Outdoor parking space intelligent detection method, system, equipment and medium

Abstract

The invention relates to the technical field of outdoor parking space detection, and provides an outdoor parking space intelligent detection method, system and device and a medium, the outdoor parking space intelligent detection method comprises the following steps: S1, constructing an OP-RTDETR model; s2, intelligent detection of outdoor parking is realized through an OP-RTDETR model; wherein in the step S1, a specific method for constructing the OP-RTDETR model comprises the following steps: S10, creating a GC-VanilaNet as a feature extraction network; s11, using an Officient Hybrid Encoder of the RTDETR as an encoder of the OP-RTDETR model, and using the Officient Hybrid Encoder of the RTDETR as an encoder of the OP-RTDETR model; s12, a query selection strategy is minimized through the uncertainty of the RT-DETR, and the Bayesian uncertainty is optimized; and S13, training the decoder through a GCDN method. According to the invention, by constructing the OP-RTDETR model, intelligent detection of outdoor parking can be realized, and the detection precision is improved while high real-time performance is realized.

Description

Technical Field

[0001] The present invention relates to the technical field of outdoor parking space detection, and in particular to an outdoor parking space intelligent detection method, system, equipment and medium. Background Art

[0002] Outdoor parking space detection is an important technology that helps reduce congestion and improve urban traffic management. By automatically identifying vacant parking spaces, traffic flow disruptions and local congestion can be reduced. Existing non-image-based parking detection methods, such as geomagnetic sensors, radio frequency, ultrasound, and radar, are expensive and have limited applicability. In contrast, image-based parking detection methods are cheaper, more accurate, and can detect more types of parking scenarios, especially for open parking lots.

[0003] Existing neural network-based outdoor parking space detection methods still have deficiencies in model architecture, resulting in the real-time problem not being completely solved. First, many methods rely on complex model structures, which increases the computational burden and affects the detection speed. For example, ShuffleDet reduces calculations through group convolution and channel shuffling techniques, but the complexity of the model architecture is still high, which makes real-time processing difficult. In addition, although methods like BCFPL reduce the computational load by reducing the image resolution, this is only a local optimization and cannot fundamentally solve the complexity of the model architecture. Complex model architectures are often inefficient when processing high-resolution images, especially in application scenarios with high real-time requirements. These complex structures will significantly slow down the detection speed.

[0004] Therefore, the current outdoor parking space detection method is still too complex in model architecture, and the real-time problem has not been fully solved. Future optimization directions should focus on further simplifying the model structure and improving the efficiency of feature extraction, so as to meet the needs of real-time detection while maintaining high detection accuracy. Summary of the invention

[0005] Based on this, in order to improve the real-time performance and accuracy of outdoor parking space detection, the present invention provides an outdoor parking space intelligent detection method, system, device and medium, and its specific technical solution is as follows:

[0006] An outdoor parking space intelligent detection method comprises the following steps:

[0007] S1, build OP-RTDETR model;

[0008] S2, realizing intelligent detection of outdoor parking through the OP-RTDETR model;

[0009] Among them, in step S1, the specific method of constructing the OP-RTDETR model includes the following steps:

[0010] S10, create GC-VanillaNet as the feature extraction network;

[0011] S11, using the Efficient Hybrid Encoder of RTDETR as the encoder of the OP-RTDETR model;

[0012] S12, optimizes Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR;

[0013] S13, train the decoder using the GCDN method.

[0014] The outdoor parking space intelligent detection method creates GC-VanillaNet as a feature extraction network, which improves the reasoning speed of the OP-RTDETR model; at the same time, the Efficient Hybrid Encoder of RTDETR is used as the encoder part of the OP-RTDETR model, and the scale-based feature interaction based on attention reduces the computational cost by performing scale-based interaction on S5 only in the single-scale Transformer encoder; through the uncertainty minimization query selection strategy of RT-DETR, the Bayesian uncertainty is optimized to ensure the feature quality of the selected encoder; the decoder is trained by the GCDN method, which further enhances the detection ability of the OP-RTDETR model in complex scenes and multi-target situations such as outdoor parking lots. That is to say, the outdoor parking space intelligent detection method can realize the intelligent detection of outdoor parking by constructing the OP-RTDETR model, and improve the detection accuracy while achieving high real-time performance.

[0015] Preferably, in step S10, the specific method of creating GC-VanillaNet as a feature extraction network includes the following steps:

[0016] VanillaNet is modified based on the group convolution technology to obtain the GC-VanillaNet.

[0017] Preferably, in step S11, the Efficient Hybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

[0018] Preferably, in step S13, the specific method of training the decoder by the GCDN method comprises the following steps:

[0019] S130, upgrades DINO’s contrast noise reduction module to the group level;

[0020] S131, perform independent self-attention calculation on the noise part and the object query in each group and perform one-to-one assignment to achieve group-level one-to-many matching.

[0021] An outdoor parking space intelligent detection system, comprising:

[0022] OP-RTDETR model for intelligent detection of outdoor parking;

[0023] The OP-RTDETR model includes a GC-VanillaNet feature extraction network, an Efficient HybridEncoder, and a decoder trained by the GCDN method.

[0024] Preferably, the Efficient Hybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

[0025] Preferably, the outdoor parking space intelligent detection system further includes:

[0026] The uncertainty minimization query selection module is used to optimize the Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR.

[0027] An outdoor parking space intelligent detection device, comprising:

[0028] Controller;

[0029] A memory storing executable instructions;

[0030] Among them, the executable instructions can be run on the controller and implement the outdoor parking space intelligent detection method.

[0031] A computer-readable storage medium stores a computer program, which implements the outdoor parking space intelligent detection method when the computer program is executed. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] The present invention can be further understood from the following description in conjunction with the accompanying drawings. The components in the figures are not necessarily drawn to scale, but the emphasis is placed on illustrating the principles of the embodiments. In different views, the same reference numerals designate corresponding parts.

[0033] Figure 1 It is a flowchart of a specific method for constructing an OP-RTDETR model in one embodiment of the present invention;

[0034] Figure 2 is a comparison diagram of standard convolution and group convolution in one embodiment of the present invention;

[0035] Figure 3 is a comparison diagram of DINO's contrastive denoising (CDN) and the proposed group contrastive denoising (GCDN) in one embodiment of the present invention;

[0036] Figure 4 It is a schematic diagram of the overall architecture of the OP-RTDETR model in one embodiment of the present invention;

[0037] Figure 5 is a chart of evaluation results of different detection algorithms in a specific experiment of the present invention;

[0038] Figure 6 It is the AP chart of RT-DETR under different backbone networks in the specific experiment of the present invention;

[0039] Figure 7 It is a graph of the computational complexity and speed of different networks during the inference process in the specific experiments of the present invention;

[0040] Figure 8 is a performance comparison chart of models with and without GCDN in a specific experiment of the present invention. DETAILED DESCRIPTION

[0041] In order to make the purpose, technical solution and advantages of the present invention more clearly understood, the present invention is further described in detail below in conjunction with its embodiments. It should be understood that the specific implementation methods described herein are only used to explain the present invention and do not limit the protection scope of the present invention.

[0042] It should be noted that when an element is referred to as being "fixed to" another element, it may be directly on the other element or there may be a central element. When an element is considered to be "connected to" another element, it may be directly connected to the other element or there may be a central element at the same time. The terms "vertical", "horizontal", "left", "right" and similar expressions used herein are for illustrative purposes only and do not represent the only implementation method.

[0043] Unless otherwise defined, all technical and scientific terms used herein have the same meaning as those commonly understood by those skilled in the art of the present invention. The terms used herein in the specification of the present invention are only for the purpose of describing specific embodiments and are not intended to limit the present invention. The term "and / or" used herein includes any and all combinations of one or more related listed items.

[0044] The “first” and “second” mentioned in the present invention do not represent specific quantities and orders, but are merely used to distinguish names.

[0045] For outdoor parking, it is crucial to provide available parking space information in real time. During peak hours and high-traffic areas, parking demand increases suddenly, which requires the outdoor parking detection model to process and provide parking data in real time. Real-time parking information can help drivers find parking spaces faster and reduce traffic congestion. This not only improves road efficiency and eases urban traffic pressure, but also reduces fuel consumption and exhaust emissions, bringing significant environmental benefits.

[0046] Non-image recognition methods for outdoor parking space detection mainly include geomagnetic sensors, radio frequency identification (RFID), ultrasonic sensors, and radar. These methods have their own characteristics, but they generally have the problems of high cost and poor versatility. Geomagnetic sensors detect the presence of vehicles by detecting changes in the earth's magnetic field. By installing a geomagnetic sensor in each parking space, it is possible to provide high detection accuracy. RFID technology detects vehicles by setting up RFID readers and writers in the parking lot and installing RFID tags on the vehicles. Ultrasonic sensors detect the presence of objects by emitting and receiving ultrasonic waves. This method works well when detecting at close range, but may be interfered with in complex environments. Radar technology uses electromagnetic waves to detect the use of parking spaces. Radar can work stably in all weather conditions and has good detection performance. However, its complex installation and high cost limit its application in large-scale parking lots.

[0047] In contrast, the image recognition-based method uses cameras to collect images of parking lots and uses computer vision technology to analyze the images to detect the status of parking spaces, which can save costs and maintain high accuracy under the premise of large-scale detection. Traditional object detection models, such as YOLOv5 (You Only Look Once version 5), are known for their real-time detection capabilities. By inputting the entire image into the network for processing at one time, they achieve fast and efficient object detection; while DETR (Detection Transformer) introduces a self-attention mechanism and applies the Transformer architecture to object detection, achieving end-to-end object detection and more robust results. RT-DETR (Real-Time DETEction TRansformer, a real-time end-to-end detector based on the DETR architecture) is an improved version of DETR. By introducing an Efficient Hybrid Encoder and an optimized query selection strategy, the detection speed and accuracy are further improved. RT-DETR uses a hybrid encoder structure for feature extraction, and adopts a query selection strategy that minimizes uncertainty in the decoder part, which effectively improves the real-time performance and accuracy of object detection, especially in complex scenes.

[0048] The following models have made model innovations for outdoor parking detection based on traditional models. BCFPL (Binary classification ConvNet based Fast Parking space recognition with Lowresolution image) proposes to use reduced image resolution and bilinear interpolation to generate low-resolution images for parking space recognition. At the same time, a lightweight binary classification convolutional neural network is proposed to identify parking spaces in low-resolution images, which improves the processing speed while maintaining the same accuracy as high-resolution images.

[0049] The ShuffleDet model uses channel shuffling and grouped convolution to build a lightweight network structure, combined with the improved Inception module and deformable convolution, to achieve efficient feature extraction and geometric adaptability. At the same time, the domain adaptation block (DAB) further enhances the performance of the pre-trained model in new domains. These innovations enable ShuffleDet to achieve higher detection accuracy and computational efficiency.

[0050] T-YOLO (Tiny YOLO, a lightweight simplified version of YOLO) proposes a lightweight deep object detection model based on the YOLO-v5 architecture, specifically for detecting tiny vehicles in large parking lots. By introducing a multi-scale module based on an efficient neural network (ENet, Efficient Net), the model is able to learn deep discriminative feature representations at different scales and automatically determine the scale that is most suitable for detecting objects in the scene. In addition, combined with the channel / spatial attention mechanism, the model significantly improves speed and accuracy without increasing the number of training parameters.

[0051] However, the existing neural network-based outdoor parking space detection methods still have deficiencies in model architecture, resulting in the real-time problem not being completely solved. First, many methods rely on complex model structures, which increases the computational burden and affects the detection speed. For example, ShuffleDet reduces computation through group convolution and channel shuffle techniques, but the complexity of the model architecture is still high, which makes real-time processing difficult. Similarly, T-YOLO improves the detection accuracy of small targets by introducing multi-scale modules and attention mechanisms, but these additional modules also increase the complexity of the model, thereby affecting the detection speed. In addition, although methods such as BCFPL reduce the computational load by reducing the image resolution, this is only a local optimization and cannot fundamentally solve the complexity of the model architecture. Complex model architectures are often inefficient when processing high-resolution images, especially in application scenarios with high real-time requirements, these complex structures will significantly slow down the detection speed.

[0052] Therefore, the current outdoor parking space detection method is still too complex in model architecture, and the real-time problem has not been fully solved. Future optimization directions should focus on further simplifying the model structure and improving the efficiency of feature extraction, so as to meet the needs of real-time detection while maintaining high detection accuracy.

[0053] In order to solve the problems of insufficient real-time and accuracy of the existing outdoor parking space detection method based on neural network, the present invention provides an outdoor parking space intelligent detection method which can improve the real-time and accuracy of outdoor parking space detection.

[0054] The outdoor parking space intelligent detection method comprises the following steps:

[0055] S1, build the OP-RTDETR (Outdoor Parking Real-Time detection transformer) model.

[0056] S2, realizing intelligent detection of outdoor parking through the OP-RTDETR model. Specifically, the collected outdoor parking lot image is input into the OP-RTDETR model, the outdoor parking lot image is analyzed and processed through the OP-RTDETR model, the parking spaces in the outdoor parking lot are identified, and the intelligent detection of outdoor parking is realized.

[0057] Among them, Figure 1 As shown, in step S1, the specific method of constructing the OP-RTDETR model includes the following steps:

[0058] S10, create GC-VanillaNet as the feature extraction network.

[0059] S11, use the Efficient Hybrid Encoder of RTDETR as the encoder of the OP-RTDETR model.

[0060] S12, optimizes Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR.

[0061] S13, train the decoder using the GCDN method.

[0062] The outdoor parking space intelligent detection method improves the reasoning speed of the OP-RTDETR model by creating GC-VanillaNet as the feature extraction network; at the same time, the Efficient Hybrid Encoder of RTDETR is used as the encoder part of the OP-RTDETR model, and the scale-based feature interaction based on attention reduces the computational cost by performing scale-based interaction on S5 only in the single-scale Transformer encoder; the Bayesian uncertainty is optimized through the uncertainty minimization query selection strategy of RT-DETR, and the feature quality of the selected encoder is ensured; the decoder is trained by the GCDN method, which further enhances the detection ability of the OP-RTDETR model in complex scenes and multi-target situations such as outdoor parking lots. That is to say, the outdoor parking space intelligent detection method can realize the intelligent detection of outdoor parking by constructing the OP-RTDETR model, and improves the detection accuracy while achieving high real-time performance.

[0063] As a preferred technical solution, in step S10, the specific method of creating GC-VanillaNet as a feature extraction network includes the following steps:

[0064] The VanillaNet is modified based on the group convolution technology to obtain the GC-VanillaNet (Groupconvolution VanillaNet).

[0065] VanillaNet is a neural network architecture that emphasizes minimalist design. VanillaNet achieves a compact and intuitive hierarchical structure by avoiding excessive depth, skip connections, and complex operations. The nonlinearity and performance of the network are improved by designing innovative deep training strategies and sequence-based activation functions, which makes VanillaNet comparable to deep neural networks in performance, but with much less depth and latency. Grouped convolution is a technique commonly used in convolutional neural networks (CNN) to improve computational efficiency and reduce model complexity. The basic idea is to divide the input feature map into multiple groups, each of which is convolved with only part of the convolution kernels. This approach reduces model parameters and computational costs, allowing the network to process input data more efficiently without significantly reducing accuracy.

[0066] By applying the group convolution technique to VanillaNet, Figure 2 As shown, this embodiment proposes a new network variant, GC-VanillaNet. While retaining the simplicity and efficiency of VanillaNet, GC-VanillaNet further improves the processing speed of the model by grouping the workload of convolution operations. The number of convolution operations in each group is reduced, which reduces the overall computational burden, making GC-VanillaNet particularly suitable for real-time tasks or application scenarios that require fast reasoning. Experiments show that after using GC-VanillaNet, the speed of the OP-RTDETR model is significantly improved compared to using VanillaNet.

[0067] As a preferred technical solution, in step S11, the Efficient Hybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

[0068] The Efficient Hybrid Encoder consists of two modules, namely, Attention-based Intra-scale Feature Interaction (AIFI) and CNN-based Cross-scale Feature Fusion (CCFF). AIFI further reduces the computational cost by performing intra-scale interaction on S5 only in a single-scale Transformer encoder.

[0069] As a preferred technical solution, in step S13, the specific method of training the decoder by the GCDN method includes the following steps:

[0070] S130, upgrades DINO's contrast noise reduction module to the group level.

[0071] S131, perform independent self-attention calculation on the noise part and the object query in each group and perform one-to-one assignment to achieve group-level one-to-many matching.

[0072] In order to further improve the performance of the OP-RTDETR model, this embodiment refers to the group-level allocation strategy of GroupDETR and proposes a Group-Wise Contrast DeNoising (GCDN) method, such as Figure 3 As shown in the figure. GroupDETR significantly accelerates the training convergence speed and improves the performance of various DETR-based models by introducing a one-to-many allocation method at the group level. Drawing on this idea, GCDN promotes the contrast denoising module of DINO (DETR with Improved DeNoising Anchor Boxes) to the group level during model training, performs independent self-attention calculations on the noise part in each group and the object query, and performs a one-to-one allocation, thus achieving a one-to-many matching at the group level and further improving the detection accuracy and training efficiency. In the model inference stage, the process is consistent with the conventionally trained DETR, requiring only one set of queries without modifying the architecture, thus ensuring the inference speed of the model.

[0073] This embodiment also provides an outdoor parking space intelligent detection system. Figure 4 As shown, it includes: an OP-RTDETR model for realizing intelligent detection of outdoor parking.

[0074] The OP-RTDETR model includes a GC-VanillaNet feature extraction network, an Efficient HybridEncoder, and a decoder trained by the GCDN method.

[0075] Specifically, the Efficient Hybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

[0076] The outdoor parking space intelligent detection system also includes: an uncertainty minimization query selection module, which is used to optimize Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR. By optimizing Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR, the initial query quality of the decoder can be improved and the quality of the selected encoder features can be ensured.

[0077] The decoder trained by the GCDN method further enhances the model's detection capabilities in complex scenes and multi-target situations such as outdoor parking lots.

[0078] Through the OP-RTDETR model, the outdoor parking space intelligent detection system can realize the intelligent detection of outdoor parking, while achieving high real-time performance and improving the detection accuracy.

[0079] This embodiment also provides an outdoor parking space intelligent detection device, which includes: a controller; a memory storing executable instructions; wherein the executable instructions can be run on the controller and implement the outdoor parking space intelligent detection method.

[0080] This embodiment also provides a computer-readable storage medium storing a computer program, which implements the outdoor parking space intelligent detection method when the computer program is executed.

[0081] 1. The following is a performance comparison based on experiments conducted on the outdoor parking space intelligent detection method, system, equipment and medium:

[0082] This paper compares a variety of advanced object detection models, including OP-RTDETR, RTDETR, YOLOv8n, YOLOv8s, YOLOv9-S and YOLOv9-M, and evaluates their performance on different accuracy indicators and inference speed. RT-DETR and YOLO detectors share the same input size (640x640), while OP-RTDETR uses an input size suitable for VanillaNet (224x224). All models are trained and tested under the same experimental settings to ensure the comparability of the results. Specifically, the three key indicators of the number of parameters of each model, the average precision at different IoU thresholds (AP@IoU0.50:0.95 and AP@IoU0.50), and the number of frames per second (FPS) are compared. The YOLO detector uses the official pre-trained model based on the end-to-end speed benchmark.

[0083] From the experimental results, if Figure 5As shown in the figure, the YOLOv8 and YOLOv9 series models showed a good performance balance in this experiment. In particular, YOLOv8n achieved an AP@IoU0.50:0.95 of 0.969 and an inference speed of 207FPS with only 3.2M parameters, demonstrating its advantage in lightweight model design. YOLOv9-M performed well in both AP@IoU0.50 and AP@IoU0.50:0.95, reaching 0.995 and 0.993 respectively, and the FPS was 195, proving that it has achieved a good balance between accuracy and speed.

[0084] The OP-RTDETR model performs extremely well in both AP@IoU0.50:0.95 and AP@IoU0.50 accuracy indicators, outperforming RTDETR and other YOLO-like models in both speed and accuracy, especially the VanillaNet_5 version of OP-RTDETR. When VanillaNet_5 is used as the backbone, AP@IoU0.50:0.95 reaches 0.993, which is 0.3% higher than the GC-VanillaNet version. This is because in outdoor scenes, the boundaries of parking spaces may become blurred due to various factors (such as lighting, blurred ground signs, camera angles, etc.). Grouped convolution reduces the number of convolution kernels in each group, which may cause the model to lose its ability to capture these subtle features, thereby affecting the recognition accuracy. However, outdoor parking space recognition usually requires processing a large number of high-definition images, and real-time processing is often required in practical applications. Grouped convolution can improve the inference speed by reducing the amount of calculation, so that the model can achieve faster response time under limited computing resources and meet the needs of real-time processing. Therefore, when GC-VanillaNet is used as the backbone, the FPS is increased by 34, reaching a maximum of 405. In comparison, the two versions of RTDETR performed relatively poorly in terms of FPS, with the Resnet50 version at 233FPS and the Resnet101 version at only 167FPS. This result shows that although the increase in the number of parameters of Resnet101 (specifically 76M) has improved the accuracy of the model, it has also significantly reduced the inference speed. The two versions of RTDETR performed relatively weakly at AP@IoU0.50:0.95, 0.952 and 0.954 respectively. Taking into account the accuracy and inference speed of the model, the OP-RTDETR model showed the most balanced performance in this experiment, especially its GC-VanillaNet version, which has the advantages of high accuracy and efficient inference, and is very suitable for tasks that require real-time processing and high-precision detection.

[0085] 2. The following is the data obtained from the ablation experiment:

[0086] First, if Figure 6 As shown, the present invention compares the performance of RT-DETR when using different backbone networks (ResNet50, GC-VanillaNet and VanillaNet_5) and maintaining the same input size. By analyzing the AP@IoU0.50:0.95 and AP@IoU0.50 indicators, it is found that the RT-DETR model with VanillaNet_5 as the backbone network performs best in terms of accuracy, with AP@IoU0.50:0.95 reaching 96.1 and AP@IoU0.50 reaching 99.3, showing high accuracy and good inference speed. Since group convolution is used in GC-VanillaNet, the number of convolution kernels in each group is reduced, resulting in a slight decrease in accuracy compared to the original VanillaNet_5 model.

[0087] However, if Figure 7 As shown in the figure, the group convolution of GC-VanillaNet significantly improves the detection speed. GC-VanillaNet achieves an amazing speed of 405FPS during inference, which is much higher than VanillaNet_5, making it suitable for application scenarios with high requirements for real-time performance. In contrast, RT-DETR uses ResNet50 and ResNet101 as the backbone network and performs poorly in the FPS indicator.

[0088] Next, we analyze the effect of introducing the GCDN method in RT-DETR (with ResNet50 as the backbone). Figure 8 It can be seen that the introduction of GCDN significantly improves AP@IoU0.50:0.95 from 95.2 to 96.5, and AP@IoU0.50 from 99.0 to 99.2, which proves the effectiveness of GCDN in improving detection accuracy.

[0089] Finally, the overall performance of RT-DETR (ResNet50 without GCDN) and OP-RTDETR model combining GC-VanillaNet and GCDN, including speed and accuracy, is compared. Figure 8 As shown in the figure, OP-RTDETR outperforms RT-DETR (ResNet50 without GCDN) in both accuracy and inference speed. Specifically, OP-RTDETR (GC-VanillaNet+GCDN) achieves AP@IoU0.50:0.95 of 99.0, AP@IoU0.50 of 99.0, and inference speed of 405FPS, demonstrating the high efficiency of GC-VanillaNet and further verifying the contribution of GCDN in improving detection accuracy.

[0090] In summary, the GCDN method is crucial to improving the accuracy of the target detection model. Although the accuracy of GC-VanillaNet is slightly reduced, its significant speed advantage makes it very suitable for tasks that require efficient real-time processing.

[0091] In summary, the outdoor parking space intelligent detection method, system, device and medium of the present invention have the following beneficial effects:

[0092] 1. GC-VanillaNet feature extraction network based on VanillaNet: GC-VanillaNet uses group convolution technology to significantly improve the model's reasoning speed while maintaining feature extraction performance, thereby adapting to real-time detection needs.

[0093] 2. Efficient Hybrid Encoder structure based on RTDETR: The encoder consists of two parts: attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion, which significantly reduces the computational cost of the model and improves the efficiency of feature extraction.

[0094] 3. Uncertainty minimization query selection strategy: The decoder part adopts the uncertainty minimization query selection strategy of RT-DETR to optimize the quality of the initial query and improve the detection accuracy of the model in complex scenes.

[0095] 4. Group Contrast Noise Reduction (GCDN) Technology: The GCDN method performs noise reduction training at the group level, which significantly improves detection accuracy and training efficiency, making the model more stable and accurate when dealing with small targets and complex scenes.

[0096] 5. OP-RTDETR model for real-time detection of outdoor parking spaces: By integrating GC-VanillaNet, Efficient Hybrid Encoder and GCDN, the OP-RTDETR model achieves a balance between high precision and high real-time performance in outdoor parking space detection, and is particularly suitable for dynamic detection scenarios in urban open parking lots.

[0097] The technical features of the above-described embodiments may be arbitrarily combined. To make the description concise, not all possible combinations of the technical features in the above-described embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0098] The above-mentioned embodiments only express several implementation methods of the present invention, and the descriptions thereof are relatively specific and detailed, but they cannot be understood as limiting the scope of the invention patent. It should be pointed out that, for ordinary technicians in this field, several variations and improvements can be made without departing from the concept of the present invention, and these all belong to the protection scope of the present invention. Therefore, the protection scope of the patent of the present invention shall be subject to the attached claims.

Claims

1. An outdoor parking space intelligent detection method, characterized in that: The outdoor parking space intelligent detection method comprises the following steps: S1, build OP-RTDETR model; S2, realizing intelligent detection of outdoor parking through the OP-RTDETR model; Among them, in step S1, the specific method of constructing the OP-RTDETR model includes the following steps: S10, create GC-VanillaNet as the feature extraction network; S11, using the Efficient Hybrid Encoder of RTDETR as the encoder of the OP-RTDETR model; S12, optimizes Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR; S13, train the decoder using the GCDN method.

2. The outdoor parking space intelligent detection method according to claim 1, characterized in that: In step S10, the specific method of creating GC-VanillaNet as a feature extraction network includes the following steps: VanillaNet is modified based on the group convolution technology to obtain the GC-VanillaNet.

3. The outdoor parking space intelligent detection method according to claim 1, characterized in that: In step S11, the Efficient Hybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

4. The outdoor parking space intelligent detection method according to claim 1, characterized in that: In step S13, the specific method of training the decoder by the GCDN method includes the following steps: S130, upgrades DINO’s contrast noise reduction module to the group level; S131, perform independent self-attention calculation on the noise part and the object query in each group and perform one-to-one assignment to achieve group-level one-to-many matching.

5. An outdoor parking space intelligent detection system, characterized in that: The outdoor parking space intelligent detection system comprises: OP-RTDETR model for intelligent detection of outdoor parking; The OP-RTDETR model includes a GC-VanillaNet feature extraction network, an Efficient HybridEncoder, and a decoder trained by the GCDN method.

6. An outdoor parking space intelligent detection system as claimed in claim 5, characterized in that: The EfficientHybrid Encoder includes attention-based intra-scale feature interaction and CNN-based cross-scale feature fusion.

7. An outdoor parking space intelligent detection system as claimed in claim 5, characterized in that: The outdoor parking space intelligent detection system also includes: The uncertainty minimization query selection module is used to optimize the Bayesian uncertainty through the uncertainty minimization query selection strategy of RT-DETR.

8. An outdoor parking space intelligent detection device, characterized in that: The outdoor parking space intelligent detection device comprises: Controller; A memory storing executable instructions; The executable instructions may be run on the controller and implement the outdoor parking space intelligent detection method as described in any one of claims 1 to 4.

9. A computer-readable storage medium, characterized in that: The computer-readable storage medium stores a computer program, and when the computer program is executed, the outdoor parking space intelligent detection method according to any one of claims 1 to 4 is implemented.

Images