Elevator car transportation quality detection system and method

Abstract

The invention discloses an elevator car transportation quality detection system and method. The system comprises a microcontroller, an image recognition module, a data acquisition module and an alarm unit. The image recognition module comprises a camera used for collecting image information in a lift car, an improved YOLOv7 algorithm is deployed in the image recognition module, and the improvement of the YOLOv7 algorithm comprises the steps that a global attention mechanism module GAM is introduced into a head network of the YOLOv7 algorithm, and an ELAN module is replaced with a re-parameterized efficient feature extraction module RGN in a neck network; the data acquisition module comprises an acceleration sensor, a vibration sensor and a noise sensor, and the data acquisition module is connected with the micro controller. According to the invention, the acceleration, vibration and noise information of the elevator car and whether an electric vehicle enters or not can be detected, so that the detection of the elevator transportation quality is realized.

Description

Technical Field

[0001] This application relates to the field of elevator safety technology, and in particular to an elevator car passenger transport quality detection system and method. Background Technology

[0002] Elevators have gradually become an indispensable means of transportation in people's daily lives. The quality of acceleration, vibration, and noise within the elevator car affects passenger comfort, and electric vehicles entering the car pose significant safety hazards. Electric vehicles entering the car can easily collide with the car walls and landing doors, causing damage and deformation, or even derailment of the landing doors, significantly shortening the elevator's lifespan. Furthermore, if a battery catches fire, the resulting high temperatures and toxic fumes can quickly fill the enclosed car, easily causing passenger injuries or fatalities, posing a major safety risk. Article 37 of the "Regulations on Fire Safety Management of High-Rise Civil Buildings" explicitly prohibits the parking or charging of electric bicycles in public lobbies, evacuation corridors, stairwells, and safety exits of high-rise civil buildings. However, due to inadequate management and insufficient infrastructure, incidents of electric vehicles entering elevator cars persist despite these prohibitions.

[0003] According to GB10058-2009 "Technical Conditions for Elevators", the maximum starting acceleration and braking deceleration of passenger elevators should not exceed 1.5 m / s². 2 When the rated speed of the passenger elevator is 1.0 m / s < v ≤ 2.0 m / s, the acceleration and deceleration of A95 should not be less than 0.50 m / s. 2 When the rated speed of the passenger elevator is 2.0 m / s < v ≤ 6 m / s, the acceleration and deceleration of A95 should not be less than 0.70 m / s². 2 The maximum peak-to-peak value of the vertical (Z-axis) vibration of the passenger elevator car during constant acceleration should not exceed 0.30 / s². 2 The peak value of A95 should not exceed 0.20 m / s. 2 The maximum peak-to-peak value of horizontal (X-axis and Y-axis) vibration during passenger elevator car operation should not exceed 0.20 / s. 2 The peak value of A95 should not exceed 0.15 m / s. 2 Noise levels should meet the following requirements:

[0004]

[0005] Currently, existing technologies lack a solution that can simultaneously detect acceleration, vibration, noise inside the elevator car and monitor the entry of electric vehicles. Using a combination of multiple solutions to complete the detection of elevator ride quality leads to problems such as deployment difficulties and high deployment costs. Summary of the Invention

[0006] This application provides an elevator car ride quality detection system and method, which has the advantage of being able to detect the elevator car's acceleration, vibration, noise information, and whether an electric vehicle has entered, thereby realizing the detection of elevator ride quality.

[0007] The technical solution of this application is as follows:

[0008] On the one hand, this application provides an elevator car ride quality detection system, including:

[0009] microcontroller;

[0010] The image recognition module includes a camera for acquiring image information inside the car. The image recognition module deploys an improved YOLOv7 algorithm. The improvements to the YOLOv7 algorithm include: introducing a Global Attention (GAM) mechanism module into the head network of the YOLOv7 algorithm, and replacing the ELAN module in the neck network with a highly parameterized, efficient feature extraction module (RGN). The image recognition module acquires image information inside the car and identifies the electric vehicle in the image information using the improved YOLOv7 algorithm.

[0011] The data acquisition module includes an accelerometer, a vibration sensor, and a noise sensor, and is connected to a microcontroller.

[0012] And an alarm unit, which is connected to and controlled by the microcontroller, and is used to issue an alarm when the image recognition module identifies that the electric vehicle or the sensor data obtained by the data acquisition module exceeds a preset threshold.

[0013] Furthermore, the image recognition module is also equipped with the AUGMIX image enhancement algorithm, which is used to generate training samples based on the input image information.

[0014] Furthermore, the Global Attention Mechanism (GAM) module integrates the Channel Sub-Attention Module (MC) and the Spatial Sub-Attention Module (MS). For the input feature map F1, the channel permutation module in the Channel Sub-Attention Module (MC) uses a three-dimensional arrangement to retain information in three dimensions, and then uses a two-layer Multilayer Perception (MLP) mechanism to amplify the cross-dimensional channel-space information interaction capability to obtain the intermediate state F2. The Spatial Sub-Attention Module (MS) uses two convolutional layers to fuse spatial information, corrects the intermediate state F2, and obtains the output state F3 through the Sigmoid activation function.

[0015] The intermediate state F2 and the output state F3 are as follows:

[0016]

[0017] In the formula, MC represents the channel attention map, and MS represents the spatial attention map. Indicates cascading.

[0018] Furthermore, the efficient feature extraction module RGN generates and fuses different feature maps through reparameterization; during the training phase, the efficient feature extraction module RGN performs information fusion through addition operations, places ReLU after the Add operation, and finally adds BN operations to the branches.

[0019] Furthermore, it also includes a detection box, in which the microcontroller, image recognition module, data acquisition module and alarm unit are set inside or on the detection box; wherein, the side of the detection box is provided with a mounting base for fixing it inside the elevator car; the acceleration sensor and vibration sensor are set on the bottom surface inside the detection box; the noise sensor is set in the control box near one of its sides, and one of the control boxes is provided with a sound receiving hole corresponding to the position of the noise sensor.

[0020] Furthermore, it also includes a GPRS communication module and a remote monitoring platform. The GPRS communication module is located inside the control box, connected to the microcontroller, and used to communicate with the remote monitoring platform.

[0021] On the other hand, this application provides a method for detecting the quality of elevator car ride, including the following steps:

[0022] S1: Improvements to the YOLOv7 algorithm: Introduce the Global Attention (GAM) module in the head network of the YOLOv7 algorithm, and replace the ELAN module in the neck network with the highly efficient feature extraction module RGN with reparameterization; The improved YOLOv7 algorithm is used to identify electric vehicles in image information;

[0023] S2: Train the improved YOLOv7 algorithm;

[0024] S3: Collect image information inside the elevator car and use the trained model to identify electric vehicles in the image information.

[0025] In summary, the beneficial effects of this application are as follows:

[0026] 1. This application can detect the acceleration, vibration, and noise information of the elevator car, as well as whether an electric vehicle has entered, thereby realizing the detection of elevator ride quality;

[0027] 2. This application integrates the detection of elevator car acceleration, vibration, and noise information, as well as the detection of whether an electric vehicle has entered, through a control box, making it convenient and quick to deploy in elevators;

[0028] 3. This application improves the accuracy of the YOLOv7 algorithm for electric vehicle recognition by making improvements to it. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the elevator car passenger transport quality detection system of this application;

[0030] Figure 2 This is a circuit diagram of an acceleration sensor in one embodiment of this application;

[0031] Figure 3 This is a vibration detection circuit diagram in one embodiment of this application;

[0032] Figure 4 This is a schematic diagram of a noise detection circuit in one embodiment of this application;

[0033] Figure 5 This is a data transmission circuit diagram of the SIM800C module in one embodiment of this application;

[0034] Figure 6 This is a schematic diagram of the connection circuit between the microcontroller and each module in one embodiment of this application;

[0035] Figure 7 This is a schematic diagram of the network structure of the improved YOLOv7 algorithm model in one embodiment of this application;

[0036] Figure 8 This is a comparison diagram of the RGN and GhostNet structures in one embodiment of this application;

[0037] Figure 9 This is a diagram of the GAM network structure in one embodiment of this application;

[0038] Figure 10 This is a schematic diagram of a portion of the dataset in one embodiment of this application;

[0039] Figure 11 This is a visualized convergence curve of the improved YOLOv7 algorithm in one embodiment of this application;

[0040] Figure 12 This is a comparison chart of the PR curves of YOLOv7 before and after the improvement in one embodiment of this application. The value shown in the chart is the mAP@0.5 value before the improvement.

[0041] Figure 13 This is a comparison chart of the PR curves of YOLOv7 before and after the improvement in one embodiment of this application. The value shown in the chart is the mAP@0.5 value after the improvement.

[0042] Figure 14 This is a test result diagram of an electric vehicle according to an embodiment of this application;

[0043] Figure 15 This is a schematic diagram of real-time detection data of elevator car operation in one embodiment of this application. Detailed Implementation

[0044] The specific embodiments of this application are described in detail below with reference to the accompanying drawings.

[0045] Example: An elevator car passenger transport quality detection system, such as Figure 1 As shown, it includes a microcontroller, an image recognition module, a data acquisition module, and an alarm unit.

[0046] It also includes a detection box, in which the microcontroller, image recognition module, data acquisition module and alarm unit are set inside or on the detection box; wherein, the side of the detection box is provided with a mounting base for fixing it inside the elevator car; the acceleration sensor and vibration sensor are set on the bottom surface inside the detection box; the noise sensor is set in the control box near one of its sides, and one of the control boxes is provided with a sound receiving hole corresponding to the position of the noise sensor.

[0047] It also includes a GPRS communication module and a remote monitoring platform. The GPRS communication module is located inside the control box, connected to the microcontroller, and used to communicate with the remote monitoring platform.

[0048] The data acquisition module includes an accelerometer, a vibration sensor, and a noise sensor, and is connected to a microcontroller.

[0049] The accelerometer uses the MPU6050 module, which integrates a 3-axis gyroscope and a 3-axis accelerometer, capable of detecting the angular velocity of an object around three coordinate axes and the acceleration along three axes. The MPU6050 also integrates a Digital Motion Processor (DMP), which features a hardware acceleration engine and can output 6-axis or 9-axis attitude calculation data. Furthermore, the MPU6050 includes two I2C interfaces. The first I2C interface serves as the main interface for transmitting data to the microcontroller; the second I2C interface is used to connect third-party digital sensors, such as external magnetometers. By connecting external sensors, a complete 9-axis signal can be output, thereby acquiring more comprehensive motion data.

[0050] The circuit diagram of the accelerometer is attached. Figure 2As shown, pin 1 (VCC) of the MPU6050 module is the power supply pin, with an input voltage range of 3.3-5.5V; pin 3 (SCL) is the serial clock line, used to synchronize data transmission between the MPU6050 and the main controller or other digital ICs, ensuring data transmission under the correct clock signal; pin 4 (SDA) is the serial data line, used for data communication between the module and the main controller, allowing the MPU6050 to send accelerometer and gyroscope data to the main controller; pin 6 (INT) is the interrupt pin, which outputs a corresponding level change when a specific event occurs in the MPU6050 (such as data readiness or exceeding a threshold) to notify the main controller for processing. In addition, the MPU6050 module also has pins related to the three-axis gyroscope and three-axis accelerometer for outputting measurement data for each axis.

[0051] The vibration sensor used is the TW-123 vibration sensor, which integrates the sensor, power management, and other necessary components within the module. The TW-123 vibration sensor module operates at 3.3-5V; this embodiment uses a 3.3V power supply. The TW-123 module communicates with the host computer using the I2C communication protocol. Therefore, the module's SCL (clock line) and SDA (data line) pins need to be connected to the corresponding I2C interfaces on the host computer. Ensure correct pin alignment during connection and adherence to I2C communication specifications. The vibration detection circuit is shown in the attached diagram. Figure 3 As shown.

[0052] The noise sensor uses the SN-ZS-BZ noise detection module, which employs a high-performance pre-polarized back-electret electret condenser microphone, characterized by a wide dynamic range and stable performance. When the module detects ambient noise, its electrical signal changes. This signal is then calibrated through a signal conditioning sampling circuit to output the corresponding noise value. The SN-ZS-BZ module internally uses an amplification and acquisition circuit to output the calculated noise value via a serial port, making noise measurement faster and more convenient. A schematic diagram of the noise detection circuit is attached. Figure 4 As shown. Ground pin 1 of the noise detection module; pin 2 is the analog signal output terminal, connected to a general I / O port; pin 5 (VCC) is connected to 5V. Since the serial port signal output by this module is 5V, connect the serial port pin of the module to serial port 3 of the microcontroller (serial port 3 of the STM32 microcontroller supports 5V level).

[0053] The GPRS communication module selected is the SIM800C module. This module has rich built-in network protocols and achieves information exchange with the main control chip through a serial port. While possessing excellent performance, this module also features simple operation and low cost, and is now widely used in industrial production and daily life. The module's RXD and TXD are serial communication pins, and VIN is the power supply pin. The core board's power supply voltage is 5V-16V, and the input voltage must be ensured not to be lower than 5V. To prevent voltage drops from causing the core board to restart, when using a 5V power supply, the input current must not be less than 2A. The SIM800C module's data transmission circuit is shown in the attached diagram. Figure 5 As shown. Pin 1 of the SIM800C module is the power supply pin, connected to 5V; pin 5 is grounded; pin 2 is the serial port selection pin, connected to 3.3V to set the communication level to 3.3V; pins 3 and 4 are the serial port transmit / receive pins, connecting TX and RX to the microcontroller's serial port 2.

[0054] The image recognition module includes a camera for capturing images inside the elevator car. The module incorporates an improved YOLOv7 algorithm to identify electric vehicles within the images. The image recognition module utilizes the K210 chip, which integrates a low-power, high-performance AI chip designed specifically for edge computing. By combining the K210 chip with other necessary hardware, the K210 image recognition module can be easily embedded into various IoT devices to enable AI functionality.

[0055] The microcontroller uses an STM32F103C8T6 as the system's main control chip. This chip has three serial ports to meet the requirements of GPRS communication, noise detection, and K210 image data transmission. It also has eight timers and multiple I / O ports, which meet the design requirements of this system. The connection circuit between the microcontroller and each module is as follows: Figure 6 As shown.

[0056] The image recognition module is also equipped with the AUGMIX image enhancement algorithm, which is used to generate training samples based on the input image information.

[0057] The AUGMIX image enhancement algorithm increases the diversity of model training data by combining a variety of different image enhancement techniques to generate new training samples.

[0058] In the AUGMIZIX image augmentation algorithm, the model is trained not only on the original image but also on the augmented images. These augmented images are generated by applying a series of randomly selected image processing operations (called "augmentation operations") to the original image. These augmented images are then blended with the original image in a specific way to form the final training samples.

[0059] The key steps of AUGMIZIX can be summarized as follows:

[0060] 1. Select the original image x and perform k enhancement operations, each denoted as opk. These operations are specific changes to the image, such as rotation, cropping, and color transformation. This technique improves the recognition rate of electric vehicles.

[0061] 2. Apply these operations to the original image x to generate a set of enhanced images {x1, x2, ..., xk}. Each enhanced image xi is obtained by applying the corresponding enhancement operation opi to the original image x, i.e., xi = opi(x).

[0062] 3. These enhanced images are linearly blended to generate the final training samples. This blending process can be expressed by the following formula:

[0063]

[0064] In this formula, (w0,w1,...,w k ) are mixed weights, which are randomly selected and satisfy .

[0065] The blended images will then serve as new training samples. In this way, AUGMIZIX can introduce a large number of image variations during training, thereby improving the model's generalization ability.

[0066] Improvements to the YOLOv7 algorithm include: First, addressing the issue of image information loss due to dim lighting and target occlusion in elevator cars, a Global Attention (GAM) module is introduced into the head network of the YOLOv7 algorithm. This aims to enhance the saliency of targets in elevator scenes, thereby improving detection accuracy. Second, in the neck network, the ELAN module is replaced with a highly parameterized and efficient feature extraction module, RGN. By effectively reusing feature maps and optimizing the cascaded arithmetic unit structure, resource consumption is reduced while decreasing computational complexity during model inference. The improved YOLOv7 algorithm model network structure is as follows: Figure 7 As shown, the image recognition module acquires image information inside the car and identifies the electric vehicle in the image information using an improved YOLOv7 algorithm.

[0067] The alarm unit is connected to and controlled by the microcontroller, and is used to issue an alarm when the image recognition module identifies that the electric vehicle or the sensor data obtained by the data acquisition module exceeds a preset threshold.

[0068] In current object detection, both DenseNet and GhostNet employ the Concat operation to generate an increasing number of feature maps for feature reuse. Although the Concat operation is a zero-parameter, zero-float operation that adds data across channels, it is still significantly less efficient than simple addition. Furthermore, the computational cost of the Concat operation on hardware devices is substantial. To improve hardware efficiency and reduce computational complexity during inference, a reparameterized, high-efficiency feature extraction module (RepGhostNeXt, RGN) is used to achieve implicit feature reuse. By reparameterizing and fusing different feature maps, the model's global information representation capability and interaction efficiency are enhanced. This not only maintains the high-precision detection performance required by YOLOv7 but also further reduces the model's computational complexity and number of parameters. A comparison of the RGN and GhostNet structures is shown below. Figure 8 As shown.

[0069] The RGN module only involves standard convolutional layers and the ReLU activation function, performing fusion in the weight space rather than the feature space, thus ensuring high hardware efficiency. A comparison of the RGN and GhostNet network structures is attached. Figure 8 As shown. During the training phase, RGN abandons the inefficient Concat operation and uses an Add operation for information fusion, placing ReLU after the Add operation and finally adding a BN operation in the branch. This not only avoids the need for additional convolutional layers but also allows the module to meet the structural reparameterization rules, accelerating fusion and inference speeds. During the inference phase, the RGN module only involves adding the feature maps generated by the two branches, thus saving memory resources.

[0070] In target recognition scenarios within elevator cars, target occlusion is a significant issue, often resulting in multiple targets being closely packed together, which increases the complexity of the detection task. While traditional attention mechanisms have contributed to preserving channel and spatial information, they still fall short in enhancing cross-dimensional interaction. Therefore, this study adopts a Global Attention Mechanism (GAM) module in the head network. By reducing information diffusion within the network and amplifying the expression of global interaction features, it highlights the characteristics of electric vehicles, bicycles, and pedestrians, preserving key information while filtering out background noise. This allows for effective focusing on target objects within these confined spaces, thereby improving detection accuracy and overall network performance. The network structure of the GAM attention module is shown in the attached figure. Figure 9 As shown.

[0071] The GAM module is a global attention mechanism that integrates the channel sub-attention module (MC) and the spatial sub-attention module (MS). Compared to common attention mechanisms such as CBAM, ABN, and TAM, the GAM module offers better robustness and deep learning performance. Specifically, for the input feature map F1, the channel permutation module in the MC module uses a three-dimensional arrangement to preserve information in all three dimensions. Then, a two-layer multilayer perceptron (MLP) is used to amplify the cross-dimensional channel-space information interaction capability, resulting in the intermediate state F2. The MS module uses two convolutional layers to fuse spatial information, corrects the intermediate state F2, and then uses the sigmoid activation function to obtain the output state F3. (The diagram shows the intermediate state F2 and the output state F3.)

[0072] The intermediate state F2 and the output state F3 are as follows:

[0073]

[0074]

[0075] In the formula, MC represents the channel attention map, and MS represents the spatial attention map. Indicates cascading.

[0076] Ultimately, experimental results demonstrate that adding the RGN module and GAM attention mechanism helps the model pay more attention to the characteristics of the elevator car environment, resulting in a significant improvement in model performance.

[0077] The experimental verification process and results are as follows:

[0078] (1) Experimental Environment

[0079] The operating system is: Ubuntu 20.04.4;

[0080] CPU: Intel(R) i9-13900K;

[0081] GPU: NVIDIA RTX A6000;

[0082] Deep learning framework: PyTorch 1.11.0;

[0083] Editing language: Python 3.8;

[0084] CUDA version: 11.3.0;

[0085] Optimizer: Adam;

[0086] Number of iterations (epochs): 200;

[0087] Training batch size: 8.

[0088] (2) Dataset creation

[0089] The elevator electric vehicle recognition dataset is a self-built image library collected from a residential community in Nanjing, containing 9345 images. The dataset is divided into three target categories: electric vehicles, bicycles, and passengers. It covers the presence of electric vehicles under different conditions inside the elevator, including but not limited to different models, sizes, placement positions, and postures of the vehicles, as well as variables such as lighting conditions and crowding levels within the elevator. The aim is to provide a comprehensive and rich data foundation for model training. A sample dataset is attached. Figure 10 As shown, the dataset was labeled using the labelimg tool, and then divided into training, validation, and test sets in a 6:2:2 ratio for experiments.

[0090] (3) Evaluation indicators

[0091] To evaluate the performance of the improved algorithm model in recognizing electric vehicles inside elevators, accuracy, recall, mean precision, and floating-point computation were selected as evaluation metrics.

[0092] Precision (P) represents the proportion of samples that were correctly predicted as positive out of all predicted samples.

[0093]

[0094] Recall (R) represents the proportion of samples that were correctly predicted as positive out of all positive samples.

[0095]

[0096] Average Precision (AP) is represented by the area under the PR curve, with recall on the x-axis and precision on the y-axis.

[0097]

[0098] Mean Average Precision (mAP) is divided into mAP0.5 and mAP0.5:0.95. The former refers to the mAP value when the IoU threshold is equal to 50%, while the latter refers to the average of 10 mAP values ​​when the IoU threshold is between 50%, 55%, and 95%.

[0099]

[0100] In the formula, TP is the number of positive samples that are correctly identified, FP is the number of negative samples that are falsely detected, FN is the number of positive samples that are missed, n represents the total number of categories in the dataset, and i represents the number of detections.

[0101] Floating-point operations (FLOPs) are used to measure the complexity of an algorithm by calculating its time complexity, and the unit is G.

[0102] (4) Experimental Results

[0103] ① Analysis of the results of the improved YOLOv7 algorithm

[0104] Experiments were conducted using the improved YOLOv7 algorithm on the dataset presented in this paper. Visualized convergence curves for different metrics of the algorithm are attached. Figure 11 As shown, the curves mainly involve the confidence loss, bounding box loss, and class loss of the validation set and training set, as well as the convergence curves of the following four performance metrics (P, R, mAP0.5, mAP0.5:0.95), where the horizontal axis represents the number of iterations (epochs).

[0105] As can be seen from the loss curves of the model on the training and validation sets, the loss values ​​of the proposed model are relatively low. Overall, the bounding box loss is around 0.004, and the class loss remains at 0.002. After 200 iterations, the classification loss curve approaches 0. The four performance metrics curves show that the proposed model converges very quickly across all metrics, with mAP 0.5 stabilizing at 89.9%; accuracy approaching 91.2%; and recall approaching 85.8%. To explore the specific improvements of the improved algorithm for detecting three types of objects in the dataset, experiments were conducted comparing the proposed algorithm with the original YOLOv7 model. The PR curves of these two algorithms on the test set are attached. Figure 12-13 As shown.

[0106] ② Ablation test

[0107] To verify the individual and combined application effects of the RGN and GAM modules in the elevator and electric vehicle recognition task, ablation experiments were conducted on the aforementioned dataset. The experimental results are shown in Table 1. "√" indicates that the method was used in the experiment, and "-" indicates that the method was not used. The experimental results show that the RGN and GAM modules can each improve the model's performance on the key performance indicators mAP0.5 and mAP0.5:0.95. In particular, the model performance is most outstanding when these two strategies are used in combination, specifically, mAP0.5 increases to 89.9% and mAP0.5:0.95 increases to 73.8%.

[0108] ③ Comparative Experiment

[0109] To verify the performance of the improved algorithm, it was compared with current mainstream detection models, including SSD, Faster R-CNN, and the YOLO series. The experimental results are shown in Table 2. The results show that, in terms of accuracy, the improved algorithm model outperforms all the comparison models on mAP 0.5. Although it is slightly inferior to the YOLOv5L model in the more stringent mAP 0.5:0.95 evaluation, the new model performs outstandingly on several key metrics, highlighting its advantages in handling complex detection tasks. Furthermore, the improved model also performs excellently in recall and precision evaluations, reaching 85.8% and 91.2% respectively, demonstrating its high efficiency in identifying positive samples and reducing false positives. In terms of lightweight design, the model complexity is 97.4G, the lowest among all comparison models, and 5.8G lower than the YOLOv7 model, demonstrating sufficient lightweight advantage.

[0110]

[0111] ③ Comparative Experiment

[0112] To verify the performance of the improved algorithm, it was compared with current mainstream detection models, including SSD, Faster R-CNN, and the YOLO series. The experimental results are shown in Table 2. The results show that, in terms of accuracy, the improved algorithm model outperforms all the comparison models on mAP 0.5. Although it is slightly inferior to the YOLOv5L model in the more stringent mAP 0.5:0.95 evaluation, the new model performs outstandingly on several key metrics, highlighting its advantages in handling complex detection tasks. Furthermore, the improved model also performs excellently in recall and precision evaluations, reaching 85.8% and 91.2% respectively, demonstrating its high efficiency in identifying positive samples and reducing false positives. In terms of lightweight design, the model complexity is 97.4G, the lowest among all comparison models, and 5.8G lower than the YOLOv7 model, demonstrating sufficient lightweight advantage.

[0113] Table 2 Comparison of different algorithms

[0114]

[0115] The improved algorithm achieved an mAP0.5 value of 89.9%, an electric vehicle recognition accuracy of 97.9%, and reduced model complexity to 97.4 Gbps. The actual detection results of the improved algorithm are shown in the figure below. Figure 13 As shown. From the appendix Figure 13 As can be seen, the improved YOLOv7 exhibits higher detection accuracy.

[0116] The detection of elevator car acceleration, vibration, and noise information was conducted using an elevator in a residential community in Nanjing. Some test results are attached. Figure 15 As shown.

[0117] Example 2: A method for detecting the quality of elevator car ride, comprising the following steps:

[0118] S1: Improvements to the YOLOv7 algorithm: Introduce the Global Attention (GAM) module in the head network of the YOLOv7 algorithm, and replace the ELAN module in the neck network with the highly efficient feature extraction module RGN with reparameterization; The improved YOLOv7 algorithm is used to identify electric vehicles in image information;

[0119] S2: Train the improved YOLOv7 algorithm;

[0120] S3: Collect image information inside the elevator car and use the trained model to identify electric vehicles in the image information.

[0121] The above description is only a preferred embodiment of this application. It should be noted that for those skilled in the art, several modifications and improvements can be made without departing from the inventive concept of this application, and these all fall within the protection scope of this application.

Claims

1. An elevator car passenger transport quality detection system, characterized in that, include: microcontroller; The image recognition module includes a camera for acquiring image information inside the car. The image recognition module deploys an improved YOLOv7 algorithm. The improvements to the YOLOv7 algorithm include: introducing a Global Attention (GAM) mechanism module into the head network of the YOLOv7 algorithm, and replacing the ELAN module in the neck network with a highly parameterized, efficient feature extraction module (RGN). The image recognition module acquires image information inside the car and identifies the electric vehicle in the image information using the improved YOLOv7 algorithm. The data acquisition module includes an accelerometer, a vibration sensor, and a noise sensor, and is connected to a microcontroller. And an alarm unit, which is connected to and controlled by the microcontroller, and is used to issue an alarm when the image recognition module identifies that the electric vehicle or the sensor data obtained by the data acquisition module exceeds a preset threshold.

2. The elevator car passenger transport quality detection system according to claim 1, characterized in that, The image recognition module is also equipped with the AUGMIX image enhancement algorithm, which is used to generate training samples based on the input image information.

3. The elevator car passenger transport quality detection system according to claim 1, characterized in that, The Global Attention Mechanism (GAM) module integrates the Channel Sub-Attention Module (MC) and the Spatial Sub-Attention Module (MS). For the input feature map F1, the channel permutation module in the Channel Sub-Attention Module (MC) uses a three-dimensional arrangement to retain information in three dimensions. Then, a two-layer Multilayer Perception (MLP) mechanism is used to amplify the cross-dimensional channel-space information interaction capability to obtain the intermediate state F2. The Spatial Sub-Attention Module (MS) uses two convolutional layers to fuse spatial information, corrects the intermediate state F2, and obtains the output state F3 through the Sigmoid activation function. The intermediate state F2 and the output state F3 are as follows: In the formula, MC represents the channel attention map, and MS represents the spatial attention map. Indicates cascading.

4. The elevator car passenger transport quality detection system according to claim 1, characterized in that, The efficient feature extraction module RGN generates and fuses different feature maps through reparameterization; during the training phase, the efficient feature extraction module RGN performs information fusion through addition operations, places ReLU after the Add operation, and finally adds BN operations to the branches.

5. The elevator car passenger transport quality detection system according to claim 1, characterized in that, It also includes a detection box, in which the microcontroller, image recognition module, data acquisition module and alarm unit are set inside or on the detection box; wherein, the side of the detection box is provided with a mounting base for fixing it inside the elevator car; the acceleration sensor and vibration sensor are set on the bottom surface inside the detection box; the noise sensor is set in the control box near one of its sides, and one of the control boxes is provided with a sound receiving hole corresponding to the position of the noise sensor.

6. The elevator car passenger transport quality detection system according to claim 1, characterized in that, It also includes a GPRS communication module and a remote monitoring platform. The GPRS communication module is located inside the control box, connected to the microcontroller, and used to communicate with the remote monitoring platform.

7. A method for detecting the quality of elevator car ride, characterized in that, Includes the following steps: S1: Improvements to the YOLOv7 algorithm: Introduce the Global Attention (GAM) module in the head network of the YOLOv7 algorithm, and replace the ELAN module in the neck network with the highly efficient feature extraction module RGN with reparameterization; The improved YOLOv7 algorithm is used to identify electric vehicles in image information; S2: Train the improved YOLOv7 algorithm; S3: Collect image information inside the elevator car and use the trained model to identify electric vehicles in the image information.

Images