the model in practical applications.

[0007] 3. There is a lack of efficient continuous learning strategies suitable for edge computing environments, which fails to fully utilize edge computing power to achieve timely model iteration. Summary of the Invention

[0008] The technical problem to be solved by this invention is: In view of the technical problems existing in the prior art, this invention provides a post-fusion incremental training method and system for roadside target detection, which aims to effectively utilize the weight file encapsulated at the edge and integrate the inference post-processing strategy into the training process, so as to achieve efficient and accurate incremental model updates on edge devices, and overcome the problems of dependence on complete initial weights, inconsistency between training and inference, and low update efficiency of the prior art.

[0009] To solve the above-mentioned technical problems, the technical solution proposed by this invention is as follows: A post-fusion incremental training method for roadside target detection includes the following steps: Step S1: Parse the existing encapsulation format weight file of the edge device, and initialize the first target detection model based on the weight file; Step S2: Use the first object detection model to infer the input image to obtain a first detection result set; use the second object detection model to detect the same input image to obtain a second detection result set; fuse the first detection result set and the second detection result set to generate a fused detection result, and use the fused detection result as a supervision signal for training the second object detection model; Step S3: Using the supervision signal as the target, train the second target detection model using a loss function that incorporates non-maximum suppression post-processing logic; the training steps of the second target detection model include: Initial training phase: Using the initial dataset, train the parameters of the second object detection model; Incremental training phase: When a new incremental dataset is obtained, a hierarchical partial unfreeze strategy is adopted, which updates only some network layer parameters of the second object detection model while keeping the remaining parameters frozen.

[0010] As a further improvement to the method of the present invention: step S1 includes: Step S101: Read the .wts weight file stored at the edge computing terminal and obtain the network structure configuration file corresponding to the weight file; Step S102: Using the .wts file and the corresponding network structure configuration file as input, call the toolchain to perform model conversion and optimization, and generate a serialized .engine optimized model file with FP16 accuracy; Step S103: Based on the .engine optimized model file, construct the inference engine context of the toolchain and complete the runtime environment initialization before inference; Step S104: Input the preprocessed image data into the inference engine context, perform forward computation to complete target detection, and parse the network output to obtain a first detection result set containing bounding box coordinates, category labels, and confidence scores.

[0011] As a further improvement to the method of the present invention: In step S2, the method of post-fusion of the first detection result set and the second detection result set is to calculate the intersection-union ratio (IOU) of the bounding boxes in the first detection result set and the second detection result set, and to perform box matching and fusion based on a preset IOU threshold.

[0012] As a further improvement to the method of the present invention: In step S3, the loss function that integrates non-maximum suppression post-processing logic includes one or more of the following: basic bounding box loss term, NMS constraint loss term, and redundancy penalty loss term; the basic bounding box loss term is used to calculate the deviation between the predicted box and the true box; the NMS constraint loss term is used to evaluate the quality of the predicted boxes retained after non-maximum suppression processing, and to impose a loss penalty on the predicted boxes whose evaluation results do not meet the preset quality standard; the redundancy penalty loss term is used to penalize the predicted boxes that highly overlap with the high-scoring predicted boxes, so as to suppress the tendency of the model to generate multiple duplicate detection results for the same target.

[0013] As a further improvement to the method of the present invention: the total loss function is the weighted sum of the basic bounding box loss term, the NMS constraint loss term and the redundancy penalty loss term, and the weight coefficients corresponding to each loss term are configurable parameters.

[0014] As a further improvement to the method of the present invention: in step S3, the initial training phase adopts a mixed precision training strategy, the forward computation of the model adopts FP16 precision to reduce memory usage, and the gradient update adopts FP32 precision to ensure numerical stability; after training, the model parameters with FP32 precision are saved.

[0015] As a further improvement to the method of the present invention: in step S3, the incremental training phase specifically includes: Unfreeze the detection head network layer of the second target detection model, while keeping the parameters of the backbone network layer of the second target detection model frozen; A dynamic learning rate scheduling strategy is adopted to fine-tune the parameters of the second target detection model. The initial value of the dynamic learning rate is set as a specified proportion of the final learning rate in the first training phase, and it decays according to the cosine annealing law. The parameters of the unfrozen detection head network layer of the second object detection model are updated using a new incremental dataset and a post-fusion loss function.

[0016] As a further improvement to the method of the present invention, the method further includes a model deployment step: converting the model weights obtained after the incremental training phase into FP16 precision and saving them as .pt format files for inference deployment on edge computing devices.

[0017] The present invention also provides a post-fusion incremental training system for roadside target detection, comprising a microprocessor and a memory interconnected thereto, the microprocessor being programmed or configured to execute the post-fusion incremental training method for roadside target detection.

[0018] The present invention also provides a computer-readable storage medium storing a computer program / instructions programmed or configured to execute the post-fusion incremental training method for roadside target detection via a processor.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows: This invention initializes the training process by directly parsing the encapsulated weight files (e.g., .wts format) locally on the edge device, constructing a discrete training system that does not rely on a complete cloud model. This significantly reduces communication overhead and update latency while effectively maintaining the confidentiality and lightweight nature of the model weights. By integrating post-processing logic (e.g., NMS) into the training loss function, it innovatively aligns the training objective with the inference process, significantly improving the model's detection accuracy and generalization ability in real-world deployment scenarios. By constructing a new convolutional neural network and employing a layered, partially unfrozen incremental training mechanism, it fully leverages local edge data and computing resources to achieve continuous autonomous model optimization while avoiding catastrophic forgetting. Finally, through resource optimization techniques such as mixed-precision training and FP16 inference storage, a complete and efficient training scheme adapted to the stringent constraints of edge computing environments is formed, providing a reliable technical path for the continuous evolution of edge intelligent applications. Attached Figure Description

[0020] Figure 1 This is a flowchart of the post-fusion incremental training method for roadside target detection in an embodiment of the present invention. Detailed Implementation

[0021] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0022] like Figure 1 As shown, this embodiment provides a post-fusion incremental training method for roadside target detection, including the following steps: Step S1: Parse the existing encapsulation format weight file of the edge device and initialize the first target detection model based on the weight file.

[0023] In this embodiment, parsing the existing encapsulation format weight file of the edge device and initializing the first target detection model based on the weight file specifically includes: Step S101: Read the .wts weight file stored at the edge computing terminal and obtain the network structure configuration file corresponding to the weight file; Step S102: Using the .wts weight file and the corresponding network structure configuration file as input, call the toolchain to perform model conversion and optimization, and generate a serialized .engine optimized model file with FP16 accuracy; Step S103: Optimize the model file based on the .engine file, build the inference engine context of the toolchain, and complete the runtime environment initialization before inference; Step S104: Input the preprocessed image data into the inference engine context, perform forward computation to complete object detection, and parse the network output to obtain the first detection result set containing bounding box coordinates, class labels and confidence scores.

[0024] This embodiment directly reads the .wts weight file locally on the edge device and uses an edge computing toolchain (e.g., using NVIDIA TensorRT as an optimization example) to convert it into an FP16 precision .engine optimized model. This successfully breaks the dependence of model training on complete initial weights, laying the foundation for autonomous edge updates. Based on this optimized model, an inference engine context is built, which can quickly complete object detection inference for images and output a structured result set A containing bounding boxes, categories, and confidence scores. This process fully utilizes the efficiency of the edge inference engine, significantly reducing memory usage and computational latency while ensuring the generation of high-quality supervision signals, perfectly adapting to the resource-constrained environment of the edge.

[0025] Step S2: Use the first object detection model to infer the input image to obtain the first detection result set; use the second object detection model to detect the same input image to obtain the second detection result set; fuse the first detection result set and the second detection result set to generate the fused detection result, and use the fused detection result as the supervision signal for training the second object detection model.

[0026] In this embodiment, the method for post-fusion of the first detection result set and the second detection result set is to calculate the intersection-union ratio (IOU) of the bounding boxes in the first detection result set and the second detection result set, and then perform box matching and fusion based on a preset IOU threshold.

[0027] In this embodiment, the Intersection over Union (IOU) is defined as the ratio of the intersection area of ​​two bounding boxes to their union area, i.e.: IOU = (intersection area of ​​predicted box and ground truth box) / (union area of ​​predicted box and ground truth box). During the post-fusion process, the bounding boxes in the first detection result set (A) are used as the matching benchmark, and the bounding boxes in the second detection result set (B) are used as the objects to be matched. The IoU value between the two is calculated and compared with a preset threshold to complete the box matching and fusion operation.

[0028] In a specific application embodiment, an efficient "edge self-supervised" training mechanism is achieved by initializing a new model (i.e., the second object detection model) and performing post-fusion with the baseline model. Specifically, this method calculates the Intersection over Union (IOU) between the output (B) of the new model and the reliable baseline (A) and performs intelligent fusion based on a threshold, dynamically transforming the static, encapsulated inference weight file into a powerful, iterative "signal generator." This process fundamentally solves the core pain point of the edge side being unable to perform continuous learning due to the lack of high-quality labeled data and complete training models. It enables continuous incremental optimization of the model using local edge data without relying on cloud support, greatly improving the model's adaptability in real-world scenarios and the overall intelligent and autonomous level of the system.

[0029] Step S3: Using the supervision signal as the target, train the second object detection model through a loss function that incorporates non-maximum suppression post-processing logic; the model training steps include: Initial training phase: Using the initial dataset, train the parameters of the second object detection model; Incremental training phase: When a new incremental dataset is obtained, a hierarchical partial unfreeze strategy is adopted, which updates only some network layer parameters of the second object detection model while keeping the remaining parameters frozen.

[0030] In this embodiment, the loss function that integrates non-maximum suppression post-processing logic includes one or more of the following: basic bounding box loss term, NMS constraint loss term, and redundancy penalty loss term. The basic bounding box loss term is used to calculate the deviation between the predicted box and the ground truth box. The NMS constraint loss term is used to evaluate the quality of the predicted boxes retained after non-maximum suppression processing and to impose a loss penalty on the predicted boxes whose evaluation results do not meet the preset quality standard. The redundancy penalty loss term is used to penalize predicted boxes that highly overlap with high-scoring predicted boxes to suppress the tendency of the model to generate multiple duplicate detection results for the same target.

[0031] In this embodiment, the total loss function is a weighted sum of the basic bounding box loss term, the NMS constraint loss term, and the redundancy penalty loss term, and the weight coefficients corresponding to each loss term are configurable parameters.

[0032] Specifically, the core of the loss function for the fusion inference post-processing logic lies in simulating and optimizing the processing effect of non-maximum suppression (NMS), which includes three synergistic loss terms: Positioning accuracy guarantee: CIoU (Complete-IoU) loss is used as the basic bounding box loss (L_box) to calculate the deviation between the predicted box and the ground truth box, ensuring the basic positioning accuracy of the predicted box.

[0033] Output quality optimization term: Introduce NMS constraint loss (L_nms) to review the IoU quality of the prediction boxes retained after NMS filtering, and penalize the retained results with low quality (such as IoU below 0.7), thereby improving the quality of the final output set.

[0034] Output simplicity optimization term: Introduce redundancy penalty loss (L_redundant) to penalize predicted boxes that are judged as redundant in NMS (such as those with an IoU greater than 0.5 with the reserved box), thereby driving the model to reduce repeated predictions from the source and making the output simpler.

[0035] The total loss function in this embodiment is a dynamic weighted sum of the above three factors: Loss = α×L_box + β×L_nms + γ×L_redundant (1) α, β, and γ are weight coefficients that can be dynamically adjusted according to the training phase to adapt to different optimization objectives.

[0036] In this embodiment, a mixed-precision training strategy is adopted in the initial training phase. FP16 precision is used for forward model computation to significantly reduce memory usage by approximately 50%, enabling smooth model training on resource-constrained edge devices. Simultaneously, FP32 precision is used for gradient updates to effectively ensure numerical stability during training, avoiding gradient vanishing or convergence oscillations caused by precision truncation, thus ensuring the reliability of model optimization. The entire training process is driven by a loss function incorporating NMS post-processing logic, guiding the model to output results more closely aligned with actual inference scenarios from the outset. After training, the complete FP32 precision model is saved, preserving the rich feature information carried by the full-precision weights and providing a high-precision, highly stable starting point for subsequent incremental learning phases.

[0037] In this embodiment, the incremental training phase specifically includes: Unfreeze the detection head network layer of the second target detection model, while keeping the parameters of the backbone network layer of the second target detection model frozen; A dynamic learning rate scheduling strategy is adopted to achieve stable fine-tuning of the parameters of the second target detection model. The initial value of the dynamic learning rate is set as a specified proportion of the final learning rate in the first training phase, and it decays according to the cosine annealing law. The parameters of the unfrozen detection head network layer of the second object detection model are updated using a new incremental dataset and a post-fusion loss function.

[0038] The method in this embodiment also includes a model deployment step: converting the model weights obtained after the incremental training phase into FP16 precision and saving them as .pt format files for inference deployment on edge computing devices.

[0039] In summary, this embodiment initializes the model by directly parsing the existing .wts weight files on the edge device, eliminating the need to obtain the complete .pt model from the central server, thus ensuring the lightweight nature and confidentiality of the weight files from the source. At the training optimization level, by introducing a post-fusion loss function that integrates non-maximum suppression (NMS) and intersection-over-union (IoU) constraints, the inference logic is embedded into the training objective, effectively bridging the gap between training and inference, thereby significantly improving the model's detection accuracy and robustness. To achieve continuous learning, this invention constructs a new convolutional neural network and adopts a hierarchical partial unfreezing strategy, supporting incremental optimization of the model using only new data from the edge. This maximizes the use of local data resources and fully utilizes the idle computing power of edge devices. Furthermore, the mixed precision strategy and FP16 model storage used in the training and deployment phases significantly reduce memory usage and inference latency while ensuring model performance, making this solution particularly suitable for resource-constrained edge environments.

[0040] This embodiment also provides a post-fusion incremental training system for roadside target detection, including a microprocessor and a memory interconnected, wherein the microprocessor is programmed or configured to perform a post-fusion incremental training method for roadside target detection.

[0041] In this embodiment, the system includes: Weight parsing module: Deployed on edge devices, used to parse .wts files and generate inference models; Construct a new convolutional neural network to achieve result set fusion, and use the fused result set as a supervision set; Loss function construction module: used to implement the calculation logic of the post-fusion loss function; Training module: Includes a mixed-precision training engine and an incremental update controller, used to adapt to the computing power of edge devices; Model saving module: Used to convert the model to FP16 precision and output a .pt file.

[0042] This embodiment also provides a computer-readable storage medium storing a computer program / instructions that are programmed or configured to execute a post-fusion incremental training method for roadside target detection via a processor.

[0043] Those skilled in the art will understand that the above embodiments of this application can be provided as methods, systems, or computer program products. Therefore, this application can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, this application can take the form of a computer program product embodied on one or more computer-readable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code. This application is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of this application. It should be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, create a machine for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1One or more processes and / or boxes Figure 1 The functions specified in one or more boxes. These computer program instructions may also be loaded onto a computer or other programmable data processing apparatus to cause a series of operational steps to be performed on the computer or other programmable apparatus to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable apparatus for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the functions specified in one or more boxes. The above are merely preferred embodiments of the present invention and are not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

[0044] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the invention. Therefore, any simple modifications, equivalent changes, and alterations made to the above embodiments based on the technical essence of the present invention without departing from the scope of the present invention should fall within the protection scope of the present invention.

Claims

1. A post-fusion incremental training method for roadside target detection, characterized in that, Includes the following steps: Step S1: Parse the existing encapsulation format weight file of the edge device, and initialize the first target detection model based on the weight file; Step S2: Use the first target detection model to perform inference on the input image to obtain the first detection result set; The second object detection model is used to detect the same input image to obtain a second detection result set; The first detection result set and the second detection result set are then fused to obtain the fused detection result, and the fused detection result is used as the supervision signal for training the second target detection model. Step S3: Using the supervision signal as the target, train the second target detection model by using a loss function that incorporates non-maximum suppression post-processing logic; The training steps for the second object detection model include: Initial training phase: Using the initial dataset, train the parameters of the second object detection model; Incremental training phase: When a new incremental dataset is obtained, a hierarchical partial unfreeze strategy is adopted, which updates only some network layer parameters of the second object detection model while keeping the remaining parameters frozen; Step S1 includes: Step S101: Read the .wts weight file stored at the edge computing terminal and obtain the network structure configuration file corresponding to the .wts weight file; Step S102: Using the .wts weight file and the corresponding network structure configuration file as input, call the toolchain to perform model conversion and optimization, and generate a serialized .engine optimized model file with FP16 accuracy; Step S103: Based on the .engine optimized model file, construct the inference engine context of the toolchain and complete the runtime environment initialization before inference; Step S104: Input the preprocessed image data into the inference engine context, perform forward computation to complete target detection, and parse the network output to obtain a first detection result set containing bounding box coordinates, category labels and confidence scores; In step S3, the loss function that incorporates the non-maximum suppression post-processing logic includes: The total loss function is a weighted sum of the basic bounding box loss term, the NMS constraint loss term, and the redundancy penalty loss term, and the weight coefficients corresponding to each loss term are configurable parameters. The basic bounding box loss term is used to calculate the deviation between the predicted box and the ground truth box. The NMS constraint loss term is used to evaluate the quality of the predicted boxes retained after non-maximum suppression processing, and to impose a loss penalty on the predicted boxes whose evaluation results do not meet the preset quality standards. The redundancy penalty loss term is used to penalize predicted boxes that highly overlap with high-scoring predicted boxes, so as to suppress the tendency of the model to generate multiple duplicate detection results for the same target.

2. The post-fusion incremental training method for roadside target detection according to claim 1, characterized in that, In step S2, the method for post-fusion of the first detection result set and the second detection result set is to calculate the intersection-union ratio (IOU) of the bounding boxes in the first detection result set and the second detection result set, and to perform box matching and fusion based on a preset IOU threshold.

3. The post-fusion incremental training method for roadside target detection according to claim 1, characterized in that, In step S3, the initial training phase adopts a mixed precision training strategy. The forward computation of the model uses FP16 precision to reduce memory usage, and the gradient update uses FP32 precision to ensure numerical stability. Save the model parameters with FP32 accuracy after training is complete.

4. The post-fusion incremental training method for roadside target detection according to claim 1, characterized in that, In step S3, the incremental training phase specifically includes: Unfreeze the detection head network layer of the second target detection model, while keeping the parameters of the backbone network layer of the second target detection model frozen; A dynamic learning rate scheduling strategy is adopted to fine-tune the parameters of the second target detection model. The initial value of the dynamic learning rate is set as a specified proportion of the final learning rate in the first training phase, and it decays according to the cosine annealing law. The parameters of the unfrozen detection head network layer of the second object detection model are updated using a new incremental dataset and a post-fusion loss function.

5. The post-fusion incremental training method for roadside target detection according to claim 1, characterized in that, The method also includes a model deployment step: converting the model weights obtained after the incremental training phase into FP16 precision and saving them as .pt format files for inference deployment on edge computing devices.

6. A post-fusion incremental training system for roadside target detection, comprising a microprocessor and a memory interconnected, characterized in that, The microprocessor is programmed or configured to execute the post-fusion incremental training method for roadside target detection as described in any one of claims 1 to 5.

7. A computer-readable storage medium storing a computer program / instructions, characterized in that, The computer program / instructions are programmed or configured to execute the post-fusion incremental training method for roadside target detection as described in any one of claims 1 to 5 via a processor.

Images