Construction site small-scale target detection method, system and equipment based on improved YOLO and medium

Abstract

The invention discloses a construction site small-scale target detection method, system and equipment based on improved YOLO, and a medium, and the method comprises the steps: constructing a construction site target detection data set SODA containing multiple types of construction targets, and dividing the SODA into a training set, a verification set and a test set; the YOLO is used as a reference model, a multi-path multi-scale cavity convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF) are integrated, and a construction site small-scale target detection model based on the improved YOLO is constructed; training, verifying and testing are carried out to finally obtain an efficient small-scale target detection model adaptive to the construction site edge computing equipment deployment requirements; the system, the equipment and the medium are used for implementing the method. The method has the technical effects that the detection frame is more accurate, the small target detection precision is higher, the calculation complexity is low, and the cross-scene generalization ability is high.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of computer vision and construction engineering monitoring, and particularly relates to a construction site small-scale target detection method, system, device and medium based on an improved YOLO. BACKGROUND

[0002] Current construction site target detection mainly relies on a remote monitoring system composed of high-altitude cameras and fixed monitoring equipment, which has improved scene coverage and data collection efficiency, but the backend data processing still has significant defects: artificial interpretation is highly subjective and has poor real-time performance, and existing detection algorithms have insufficient detection accuracy for small-scale targets in complex construction environments, which are prone to missed detection and misjudgment. Although the target detection algorithm based on the YOLO series has shown good performance in the industrial detection field, it still faces multiple challenges in small-scale target detection in the construction site: first, the construction site targets generally exhibit small-scale characteristics due to the high-altitude collection angle, and traditional convolution downsampling leads to excessive aggregation of small target features, weak expression ability of deep features, and difficulty for the detector to effectively distinguish; second, there is a significant spatial resolution and semantic difference between different scale feature maps when they are fused, and the feature misalignment problem causes small-scale target features to be easily swallowed by large target features, seriously affecting the detection reliability; third, existing methods focus on single-type targets or large-scale objects, and have insufficient detection capability for multiple small-scale targets such as workers, tools, and safety signs, which cannot provide comprehensive resource distribution information for construction management. These problems restrict the advancement of intelligent management in the construction site, and there is an urgent need to design an efficient detection framework that takes into account small target feature expansion and cross-scale feature alignment.

[0003] The patent application file with publication number CN119785184A discloses a lightweight real-time target detection method based on YOLO-FESD, which improves target detection efficiency by adding a C2f_Faster_ELA module group in the backbone network of YOLOv8 network, adding a GSConv layer and VoV-GSCSP module in the neck network, and adding a deep separable convolution layer in the head network, but this scheme has the problem of insufficient extraction of subtle features due to the easy aggregation and loss of small-scale target feature information in complex scenes and cross-scale feature misalignment, and the traditional YOLOv8 has insufficient optimization for small target detection and high computational complexity, resulting in low detection accuracy in complex scenes and affected model real-time performance.

[0004] The patent application file with publication number CN119810801A discloses an unmanned aerial vehicle runway obstacle identification method and obstacle avoidance method based on an improved YOLOv5, which improves the unmanned aerial vehicle runway obstacle identification efficiency by replacing the target identification sub-model in the backbone network Backbone with CenterNet, deleting the MSN component in the detection head Head, and applying a sky segmentation method, but this scheme still has the problem of recognition accuracy caused by the insufficient feature extraction of CenterNet for dense small targets, and the limitations of the improved YOLOv5 visual model with low sensitivity to obstacle edge features in complex environments and image information loss affecting the detection effect after sky segmentation. SUMMARY

[0005] In order to overcome the shortcomings of the prior art in small-scale target detection on construction sites, the purpose of the present application is to provide a small-scale target detection method, system, device and medium based on improved YOLO for construction sites, by constructing a data set SODA containing multiple types of construction targets, using YOLO as a benchmark model and integrating a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF) to optimize the model structure, which can enhance the small-scale target feature expression ability, alleviate the cross-scale feature misplacement problem, and adapt to complex construction environments and improve detection real-time performance, with more accurate detection frame, higher small target detection accuracy, low computational complexity, strong cross-scene generalization ability and other technical effects.

[0006] In order to achieve the above purpose, the technical solution adopted by the present application is:

[0007] A small-scale target detection method based on improved YOLO for construction sites, comprising the following steps:

[0008] S1, a construction site target detection data set SODA containing multiple types of construction targets is constructed, and is divided into a training set, a validation set and a test set;

[0009] S2, using YOLO as a benchmark model, integrating a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF), constructing a small-scale target detection model for construction sites based on improved YOLO;

[0010] S3, training the improved YOLO-based construction site small-scale target detection model constructed in step S2 using the training set divided in step S1, obtaining the trained improved YOLO-based construction site small-scale target detection model through training set optimization; verifying the trained improved YOLO-based construction site small-scale target detection model using the verification set divided in step S1, obtaining the optimized generalization improved YOLO-based construction site small-scale target detection model through verification set monitoring and early stopping strategy screening; inputting the test set divided in step S1 into the optimized generalization improved YOLO-based construction site small-scale target detection model for testing, finally obtaining the efficient small-scale target detection model adapted to the deployment requirements of the construction site edge computing device through test set evaluation.

[0011] The specific method of step S1 is:

[0012] By screening the disclosed construction site image data, a construction site target detection data set SODA containing multiple types of construction targets is constructed, covering high-altitude aerial view, close-range operation multi-view, and multiple lighting conditions such as sunny, cloudy, and night, and the construction site data set SODA is divided into a training set, a verification set, and a test set.

[0013] The specific method of step S2 is:

[0014] YOLO is used as a benchmark model, including an input layer, a backbone network, a neck network, and a prediction head, a multi-path multi-scale hollow convolution feature extraction module (MPFD) is embedded in the backbone network, and an adaptive multi-level feature fusion module (AMFF) is introduced in the neck network, to construct an improved YOLO-based construction site small-scale target detection model;

[0015] The multi-path multi-scale hollow convolution feature extraction module (MPFD) is used to receive input features, first performs a convolution operation to convert the number of convolution channels from c to c / 2, uses a convolution kernel size of 3, then the features are divided into three parallel paths:

[0016] The first path performs a convolution operation to convert the number of channels from c / 2 to c, using a convolution kernel size of 3 and a dilation rate of 1; the second path performs a convolution operation to convert the number of channels from c / 2 to c, using a convolution kernel size of 3 and a dilation rate of 3; the third path performs a convolution operation to convert the number of channels from c / 2 to c, using a convolution kernel size of 3 and a dilation rate of 5;

[0017] The results of the three parallel paths are spliced in features, the channel is adjusted through a 3x3 convolution layer, then a convolution operation is performed to convert the convolution channel number from 2c after splicing to c, a convolution kernel with a size of 1 is used, then the result is added to the initial input feature, and finally an MPFD multi-scale feature map is output;

[0018] The adaptive multi-level feature fusion module (AMFF) takes the MPFD multi-scale feature map as input, first adjusts the channel number through 1x1 convolution, then enters the CBS module composed of 1x1 convolution, BN layer and SILU activation function for processing, and generates a weight matrix using the Sigmoid function; based on the weight matrix, the forward weighted feature and the reverse weighted feature are calculated respectively, and the weighted features are scaled matched through bilinear interpolation, so that the features of different scales remain consistent in size; then, multi-level pre-fusion operation is performed, the original features of the current scale and the fused features after scale matching are fused again, and finally, the features of different scales are spliced together through cross-scale splicing to output the fused and enhanced features.

[0019] The specific method of step S3 is:

[0020] The training set divided in step S1 is used to optimize the network parameters of the improved YOLO-based construction site small-scale target detection model, and after a certain number of training, a trained improved YOLO-based construction site small-scale target detection model is obtained.

[0021] The verification set divided in step S1 is input into the trained improved YOLO-based construction site small-scale target detection model for verification, the overfitting risk of the improved YOLO-based construction site small-scale target detection model during training is monitored, and the hyperparameter learning rate and weight decay coefficient are dynamically adjusted. When the performance indicators of the improved YOLO-based construction site small-scale target detection model on the verification set no longer improve in continuous multiple training rounds, the convergence condition is met, the early stopping strategy is triggered, the generalization ability of the improved YOLO-based construction site small-scale target detection model in complex construction environment is ensured, the balance between small target feature preservation and large target feature fusion is balanced, and an optimized generalization improved YOLO-based construction site small-scale target detection model is obtained.

[0022] The test set divided in step S1 is input into the optimized generalization improved YOLO-based construction site small-scale target detection model for testing, the improved YOLO-based construction site small-scale target detection model generates a prediction result, the detection performance of the improved YOLO-based construction site small-scale target detection model on unseen data is evaluated through quantitative indicators, and finally an efficient small-scale target detection model that meets the deployment requirements of the construction site edge computing device is obtained.

[0023] The application also provides a construction site small-scale target detection system based on improved YOLO, comprising:

[0024] A data set construction and division module is configured to construct a construction site target detection data set SODA containing multiple types of construction targets and divide the data set into a training set, a verification set and a test set;

[0025] A construction site small-scale target detection model construction module based on improved YOLO is configured to implement a construction site small-scale target detection model based on improved YOLO by taking YOLO as a reference model, integrating a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF).

[0026] A construction site small-scale target detection model training, verification and test module based on improved YOLO is configured to train the construction site small-scale target detection model based on improved YOLO using the training set, optimize the model based on the training set, obtain a trained construction site small-scale target detection model based on improved YOLO, verify the trained construction site small-scale target detection model based on improved YOLO using the verification set, screen the model based on the verification set monitoring and early stopping strategy, obtain an optimized generalization construction site small-scale target detection model based on improved YOLO, input the test set into the optimized generalization construction site small-scale target detection model based on improved YOLO, test the model based on the test set, evaluate the model based on the test set, and finally obtain an efficient small-scale target detection model that meets the deployment requirements of a construction site edge computing device.

[0027] The application also provides a construction site small-scale target detection device based on improved YOLO, comprising:

[0028] A memory is configured to store a computer program of the construction site small-scale target detection method based on improved YOLO, and the memory is a computer-readable device.

[0029] A processor is configured to implement the construction site small-scale target detection method based on improved YOLO when the computer program is executed.

[0030] The application also provides a computer-readable storage medium, which stores a computer program, and the computer program can implement the construction site small-scale target detection method based on improved YOLO when executed by a processor.

[0031] Compared with the prior art, the application has the following beneficial effects:

[0032] 1. The application embeds an MPFD module in the backbone network, uses three parallel dilated convolution with expansion rates of 1, 3 and 5 combined with splicing, reduces the calculation amount in the compression channel, extracts small target accurate position information and large scale feature expansion respectively, significantly alleviates the feature over-aggregation caused by downsampling, expands the small target feature information distribution, strengthens the small target detail retention from the bottom feature extraction stage, and avoids the problems of feature aggregation loss and insufficient fine feature extraction in the prior art.

[0033] 2. The application introduces an adaptive multi-level feature fusion module (AMFF) in the neck network, adjusts the channel through 1x1 convolution, extracts features through CBS (1x1 convolution layer, BN layer normalization and SILU activation function), generates adaptive weighting mechanism of self-learning weight through Sigmoid, combines three-party pre-fusion strategy of bilinear interpolation scale matching and self-scale feature and forward and backward weighted features, eliminates the information imbalance and misplacement problem of cross-scale features, avoids the small target feature being swallowed by large target; meanwhile, the dataset SODA containing 15 types of construction targets, multi-view and multi-illumination conditions is integrated, the training set, the verification set and the test set are divided in the proportion of 7:1:2, the complex construction scene is covered, the limitations of dense small target missed detection, low edge feature sensitivity and poor scene adaptability in the prior art are solved, and the detection robustness of the model in multiple scenes is significantly improved.

[0034] 3. The application strengthens the small scale target feature expansion capability through the multi-path parallel dilated convolution mechanism of the multi-path multi-scale dilated convolution feature extraction module (MPFD), improves the feature alignment accuracy by combining the self-learning weight and cross-scale pre-fusion strategy of the adaptive multi-level feature fusion module (AMFF), and realizes lightweight optimization based on the YOLO benchmark model, which significantly improves the small scale target detection accuracy, feature fusion efficiency and inference real-time performance, and effectively solves the missed detection problem caused by small target feature aggregation loss and feature misplacement in the prior art.

[0035] 4. The multi-path multi-scale dilated convolution feature extraction module (MPFD) of the application adopts multi-expansion rate parallel dilated convolution design with expansion rates of 1, 3 and 5 respectively, which significantly alleviates the feature over-aggregation in the downsampling process and retains the small target detail information; the adaptive multi-level feature fusion module (AMFF) eliminates the information imbalance between different scale features through adaptive weighting and multi-level pre-fusion, and improves the saliency of small target features in the fusion process. Combined with the multi-scene dataset SODA, WOTR and Pictor-v3 covering multi-view and multi-resolution scenes in construction site, the model generalization ability is enhanced, the small target missed detection and misjudgment in complex construction environment are reduced, the calculation complexity is controlled, and the real-time detection demand is met.

[0036] 5、The application designs a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF) and integrates them in a YOLO benchmark model to construct a small-scale target detection framework suitable for construction sites, effectively expands the small target feature distribution and solves the feature misplacement problem, and improves the model's perception ability for small-scale targets such as workers and tools. The channel compression design of the multi-path multi-scale hollow convolution feature extraction module (MPFD) reduces the calculation cost while enhancing the feature expression, and the lightweight fusion structure of the adaptive multi-level feature fusion module (AMFF) avoids the swallowing of small target features by large-scale targets, so that the model still maintains high detection accuracy under complex lighting and high-altitude angles, and adapts to the deployment of edge computing devices in construction sites.

[0037] 6、The multi-path multi-scale hollow convolution feature extraction module (MPFD) of the application expands medium and large scale features while extracting small target position information through multi-path hollow convolution parallel processing, avoids the hollow effect while reducing feature redundancy, maintains network lightweight, and improves the model's ability to capture small target edges and details.

[0038] 7、The adaptive multi-level feature fusion module (AMFF) of the application dynamically adjusts the weights of features of different scales through four times of weighting processing and three times of cross-scale fusion, solves the misplacement problem of traditional fusion methods, and makes small target features remain complete in multi-scale fusion. Experiments show that the introduction of AMFF alone improves ARS by 9.3%, and further improves it to 23.7% after combining with MPFD, which significantly enhances the detection robustness of the model for dense small targets.

[0039] 8、The small-scale target detection model based on improved YOLO constructed by the application cooperates with the multi-path multi-scale hollow convolution feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), and significantly outperforms existing methods in key indicators such as APS and ARS of small-scale target detection, and maintains leading performance in cross-scene generalization test, proving its strong adaptability to complex construction environments. Compared with single-module optimization, the double-module cooperation significantly improves the feature utilization efficiency, reduces the calculation overhead, and realizes the balance between detection accuracy and real-time performance.

[0040] In summary, by constructing the MPFD and the adaptive multi-level feature fusion module (AMFF) and combining with multi-scene dataset training, the application realizes the accurate detection of small-scale targets in construction sites, has the significant advantages of more accurate detection frame positioning, higher small target detection accuracy, strong cross-scene generalization ability, low calculation complexity and adaptive edge device deployment. BRIEF DESCRIPTION OF DRAWINGS

[0041] Figure 1 The implementation method flowchart of the embodiment of the application.

[0042] Figure 2A structural schematic diagram of a multi-path multi-scale hollow convolution feature extraction module (MPFD) of the present application.

[0043] Figure 3 A structural schematic diagram of an adaptive multi-level feature fusion module (AMFF) of the present application.

[0044] Figure 4 A structural schematic diagram of a construction site small-scale target detection model based on an improved YOLO.

[0045] Figure 5 A thermal map of the ablation experiment.

[0046] Figure 6 A detection result map of the ablation experiment.

[0047] Figure 7 A thermal map of the performance comparison experiment.

[0048] Figure 8 A detection frame map of the performance comparison experiment.

[0049] Figure 9 A thermal map of the generalization experiment.

[0050] Figure 10 A detection frame map of the generalization experiment. DETAILED DESCRIPTION

[0051] To make the objectives, technical solutions and advantages of the present application clearer, the technical solutions in the present application will be described clearly and completely below. Obviously, the described embodiments are part of the embodiments of the present application, rather than all the embodiments. Based on the embodiments in the present application, all other embodiments obtained by those of ordinary skill in the art without creative labor fall within the scope of protection of the present application. The following embodiments are used to illustrate the present application, but cannot be used to limit the scope of the present application.

[0052] In this embodiment, a construction site small-scale target detection method based on an improved YOLO belongs to the technical field of computer vision and construction engineering monitoring. In view of the problems of existing YOLO model, such as small-scale target feature aggregation loss, high missing detection rate caused by cross-scale feature misplacement, and large consumption of computing resources in the construction site, the following technical solutions are proposed.

[0053] As shown in Figure 1 A construction site small-scale target detection method based on an improved YOLO includes the following steps:

[0054] S1, a construction site target detection data set SODA containing 15 types of construction targets is constructed, and is divided into a training set, a validation set and a test set;

[0055] The construction site object detection dataset SODA is integrated by screening existing public construction site image data, and specifically includes:

[0056] Step S11, data composition and screening dimension

[0057] The object categories cover 15 common target categories in the construction site, including person, helmet, vest, board, wood, rebar, brick, scaffold, handcart, cutter, ebox, hopper, hook, fence and slogan, including small-scale targets and large-scale targets, meeting the detection needs of complex scenes; The collected scenes include multi-view images such as high-altitude aerial view and close-up operation, as well as multi-illumination conditions such as sunny day, cloudy day and night, to improve the generalization ability of the model. Multi-view design makes the model adapt to the detection needs of different deployment locations, and multi-illumination scenes enhance the adaptability of all-weather detection in the construction environment.

[0058] Step S12, dataset division

[0059] The construction site object detection dataset SODA obtained by screening is divided into a training set, a validation set and a test set in a ratio of 7:1:2; Among them, the training set is used for model training and optimization of network parameters, the validation set is used for adjustment of hyperparameters and monitoring of early stopping strategy, and the test set is used for final model performance evaluation. This division ensures the sufficiency of model training and the objectivity of generalization ability evaluation, covering the diversity of construction site scenes.

[0060] S2, taking YOLOv8 as the benchmark model, integrating the multi-path multi-scale hollow convolution feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), and constructing a construction site small-scale target detection model based on improved YOLO;

[0061] YOLOv8 is used as the benchmark model, including the input layer, the backbone network (Backbone), the neck network (Neck) and the prediction head (Head), the multi-path multi-scale hollow convolution feature extraction module (MPFD) is embedded in the backbone network, and the adaptive multi-level feature fusion module (AMFF) is introduced in the neck network, to construct a construction site small-scale target detection model based on improved YOLO;

[0062] The multi-path multi-scale hollow convolution feature extraction module (MPFD) is used to relieve the problem of excessive aggregation of small-scale target features by using different expansion rate hollow convolutions in a multi-path parallel manner, expand the feature information distribution, and improve the deep feature expression capability; the adaptive multi-level feature fusion module (AMFF) is used to construct an adaptive multi-level feature fusion structure, combine a self-learning weight mechanism with a cross-scale pre-fusion strategy, solve the cross-scale feature misplacement problem, reduce the risk of small target features being swallowed by large target features, enhance the multi-scale feature alignment capability, and significantly improve the small target detection precision and inference real-time performance while maintaining lightweight.

[0063] Step S21, the backbone network and the multi-path multi-scale hollow convolution feature extraction module (MPFD)

[0064] The backbone network adopts the feature extraction architecture of YOLO, and a multi-path multi-scale hollow convolution feature extraction module (MPFD) is embedded in a key layer. The multi-path multi-scale hollow convolution feature extraction module (MPFD) includes:

[0065] As shown in Figure 2 The multi-path multi-scale hollow convolution design first performs a convolution operation on the input features with a convolution kernel size of 3, compresses the channel number from c to c / 2, and then divides the feature map into three parallel paths: the first path uses 3×3 convolution with an expansion rate of 1 to restore the channel number to c, focusing on extracting the accurate position information of small-scale targets; the second path uses 3×3 hollow convolution with an expansion rate of 3 to expand the channel number back to c, achieving the expansion of small-scale features in the medium scale; the third path uses 3×3 hollow convolution with an expansion rate of 5 to complete the channel conversion from c / 2 to c, achieving the expansion of small-scale features in the large scale. After the three paths of features are spliced, the channel information is integrated through 1×1 convolution, and finally the MFD multi-scale feature map is output by adding the residual of the input features, which fully realizes the multi-level feature capture from local details to global semantics.

[0066] Feature fusion: after the three-path feature splicing, the channel is adjusted through 3×3 convolution, and the original input features are spliced to compensate for the information loss in feature processing, and finally the MPFD multi-scale feature map containing rich small-scale target details is output; through such a process, the extraction and fusion of multi-scale features are realized, and the feature processing capability is enhanced. This design effectively alleviates the problem of excessive aggregation of features caused by downsampling, and improves the deep feature expression capability.

[0067] The backbone network improves the small-scale target feature expression capability by integrating the multi-path multi-scale hollow convolution feature extraction module (MPFD):

[0068] Multi-path and multi-scale dilated convolution feature extraction module (MPFD) in multi-path dilated convolution feature extraction structure: after the input feature is compressed by 3x3 convolution, it is divided into three parallel processing: ordinary convolution with expansion rate 1 retains position information, and dilated convolution with expansion rates 3 and 5 expands feature space distribution. After the three features are spliced and adjusted by 3x3 convolution, they are fused with the original input, and the feature map containing rich details is output.

[0069] Lightweight advantage: reduce the computational complexity, avoid the over-aggregation of small target features in deep feature maps, and improve the integrity of feature expression.

[0070] Step S22, neck network and adaptive multi-level feature fusion module (AMFF)

[0071] As shown in Figure 3 , the neck network introduces an adaptive multi-level feature fusion module (AMFF) to perform adaptive multi-level fusion on the MPFD multi-scale feature map output by the adaptive multi-level feature fusion module (AMFF), which specifically includes:

[0072] Adaptive weighting mechanism: adjust the channel of the MPFD multi-scale feature map by 1x1 convolution, extract the feature through the CBS module composed of 1x1 convolution, BN layer and SILU activation function, and generate self-learning weights through Sigmoid function, then perform adaptive weighting on the feature map and match the scale through bilinear interpolation.

[0073] Multi-level pre-fusion strategy: before cross-scale splicing, combine the original features, forward weighted features and reverse weighted features of the current scale for three-way fusion to eliminate the information imbalance and misalignment of different scale features, reduce the risk of small target features being swallowed by large target features, and enhance the multi-scale feature alignment capability.

[0074] As the core of feature fusion, the adaptive multi-level feature fusion module (AMFF) performs adaptive weighting on different scale feature maps through the adaptive weighting mechanism, and after matching the scale through bilinear interpolation, it fuses the features of the current scale, weighted features and reverse weighted features, realizes multi-level pre-fusion of cross-scale features, effectively alleviates the position misalignment problem in traditional fusion, and avoids the swallowing of small target features by large target features.

[0075] S3. Use the training set divided in step S1 to train the small-scale target detection model for construction sites based on improved YOLO constructed in step S2. Through training set optimization, a trained small-scale target detection model for construction sites based on improved YOLO is obtained. The model has learned the feature expressions of multiple categories of small-scale targets at the construction site. Use the validation set divided in step S1 to verify the trained small-scale target detection model for construction sites based on improved YOLO. Through validation set monitoring and early stopping strategy screening, an optimized generalized small-scale target detection model for construction sites based on improved YOLO is obtained. The model achieves the best balance on both the training set and the validation set. Input the test set divided in step S1 into the optimized generalized small-scale target detection model for construction sites based on improved YOLO for testing. Through test set evaluation, an efficient small-scale target detection model that adapts to the deployment requirements of edge computing equipment at the construction site is finally obtained. The model shows excellent detection performance and generalization ability in indicators such as AP50, APS, and ARS.

[0076] The specific method of step S3 is:

[0077] The training set divided in step S1 is used to optimize the network parameters of the construction site small-scale target detection model based on the improved YOLO, so that the construction site small-scale target detection model based on the improved YOLO can fully learn the feature expressions of multiple categories of small-scale targets at the construction site under multiple perspectives such as high-altitude overlooking and close-up operations, and multiple lighting scenarios such as sunny and rainy day and night on the training set, continuously strengthen the ability to extract small target detail features and cross-scale features, reduce the model loss function value, and after a certain number of rounds of training, obtain a trained construction site small-scale target detection model based on the improved YOLO;

[0078] Input the validation set divided in step S1 into the trained small-scale target detection model for construction sites based on improved YOLO for verification, monitor the overfitting risk during the training process of the small-scale target detection model for construction sites based on improved YOLO, dynamically adjust the hyperparameter learning rate and weight decay coefficient, and trigger the early stopping strategy when the performance index of the small-scale target detection model for construction sites based on improved YOLO on the validation set no longer improves in multiple consecutive training rounds and the convergence condition is met to ensure the generalization ability of the small-scale target detection model for construction sites based on improved YOLO in complex construction environments, balance the balance between small target feature retention and large target feature fusion, and obtain an optimized generalized small-scale target detection model for construction sites based on improved YOLO;

[0079] The test set divided in step S1 is input into the optimized generalization construction site small-scale target detection model based on improved YOLO for testing, the construction site small-scale target detection model based on improved YOLO generates a prediction result, and the detection accuracy and robustness of the model on unseen data are quantified by AP50, APS and ARS indicators, and the adaptability of the construction site small-scale target detection model based on improved YOLO in the following three aspects is verified:

[0080] 1) the ability to capture small-scale target edge details;

[0081] 2) the correction effect of cross-scale feature misplacement;

[0082] 3) robustness under complex lighting conditions.

[0083] Finally, an efficient small-scale target detection model that meets the deployment requirements of construction site edge computing devices in terms of detection accuracy, inference speed and resource consumption is obtained.

[0084] In step S3, the model training strategy is optimized:

[0085] Early stopping strategy: terminate training early if the performance of the validation set does not improve for 100 consecutive rounds, balancing training sufficiency and efficiency.

[0086] Real-time guarantee: through the lightweight design of the multi-path multi-scale feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), the model inference speed is maintained at 102FPS, meeting the real-time detection requirements of the construction site.

[0087] Through the above technical solutions, the present application effectively solves the problems of feature aggregation loss and feature misplacement in construction site small-scale target detection, significantly improves the small target detection accuracy and model generalization ability, and provides efficient technical support for building engineering safety management.

[0088] Through the above embodiments, the construction site small-scale target detection model based on improved YOLO effectively solves the problems of feature aggregation loss and feature misplacement in construction site small-scale target detection, significantly improves the small target detection accuracy and model generalization ability, and provides efficient technical support for building engineering safety management.

[0089] The application discloses a construction site small-scale target detection method based on an improved YOLO, and belongs to the technical field of computer vision and building engineering monitoring. In view of the problems of the existing YOLO model, such as easy aggregation loss of small-scale target features in a construction site scene, high missing detection rate caused by cross-scale feature misplacement and large consumption of computing resources, the application optimizes the model structure by integrating a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF). The multi-path multi-scale hollow convolution feature extraction module (MPFD) significantly alleviates the problem of feature over-aggregation of small-scale targets caused by down-sampling through multi-path parallel hollow convolution with different expansion rates, and expands the feature information distribution. The adaptive multi-level feature fusion module (AMFF) effectively solves the cross-scale feature misplacement problem through an adaptive weighting mechanism and a multi-level feature pre-fusion strategy, and reduces the risk of small target features being swallowed by large target features. Meanwhile, a multi-scene and multi-resolution construction site dataset SODA containing 15 types of construction targets is integrated, and a generalization verification system is formed in combination with a WOTR blind obstacle and a Pictor-v3 protective equipment dataset, so that the cross-scene adaptability of the model is effectively improved. On the basis of controlling the model parameter quantity and the calculation complexity, the small-scale target detection precision and the reasoning real-time performance are significantly improved, the construction site edge computing device deployment is adapted, and efficient technical support is provided for building engineering safety management and resource scheduling.

[0090] As shown in Figure 4 , a construction site small-scale target detection model structure based on an improved YOLO is displayed. An input layer receives construction site images in multiple visual angles such as high-altitude aerial view and close-range operation and under multiple light conditions such as sunny, rainy and day and night; a backbone network takes YOLO as a benchmark architecture, embeds a multi-path multi-scale hollow convolution feature extraction module (MPFD), extracts small target multi-scale features through three parallel paths of hollow convolution with different expansion rates after 3*3 convolution preprocessing, and then fuses and outputs; a neck network contains an adaptive multi-level feature fusion module (AMFF), performs channel adjustment, feature extraction and other operations on the output features of the multi-path multi-scale hollow convolution feature extraction module (MPFD), adaptively fuses the features, and further fuses the features in combination with a DsPAN structure; a prediction network performs target detection based on the output of the DsPAN. In addition, the model is further optimized through lightweight design and multi-scene adaptation.

[0091] Simulation conditions

[0092] The simulation experiment of the application is completed based on a high-performance computing platform, the hardware configuration includes an AMD EPYC 7402 24-core processor CPU, 24 GB RAM and an NVIDIA GeForce RTX 4090 GPU, and the software environment is Python 3.10.14, PyTorch 2.2.2 and CUDA 12.1. In the training process, a stochastic gradient descent (SGD) optimizer is used, the initial learning rate is 0.01, the momentum is 0.937, the batch size is 32, dynamic adjustment is used to optimize the memory utilization, and the maximum training round is set to 1000, combined with the early stopping strategy to prevent overfitting. The experiment takes a lightweight YOLO as a benchmark model, and does not use pre-trained weights to ensure the fairness of the comparative experiment.

[0093] Simulation content

[0094] The simulation content is divided into performance comparison and ablation study. The performance comparison experiment is performed on the SODA, WOTR and Pictor-v3 datasets, and compared with mainstream models such as YOLOv5, YOLOv6, YOLOv8, YOLOv11, FasterR-CNN, RT-DETR-r18, SPD-YOLO and SOC-YOLO, the evaluation indexes include AP50, AP50:95, APS (small target AP), ARS (small target AR), GFLOPs, parameter quantity and FPS, to analyze the detection accuracy and computing efficiency. The ablation experiment gradually introduces the multi-path multi-scale hollow convolution feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), verifies the independent contribution and synergistic effect of the modules on the model performance, and assists in the visualization analysis of the heat map and the detection result, to directly show the optimization effect of the modules on the small target feature expansion and the cross-scale feature alignment.

[0095] Simulation results

[0096] The experimental results show that Table 1 ablation experiment results and Table 2 performance comparison experiment results show the key index performance of different models.

[0097]

[0098] Table 1 Ablation experiment results

[0099]

[0100] Table 2 Performance comparison experiment results

[0101] In the performance comparison, the improved YOLO-based small-scale target detection model for construction sites achieves 80.9%, 48.6%, 15.7%, and 23.7% in AP50, AP50:95, APS, and ARS indicators, respectively, with a 4.5%, 3.3%, 9.1%, and 11.3% improvement over Baseline, a GFLOPs of 14.8G, a parameter of 3.7M, an FPS of 102, and a balance between accuracy and real-time performance. Ablation experiments show that the introduction of the multi-path multi-scale feature extraction module (MPFD) alone improves APS by 0.8% and AR by 0.7%; the introduction of the adaptive multi-level feature fusion module (AMFF) alone improves APS by 5.8% and ARS by 9.3%; and the combination of the two further improves the detection performance, verifying the complementarity between the modules. The improved YOLO-based small-scale target detection model for construction sites can accurately focus on small target areas, reduce missed and false detections, and have more complete feature responses and better edge detail retention. The lightweight design makes the model suitable for deployment on edge devices in construction sites, providing an efficient solution for real-time detection of small-scale targets in complex environments.

[0102] As Figure 5 shown in FIG. 6, the visualized heat map further verifies the independent contribution and synergistic effect of the multi-path multi-scale feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF). The heat map can intuitively reflect the model's attention to different regions, Figure 5 showing the attention distribution of each model in the construction site target detection task. It can be observed that the multi-path multi-scale feature extraction module (MPFD) effectively suppresses the feature aggregation problem and enhances the model's global perception ability, allowing it to evenly focus on each target within the entire image range, especially in small target detection. The YOLO-AF, on the other hand, strengthens the precision of feature fusion through the adaptive multi-level feature fusion module (AMFF), allowing attention to more completely cover the target area. The improved YOLO-based small-scale target detection model for construction sites combines the global perception ability of the multi-path multi-scale feature extraction module (MPFD) and the precise feature alignment ability of the adaptive multi-level feature fusion module (AMFF), focusing on the target while accurately focusing on the target location and more comprehensively covering the target area, avoiding the loss or shift of feature information. This combination of "full" and "accurate" makes the improved YOLO-based small-scale target detection model for construction sites perform better in the target detection task, especially in small-scale target detection. Figure 6The detection results of each module in the ablation experiment are shown. It can be clearly seen that with the addition of the multi-path multi-scale dilated convolutional feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), the model's missed detection rate and false detection rate have significantly decreased, further verifying the importance of their synergy in improving target detection accuracy.

[0103] like Figure 7 、 Figure 8 As shown in the figure, in order to further intuitively demonstrate the feature perception ability and detection effect of the proposed construction site small-scale target detection model based on improved YOLO in the target detection task, the present invention selects a YOLO architecture model with a similar structure to the construction site small-scale target detection model based on improved YOLO for visualization analysis, mainly including feature heat map analysis and detection result comparison. Figure 7 From the feature heat map results, it can be seen that the small-scale target detection model based on the improved YOLO at the construction site can achieve more accurate target attention in complex construction site scenes. Its attention is mainly focused on the key target area, and the coverage of the target area is more comprehensive. At the same time, the small-scale target detection model based on the improved YOLO at the construction site has a weak response in the background area, which significantly suppresses the interference information in the non-target area, reflecting excellent background suppression and feature extraction capabilities. The above results fully demonstrate that the small-scale target detection model based on the improved YOLO at the construction site has stronger feature perception and regional attention capabilities under complex background interference, which provides important support for the improvement of its overall detection performance. Further visualization of the detection results is compared. Figure 8 As shown in the figure, the improved YOLO-based small-scale target detection model for construction sites demonstrates higher detection accuracy and robustness in actual detection processes. Specifically, the improved YOLO-based small-scale target detection model for construction sites not only effectively reduces missed and false detections of small targets, but also predicts more accurate bounding box positions, shapes that more closely match the true target outline, and higher confidence scores, demonstrating stronger target positioning capabilities and detection reliability.

[0104] like Figure 9 、 Figure 10 As shown, from Figure 9The heat map of the generalization experiment can find that, compared with other advanced models, the construction site small-scale target detection model based on improved YOLO of the present application has higher accuracy and stability in the target attention area of WOTR and Pictor-v3 data sets, and can more comprehensively cover the complete contour and key area of the target. Especially in the small-scale target or complex defect target detection task, the construction site small-scale target detection model based on improved YOLO of the present application not only can accurately focus on the target body, but also can effectively suppress the interference of background noise and irrelevant areas, and has better feature perception and positioning ability. In contrast, the heat map results of other comparative models on the generalization data set generally have problems such as attention area deviation, incomplete target edge coverage, or false attention to background areas, especially in the scene where the target size is small or the occlusion is serious. The excellent performance of the heat map results of the construction site small-scale target detection model based on improved YOLO in the generalization experiment verifies its excellent feature perception ability and outstanding generalization performance in the small-scale target detection and construction site target detection tasks. Figure 10 In the WOTR and Pictor-v3 two generalization data sets, compared with other models, the construction site small-scale target detection model based on improved YOLO of the present application greatly relieves the problems of false detection and missed detection, further verifying the strong generalization performance of the construction site small-scale target detection model based on improved YOLO in the small-scale target detection field and the construction site target detection field. In addition, from the detection box visualization results of Figure 10 The detection advantage of the construction site small-scale target detection model based on improved YOLO of the present application in the WOTR and Pictor-v3 two generalization data sets is further verified. The comparison results show that the construction site small-scale target detection model based on improved YOLO of the present application greatly relieves the false detection and missed detection problems existing in other advanced detection models. Whether in the small target dense WOTR data set or in the complex background Pictor-v3 data set, the construction site small-scale target detection model based on improved YOLO has excellent target recognition performance, effectively reduces the occurrence of false detection and missed detection, and further embodies its strong feature expression ability and cross-scene robustness.

[0105] The application is particularly suitable for real-time detection and positioning of multiple types of small-scale targets such as workers, equipment and materials in complex construction environments. The application optimizes the model structure through a multi-path multi-scale hollow convolution feature extraction module and an adaptive multi-level feature fusion module, combines a multi-scene, multi-resolution construction site dataset SODA and a generalization verification dataset, significantly improves the small-scale target detection accuracy and inference real-time performance, while controlling the model parameter quantity and computational complexity, adapts to the deployment of edge computing devices in construction sites, and provides efficient technical support for construction safety management and resource scheduling.

[0106] The application also provides a construction site small-scale target detection system based on improved YOLO, comprising:

[0107] A dataset construction and division module is used to construct a construction site target detection dataset SODA containing multiple types of construction targets in step S1, and divide it into a training set, a verification set and a test set;

[0108] A construction site small-scale target detection model based on improved YOLO is constructed in step S2 by taking YOLO as a reference model, integrating a multi-path multi-scale hollow convolution feature extraction module (MPFD) and an adaptive multi-level feature fusion module (AMFF), and constructing a construction site small-scale target detection model based on improved YOLO;

[0109] The construction site small-scale target detection model based on improved YOLO is trained, verified and tested in step S3 by using the training set divided in step S1 to train the construction site small-scale target detection model based on improved YOLO constructed in step S2, optimizing the training set to obtain a trained construction site small-scale target detection model based on improved YOLO; the trained construction site small-scale target detection model based on improved YOLO is verified using the verification set divided in step S1, and an optimized generalization construction site small-scale target detection model based on improved YOLO is obtained through verification set monitoring and early stopping strategy screening; the test set divided in step S1 is input into the optimized generalization construction site small-scale target detection model based on improved YOLO for testing, and finally an efficient small-scale target detection model adapted to the deployment requirements of edge computing devices in construction sites is obtained through test set evaluation.

[0110] The application also provides a construction site small-scale target detection device based on improved YOLO, comprising:

[0111] A memory is provided for storing the computer program of the above-mentioned construction site small-scale target detection method based on improved YOLO, which is a computer-readable device;

[0112] Processor: used to execute the computer program to realize the improved YOLO-based construction site small-scale target detection method.

[0113] The application further provides a computer readable storage medium, which stores a computer program, and the computer program can realize the improved YOLO-based construction site small-scale target detection method when executed by a processor.

Claims

1. A construction site small-scale target detection method based on improved YOLO, characterized by: The following steps are involved: S1. Construct a construction site object detection dataset SODA containing multiple types of construction objects and divide it into training set, validation set and test set; S2. Using YOLO as the benchmark model, we integrated the multi-path multi-scale dilated convolutional feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF) to build a small-scale target detection model for construction sites based on the improved YOLO. S3. Use the training set divided in step S1 to train the construction site small-scale target detection model based on improved YOLO constructed in step S2, and obtain the trained construction site small-scale target detection model based on improved YOLO through training set optimization; use the validation set divided in step S1 to verify the trained construction site small-scale target detection model based on improved YOLO, and obtain the optimized generalized construction site small-scale target detection model based on improved YOLO through validation set monitoring and early stopping strategy screening; input the test set divided in step S1 into the optimized generalized construction site small-scale target detection model based on improved YOLO for testing, and finally obtain an efficient small-scale target detection model that adapts to the deployment requirements of edge computing equipment at the construction site through test set evaluation.

2. A construction site small-scale target detection method based on improved YOLO according to claim 1, characterized in that: The specific method of step S1 is: By screening publicly available construction site image data, we constructed a construction site target detection dataset SODA that contains multiple categories of construction targets, covering multiple perspectives such as high-altitude bird's-eye view and close-up operations, as well as multiple lighting conditions such as sunny days, cloudy days, and nighttime. The construction site dataset SODA is divided into training set, validation set, and test set.

3. A construction site small-scale target detection method based on improved YOLO according to claim 1, characterized in that: The specific method of step S2 is: YOLO is used as the baseline model, which includes the input layer, backbone network (Backbone), neck network (Neck) and prediction head (Head). A multi-path multi-scale dilated convolutional feature extraction module (MPFD) is embedded in the backbone network, and an adaptive multi-level feature fusion module (AMFF) is introduced in the neck network to build a small-scale target detection model for construction sites based on the improved YOLO. The multi-path multi-scale dilated convolution feature extraction module (MPFD) is used to receive input features. It first performs a convolution operation, converting the number of convolution channels from c to c / 2, using a convolution kernel size of 3. Subsequently, the features are divided into three parallel paths: The first path performs a convolution operation, changes the number of channels from c / 2 to c, and uses a convolution kernel size of 3 and a dilation rate of 1. The second path performs a convolution operation, changes the number of channels from c / 2 to c, and uses a convolution kernel size of 3 and a dilation rate of 3. The third path performs a convolution operation, changes the number of channels from c / 2 to c, and uses a convolution kernel size of 3 and a dilation rate of 5. The results of the three parallel paths are concatenated and the channels are adjusted through a 3x3 convolutional layer. Then, a convolution operation is performed to convert the number of convolution channels from 2c after the concatenation to c, and a convolution kernel size of 1 is used. Then, this result is added to the original input features to finally output the MPFD multi-scale feature map; The adaptive multi-level feature fusion module (AMFF) takes the MPFD multi-scale feature map as input, first adjusts the number of channels through 1×1 convolution, then enters the CBS module composed of 1×1 convolution, BN layer and SILU activation function for processing, and uses the Sigmoid function to generate a weight matrix. After calculating the forward weighted features and reverse weighted features based on the weight matrix, the weighted features are scale-matched through bilinear interpolation to keep the features of different scales consistent in size. Then, a multi-level pre-fusion operation is performed to fuse the original features of the current scale and the fused features after scale matching again. Finally, through cross-scale splicing, the features of different scales are spliced ​​together to output the fused and enhanced features.

4. A construction site small-scale target detection method based on improved YOLO according to claim 1, characterized in that: The specific method of step S3 is: The training set divided in step S1 is used to optimize the network parameters of the construction site small-scale target detection model based on the improved YOLO. After a certain number of rounds of training, a trained construction site small-scale target detection model based on the improved YOLO is obtained. Input the validation set divided in step S1 into the trained small-scale target detection model for construction sites based on improved YOLO for verification, monitor the overfitting risk during the training process of the small-scale target detection model for construction sites based on improved YOLO, dynamically adjust the hyperparameter learning rate and weight decay coefficient, and trigger the early stopping strategy when the performance index of the small-scale target detection model for construction sites based on improved YOLO on the validation set no longer improves in multiple consecutive training rounds and the convergence condition is met to ensure the generalization ability of the small-scale target detection model for construction sites based on improved YOLO in complex construction environments, balance the balance between small target feature retention and large target feature fusion, and obtain an optimized generalized small-scale target detection model for construction sites based on improved YOLO; The test set divided in step S1 is input into the optimized generalized small-scale target detection model based on improved YOLO at construction sites for testing. The small-scale target detection model based on improved YOLO at construction sites generates prediction results. The detection performance of the small-scale target detection model based on improved YOLO at construction sites on unseen data is evaluated through quantitative indicators. Finally, an efficient small-scale target detection model that adapts to the deployment requirements of edge computing equipment at construction sites is obtained.

5. A construction site small-scale target detection system based on an improved YOLO based on the method according to any one of claims 1 to 4, characterized in that: include: The dataset construction and partitioning module is used to construct the construction site object detection dataset SODA, which contains multiple types of construction objects, and divide it into training set, validation set and test set; A module for building a small-scale target detection model for construction sites based on an improved YOLO. This module uses YOLO as a baseline model, integrates the multi-path multi-scale dilated convolutional feature extraction module (MPFD) and the adaptive multi-level feature fusion module (AMFF), and builds a small-scale target detection model for construction sites based on an improved YOLO. The training, verification and testing module of the construction site small-scale target detection model based on improved YOLO is used to train the construction site small-scale target detection model based on improved YOLO using the training set, and obtain the trained construction site small-scale target detection model based on improved YOLO through training set optimization; the trained construction site small-scale target detection model based on improved YOLO is verified using the validation set, and the optimized generalized construction site small-scale target detection model based on improved YOLO is obtained through validation set monitoring and early stopping strategy screening; the test set is input into the optimized generalized construction site small-scale target detection model based on improved YOLO for testing, and finally an efficient small-scale target detection model that adapts to the deployment requirements of edge computing equipment on the construction site is obtained through test set evaluation.

6. A construction site small-scale target detection device based on improved YOLO, characterized in that: include: Memory: a computer-readable device storing a computer program for a construction site small-scale target detection method based on an improved YOLO as described in any one of claims 1 to 4; Processor: used to implement the construction site small-scale target detection method based on improved YOLO as described in any one of claims 1-4 when executing the computer program.

7. A computer-readable storage medium, characterized in that The computer-readable storage medium stores a computer program, which, when executed by a processor, can implement a construction site small-scale target detection method based on improved YOLO as described in any one of claims 1 to 4.

Images