Automatic hoisting identification detection method for hook arm type garbage truck

Abstract

The invention discloses a hook arm type garbage truck automatic hoisting identification detection method, and relates to the technical field of image detection, and the method comprises the steps: expanding an image sample in the hook arm type garbage truck automatic hoisting operation process based on an SD stable diffusion AI large model, and obtaining a data set; finely tuning and training the deep learning model by adopting the expanded data set; and the hook arm type garbage truck utilizes the obtained deep learning model to carry out automatic detection and identification on images collected in the hoisting operation process. According to the method, high-quality data set expansion is carried out on training samples through an SDM stable diffusion large model and technologies such as graph generation, text generation and the like, and the problem of insufficient image learning samples is solved, so that the intelligent hoisting recognition detection model is rapidly trained, and the generalization ability and the reasoning accuracy are effectively improved. And in addition, the YOLO neural network is built in a hook arm type garbage truck controller after parameter fine adjustment is carried out on the YOLO neural network, recognition and detection according to collected images in the hoisting operation are achieved, and the industrial detection level requirement of unmanned hoisting operation is met.

Description

Technical Field

[0001] The present invention relates to the technical field of image detection, in particular to an automatic hoisting identification detection method for a hook-arm type garbage truck. Background Art

[0002] There are numerous fixed collection points for domestic waste in cities. Hook-arm garbage trucks travel to these collection points, pull out removable waste compactors (which can weigh several tons of waste), hoist them onto the rear frame, and then transport them. Traditional hook-arm garbage trucks require extensive manual operation when hoisting removable waste compactors. Typically, the driver, relying on experience, parks the vehicle, manually activates and adjusts the hoisting mechanism (using a remote control with buttons and handles located in the cabin). Garbage collection point maintenance personnel then verify the correct alignment of the vehicle hook and the removable waste compactor's lifting rings. The driver then drives the vehicle and hook arm, making various adjustments to ensure the hook engages the lifting rings, allowing the removable waste compactor to be pulled onto the truck. Garbage collection point maintenance personnel also monitor the centering (or misalignment) of the waste compactor with the two guide wheels on the rear frame, ensuring the compactor remains centered, stable, and free of deviation during the hoisting process. If any tilt or deviation is detected, the personnel immediately notify the driver and adjust the vehicle and hook arm.

[0003] This method of operation presents problems: the entire lifting process requires an experienced driver to operate the vehicle and lifting device. Furthermore, garbage point maintenance personnel must closely monitor the hooks and eyelets for alignment and engagement. They must also closely monitor the compactor to ensure it is not tilting or colliding with the vehicle while being towed onto the rear frame rails. After loading, they must ensure the bottom of the compactor is fully aligned with the rear frame rails. Furthermore, garbage point maintenance personnel must communicate with the driver in real time, prompting him to continuously adjust the position of the vehicle and lifting device. This results in very low efficiency and a low success rate. After the removable compactor is hoisted onto the vehicle, the driver may make mistakes, such as failing to secure it to the rear frame or retracting the rear legs. Therefore, garbage point maintenance personnel must assist the driver in checking the locking and restoration status to avoid dangers such as the compactor falling.

[0004] In order to replace the manual inspection, coordination, and manual control in existing lifting operations with automatic lifting operations, and solve the problems of laboriousness, tediousness, and low efficiency, it is necessary to replace it with automatic image recognition, identify and detect key points in the lifting operation process, and thus drive and control the hook-arm garbage truck. However, whether the lifting operation area is safe, that is, there are no pedestrians and other dynamic moving obstacles, the hook of the hook-arm truck is aligned and hooked with the lifting ring of the garbage compactor, the hook is firm during the lifting process, and the garbage compactor maintains balance and stability, the bottom of the garbage compactor is fully aligned with the rear frame slide rail, and the rear legs of the vehicle are extended / retracted. There are not enough image learning sample training sets on the market for these key points of visual inspection judgment. There is also no ready-made large model that can meet the industrial inspection level requirements of automatic lifting operations. Summary of the Invention

[0005] The purpose of the present invention is to provide a method for automatic lifting identification and detection of hook-arm garbage trucks, which is used to solve the problem in the prior art that there are not enough image learning samples for training automatic lifting operations, and there are no ready-made large models that can meet the industrial detection level requirements of automatic lifting operations, and cannot realize hook-arm automatic lifting operations.

[0006] The present invention solves the above problems through the following technical solutions:

[0007] A method for automatically lifting, identifying, and detecting a hook-arm type garbage truck, comprising:

[0008] Step S1: Expand the image samples of the automatic lifting operation process of the hook-arm garbage truck based on the SD stable diffusion AI large model to obtain a data set;

[0009] Step S2: fine-tune the deep learning model using the expanded data set;

[0010] Step S3: The hook-arm garbage truck uses the obtained deep learning model to automatically detect and identify images collected during the lifting operation.

[0011] Furthermore, in step S1, the AI large model based on SD stable diffusion includes a text image encoder, a noise intensity variance regulator, an image information creator, and a VAE model image decoder, wherein:

[0012] Text-image encoder, used to encode the input visual image and user prompt text, and output image embedding and text embedding;

[0013] The noise intensity variance regulator is used to initialize the latent variable using the noise random probability based on the Gaussian distribution of the input, adjust the noise intensity according to the proportional variance scaling algorithm, predict the noise residual according to the time step, and update the noise intensity level latent variable;

[0014] An image information creator, configured to use a scheduler to iteratively schedule a computation on a U-NET model with an attention mechanism to generate a predicted noisy image tensor; the input of the U-NET model with the attention mechanism is the updated latent variable and the image embedding and text embedding;

[0015] The VAE model image decoder is used to reconstruct the predicted noise image tensor to generate a target image, i.e., a data set.

[0016] Furthermore, the text image encoder includes a residual network encoder and a conversion encoder, wherein the residual network encoder is used to receive an input visual image and output an image embedding, and the conversion encoder is used to receive an input user prompt word text and output a text embedding.

[0017] Furthermore, the U-NET model includes a downsampling path encoder, an upsampling path decoder, and a bottleneck layer, wherein:

[0018] The downsampling path encoder consists of four downsampling stages, each of which contains: two residual network blocks for feature extraction and temporal conditioning, a spatial deformer based on the attention mechanism for introducing context conditioning and spatial attention, and a downsampling block for reducing the resolution through strided convolution or pooling;

[0019] The upsampling path decoder consists of four upsampling stages, each of which contains: two residual network blocks for recovering details in combination with residual connections, an attention-based spatial deformer for fusing conditional information, and an upsampling block for improving resolution through interpolation or transposed convolution;

[0020] The bottleneck layer is used to connect the middle part of the encoder and decoder. It contains two residual network blocks for further extracting global features and an attention-based spatial deformer for enhancing long-range dependency modeling.

[0021] In the U-NET model, an attention module is added after each residual network block that converts latent space data. The input token embedding is incorporated into each processing stage and then sent to the next residual network block, forming a serial structure. A portion of the input of each attention module is directly sent to the final stage of the processing process in the form of an attention module.

[0022] Furthermore, the residual network block includes: a normalization layer for prompting training stability, an activation function SILU for enhancing nonlinearity, a 3×3 convolutional layer for processing spatial features, a time step embedding injection for integrating temporal information into features through a fully connected layer, and a residual connection for retaining input features and alleviating gradient disappearance.

[0023] Furthermore, the core mechanism of the attention-based spatial deformer includes:

[0024] Self-attention, modeling the relationship between regions within the image;

[0025] Cross-Attention, injecting text conditions (CLIP embedding) into the image generation process;

[0026] Position encoding, which preserves spatial position information through sinusoidal encoding.

[0027] Furthermore, the VAE model image decoder is used to gradually reconstruct an image from the latent variable as an output through upsampling. The VAE model image decoder includes a latent space input layer and a convolutional upsampling path decoder. The convolutional upsampling path decoder includes, in networking order, 1 convolution module, 1 intermediate module including 1 residual network block + 1 self-attention module + 1 residual network block, 3 interpolation modules including, in networking order, 1 for upsampling the image and amplifying the spatial resolution of the image, 1 upsampling block of the convolution layer for performing a convolution operation on the upsampled image and restoring image details, 1 residual network block for superimposing the original input features after several layers of convolution and activation processing by introducing jump connections, and 1 GSC module combination. The GSC module combination is group normalization + activation function Swish + convolution layer.

[0028] Furthermore, the step S2 specifically includes:

[0029] Define category labels;

[0030] The operation process defines small target high-precision positioning, medium target symmetry detection, large target contact surface judgment, and binary abnormality detection. The small target high-precision positioning includes hook and ring alignment; the medium target symmetry detection includes guide wheel centering status; the large target contact surface judgment includes box and slide rail fit detection; the binary abnormality detection includes rear leg status classification.

[0031] Label the data set and generate a label file in YOLO format;

[0032] Divide the dataset into training set, validation set and test set;

[0033] Use the training set to fine-tune the parameters of the pre-trained model (such as the YOLO-v8 ~ YOLO-v11 model), including:

[0034] A1, Backbone model freezes the shallow cross-stage local network CSP stage;

[0035] A2,Backbone model unfreezes the deep C2PSA stage, which adjusts high-level semantic features to identify the structural associations of mechanical parts;

[0036] A3. Neck model unfreezes all BiFPN-XL layers to enhance multi-scale feature fusion capabilities and optimize the collaborative detection of small and large objects;

[0037] A4. In the Head model, the Detect detector classification branch parameters Conv2d and CLSLoss are unfrozen and trained from scratch to adapt to the new category; the initial layer of the regression branch is frozen to retain the pre-trained positioning ability and fine-tune the local parameters to adapt to the precise offset of the mechanical parts;

[0038] Use the training set to train and optimize the pre-trained model, including:

[0039] B1. Optimize the sample enhancement configuration in the training set, including one or more of enhancing hue changes, increasing saturation, controlling brightness, increasing rotation range, and enabling MixUp enhancement;

[0040] B2. For the scene where the hook is aligned with the ring and hooked, calculate the loss function of the distance between the two bounding boxes in the target monitoring CLoU :

[0041]

[0042] Among them, CLoU Used to reflect the similarity between two bounding boxes; IoU is the intersection over union ratio between two bounding boxes, ρ 2 (b pred ,b gt ) represents the square of the center distance between the predicted box and the ground truth GT box, c is the diagonal length of the minimum envelope of the two boxes, α is the adjustment coefficient, v represents the aspect ratio of the bounding box, w is the width of the image, h is the height of the image, img_size is the image size, ω small is the small target weight coefficient;

[0043] Dynamic calculation: ω small =1.0+2*(1-(wh) / img_size);

[0044] Fixed setting: For targets with an area smaller than 32*32, ω smallAssign 2.0 to 3.0 times weight;

[0045] B3. For the case of unbalanced samples of vehicle rear outrigger status, the binary cross entropy loss function ￡ CLS :

[0046]

[0047] Among them, CLS represents the binary cross entropy loss function, α dynamic Represents the weight coefficient, p t represents the predicted probability, γ adaptive is an adaptive adjustment factor, epoch represents the number of cycles of model training, and total_epoch represents the total number of cycles of model training;

[0048] B4. Use Tensor RT deployment to convert the fine-tuned pre-trained model to FP16 precision, optimize the computational graph, and obtain a deep learning model.

[0049] B5. Use the validation set and test set to evaluate the deep learning model of B4 respectively. After the evaluation is passed, the deep learning model will be deployed to the edge computing MEC system of the hook-arm garbage truck, the vehicle-side lifting operation domain controller LODC, and the vehicle-side automatic parking domain controller APDC respectively.

[0050] Furthermore, the step S3 specifically includes:

[0051] The hook-arm garbage truck uses the obtained deep learning model to automatically detect and identify images collected during the lifting operation.

[0052] The deep learning model built into the hook-arm garbage truck receives input lifting operation video data, performs image detection and recognition, and feeds back the real-time detection results to the vehicle control unit VCU of the hook-arm garbage truck in real time, realizing unmanned and fully automatic lifting operations.

[0053] Furthermore, the hook-arm garbage truck includes: an image detection module, an intelligent central computing gateway ICG, and an automatic parking domain controller APDC connected to the intelligent central computing gateway ICG, a chassis system, a vehicle control unit VCU, a power system, a body domain controller BDC, an intelligent cockpit domain controller ICDC, and a lifting operation domain controller, wherein:

[0054] The image detection module is used to obtain video data of the hook-arm garbage truck's parking space and the surrounding environment of the lifting operation, and input it into the automatic parking domain controller APDC. It is also used to collect video data of the hook of the hook-arm garbage truck and the lifting ring of the garbage compactor, and input it into the lifting operation domain controller LODC.

[0055] The intelligent central computing gateway (ICG) is used to implement protocol stack encapsulation and data routing and forwarding between the automatic parking domain controller (APDC), chassis system, vehicle control unit (VCU), power system, body domain controller (BDC), intelligent cockpit domain controller (ICDC), and lifting operation domain controller (LODC);

[0056] The automatic parking domain controller (APDC) is used to recognize images input by the image detection module and communicate with the vehicle control unit (VCU). The VCU chassis system can then implement automatic parking, or the APDC can directly issue control commands to the chassis system to implement automatic parking. The APDC has a built-in AI deep learning neural network that trains and learns from massive amounts of images to detect, judge, and control the automatic parking process.

[0057] Power system, used to provide transmission control, speed control and power;

[0058] Vehicle control unit VCU, used to control the chassis system and power system;

[0059] Body domain controller (BDC), used to check and issue warnings on vehicle tire pressure status;

[0060] The lifting operation domain controller LODC is communicated with the image detection module and is used to judge whether the hook and the lifting ring are aligned and hooked, and whether the hook is firmly on the lifting ring based on the input video data; the lifting operation domain controller LODC has a built-in AI deep learning neural network, which trains and learns massive images to realize the detection, judgment and control of the automatic lifting operation process; it also connects the detection module and the electronic control module through a private CAN. The detection module is used to detect the hydraulic oil volume, hydraulic pipeline and quick connector of the hook-arm garbage truck, and the electronic control module is used to control the power take-off, rear support legs, main arm lock hook and hook arm extension of the hook-arm garbage truck, which is used to realize status detection before lifting operation and start lifting operation.

[0061] Furthermore, the image detection module includes a first wide-angle camera arranged at the hook of the hook-arm garbage truck and AVM surround-view cameras respectively arranged at the front, rear, left and right sides of the hook-arm garbage truck; it also includes a second wide-angle camera arranged on the top frame of the garbage room door, a third wide-angle camera on the left side of the middle of the line parking space and a fourth wide-angle camera on the right side of the middle of the line parking space. The AVM surround-view camera is used to obtain video data of the surrounding environment of the lifting operation, and the first wide-angle camera is used to collect video data of the hook and the lifting ring of the garbage compression box. The line parking space is a rectangular frame, and the side of the rectangular frame close to the garbage room is N meters away from the garbage room. The width of the rectangular frame = the inner width of the garbage room exit, and the length of the rectangular frame ≥ the length of the hook-arm garbage truck.

[0062] Furthermore, it also includes a field-side multi-access edge computing MEC system and an intelligent terminal. The field-side multi-access edge computing MEC system is communicatively connected to the second wide-angle camera, the third wide-angle camera, and the fourth wide-angle camera, and is used to process and judge the input video data of the hook-arm garbage truck parking space, and communicate with the automatic parking domain controller APDC and the lifting operation domain controller LODC through the intelligent central computing gateway ICG;

[0063] The smart terminal communicates with the intelligent central computing gateway ICG via Bluetooth or WiFi, and is used to start the lifting operation of the hook-arm garbage truck with one click.

[0064] Furthermore, the field-side multi-access edge computing MEC system includes:

[0065] an image acquisition and preprocessing module, configured to process in parallel the video data acquired by the second wide-angle camera, the third wide-angle camera, and the fourth wide-angle camera, and obtain high-quality image data frames through one or more algorithms selected from the group consisting of noise suppression, data standardization, color space transformation, and geometric distortion correction;

[0066] Environmental perception implementation reasoning module for target recognition and tracking based on input high-quality image data frames;

[0067] The logic calculation and judgment warning module is used to determine whether the hoisting operation area is safe based on the target identification and tracking results, and during the hoisting process, determine:

[0068] Whether the garbage compression box is balanced and stable, and whether there is any deviation from the center;

[0069] When the garbage compression box is fully hoisted onto the rear frame, is its bottom completely in contact with the rear frame slide rail?

[0070] Whether the vehicle's rear outriggers are deployed when the lifting operation is started and whether the vehicle's rear outriggers are restored and retracted when the lifting operation is completed;

[0071] The logic calculation and judgment warning module needs to use AI deep learning neural networks to train and learn massive images to achieve detection, judgment, and control of the safety of the lifting operation area and the lifting operation process;

[0072] The V2I communication module is used to actively broadcast the above judgment results to the intelligent central computing gateway ICG.

[0073] Furthermore, the chassis system includes an electronic parking brake EPB, an electronic mechanical brake EMB and an electronic power steering system EPS. The automatic parking domain controller APDC is communicated with the chassis system through a chassis domain CAN. The automatic parking domain controller APDC is also communicated with the electronic mechanical brake EMB and the electronic power steering system EPS through a private CAN.

[0074] Furthermore, the electromechanical brake EMB adopts a brake-by-wire dual-redundancy setting.

[0075] Furthermore, the intelligent cockpit domain controller (ICDC) is connected to the central control screen (CSD). The automatic parking domain controller (APDC) forwards the spliced image from the video data input by the image detection module to the intelligent cockpit domain controller (ICDC) via the in-vehicle Ethernet and the intelligent central computing gateway (ICG). The intelligent cockpit domain controller (ICDC) displays the image in real time on the central control screen (CSD), allowing the driver to set parameters and monitor the status of the lifting operation in real time from within the vehicle.

[0076] Furthermore, the power system includes a power domain module group, a rear motor controller integrated rotation MCRT and a rear transmission control unit TCUR. The rear motor controller integrated rotation MCRT and the rear transmission control unit TCUR are communicatively connected to the vehicle control unit VCU and the intelligent central computing gateway ICG through the power domain CAN-1; the power domain module group is communicatively connected to the vehicle control unit VCU and the intelligent central computing gateway ICG through the power domain CAN-2.

[0077] Furthermore, the body domain controller BDC is connected to the master tire pressure monitoring module TPMM and the slave tire pressure monitoring module TPMS via the body domain CAN communication.

[0078] Furthermore, the detection module includes a hydraulic oil quantity detection module, a hydraulic pipeline detection module and a quick connector detection module; the electronic control module includes a power take-off electronic control module, a rear support leg electronic control module, a main arm lock hook electronic control module and a hook arm telescopic electronic control module.

[0079] Environmental perception and recognition detection for intelligent lifting operations are implemented using deep learning models for training and reasoning.

[0080] (1) The deep learning model built into the vehicle-side automatic parking domain controller (APDC) (using the AVM surround view camera) needs to identify and detect scenarios including: linear parking spaces, pedestrians in the lifting operation area, and other dynamic obstacles;

[0081] (2) The deep learning model built into the vehicle-side lifting operation domain controller (LODC) (using a wide-angle camera at the vehicle hook) needs to identify and detect scenarios such as: whether the vehicle hook is aligned with the garbage compression box hook and whether it is secure;

[0082] (3) The deep learning model built into the yard-side MEC (using the camera on the top frame of the garbage room door and the cameras on the left and right sides of the middle of the yard-side parking space) requires auxiliary vehicle-side identification and detection scenarios including: pedestrians and other dynamic obstacles in the lifting operation area; among them: using the camera on the top frame of the garbage room door, the scenarios that need to be identified and detected include: during the lifting process, the garbage compression box and the middle position of the two guide wheels on the rear frame are aligned / deviated; when the lifting operation is started / completed, the vehicle's rear outriggers are deployed / restored; using the cameras on the left and right sides of the middle of the yard-side parking space, the scenarios that need to be identified and detected include: during the box loading process, the bottom of the garbage compression box is completely in contact with the rear frame slide rail; when the lifting operation is started / completed, the vehicle's rear outriggers are deployed / restored.

[0083] A method for automatically lifting a hook-arm garbage truck comprises: a field-side multi-access edge computing (MEC) system is communicatively connected to the image detection module to process and judge input video data of the hook-arm garbage truck's parking space, and communicates with the automatic parking domain controller (APDC) and the lifting operation domain controller (LODC) via an intelligent central computing gateway (ICG); the intelligent terminal communicates with the intelligent central computing gateway (ICG) via Bluetooth or WiFi to start the hook-arm garbage truck lifting operation with one click; the method comprises:

[0084] Step S1, automatic parking process, including:

[0085] A1. The smart terminal establishes a communication connection with the hook-arm garbage truck, and the automatic parking is activated with one button on the smart terminal.

[0086] A2, the automatic parking domain controller APDC performs line parking space detection and recognition, operating area surrounding environment perception, and obstacle recognition and positioning based on the video data input by the image detection unit;

[0087] A3, the automatic parking domain controller (APDC) generates the parking path, initiates parking, and performs collision detection, trajectory tracking control, and exception handling.

[0088] A4, the hook-arm garbage truck stops in the center of the parking space and completes automatic parking;

[0089] Step S2: The smart terminal starts the status check function before the lifting operation with one click, and the hook-arm garbage truck establishes a V2I short-range wireless connection with the multi-access edge computing MEC at the site; including:

[0090] (1) The lifting operation domain controller LODC determines whether the hydraulic oil volume, hydraulic pipeline and quick connector status reported by the detection module are normal;

[0091] (2) The lifting operation domain controller LODC determines whether the status reported by the electronic control module is normal;

[0092] (3) The body domain controller (BDC) determines whether the tire pressure data is normal;

[0093] (4) The automatic parking domain controller (APDC) confirms whether the electronic power steering system (EPS) and the electronic mechanical brake (EMB) of the chassis system are in normal condition;

[0094] (5) The automatic parking domain controller (APDC) confirms that there are no personnel or obstacles in the lifting operation area;

[0095] (6) Multi-access edge computing (MEC) at the site confirms that there are no personnel or obstacles in the lifting operation area;

[0096] If any of (1)-(4) is abnormal or there are people or obstacles in the lifting operation area, an alarm is issued and sent to the smart terminal; otherwise, the status detection before the lifting operation is completed;

[0097] Step S3, the hoisting operation process, includes:

[0098] B1. The intelligent terminal starts the unmanned lifting operation with one click. The operation command is forwarded to the vehicle control unit (VCU) and the lifting operation domain controller (LODC) via the ICG of the hook-arm garbage truck.

[0099] B2. The vehicle control unit (VCU) controls the electronic parking brake (EPB) of the chassis system to apply the vehicle's parking brake. The vehicle control unit (VCU) controls the rear transmission control unit (TCUR) of the power system to put the vehicle into neutral gear.

[0100] B3. The lifting operation domain controller LODC controls the rear outrigger electronic control module in the electronic control module to lower the vehicle's rear outriggers.

[0101] B4. The vehicle control unit (VCU) controls the power domain module group of the built-in lifting operation hydraulic system to start the vehicle engine and turn on the power of the hydraulic system. The vehicle control unit (VCU) controls the MCRT to provide power.

[0102] B5. The vehicle control unit (VCU) controls the rear transmission control unit (TCUR) to drive the synchronizer, while the lifting operation domain controller (LODC) controls the power take-off (PTO) electronic control module in the electronic control module. The two work together to complete the hydraulic pump power take-off action.

[0103] B6. The lifting operation domain controller LODC controls the main arm lock hook electric control module in the electric control module to control the main arm lock hook to be in a completely released state;

[0104] B7. The hoisting operation domain controller (LODC) controls the hook arm extension and retraction electronic control module in the electronic control module to continuously extend the main arm cylinder until the auxiliary arm hook flips over to a position slightly lower than the center of the garbage compression box hook.

[0105] B8: The vehicle control unit (VCU) controls the rear transmission control unit (TCUR) to drive the synchronizer, and the lifting operation domain controller (LODC) controls the power take-off (PTO) electronic control module in the electronic control module. The two work together to shut down the hydraulic pump PTO.

[0106] B9. The vehicle control unit (VCU) controls the rear transmission control unit (TCUR) to drive the synchronizer, shifting the vehicle into reverse gear. The VCU then controls the power system's rear motor controller to rotate the MCRT, causing the vehicle to reverse backward to the center of the hook and the trash compactor's lifting ring. The connection between the hook and the lifting ring is monitored in real time. Once the lifting operation domain controller (LODC) confirms a secure connection, the VCU controls the EMB to brake the vehicle.

[0107] B10. Hydraulic power take-off is performed again: The vehicle control unit (VCU) controls the rear transmission control unit (TCUR) to drive the synchronizer, and the hoisting operation domain controller (LODC) controls the power take-off electronic control module in the electronic control module. The two work together to complete the hydraulic pump power take-off action.

[0108] B11. The hoisting operation domain controller (LODC) controls the hook arm extension and retraction electronic control module, causing the main arm cylinder to continuously retract, pulling the garbage compactor onto the slide rails of the rear frame. The container is continuously lifted until it is completely suspended in the air and parallel to the frame, until it is completely pulled onto the rear frame.

[0109] B12: Monitor the centering between the box and the two guide wheels on the rear frame. Multi-access edge computing (MEC) on the field side performs real-time analysis and judgment. If the box is judged to have deviated from the centering position, an early warning message is immediately sent to notify the vehicle to adjust its position and angle.

[0110] B13. During the loading process, the bottom of the garbage compression box is monitored to ensure it is completely aligned with the rear frame rails. The multi-access edge computing (MEC) on the site performs real-time analysis and judgment, and sends a notification of alignment to the lifting operation domain controller (LODC).

[0111] B14. Receive a message about the fit status of the box and the rear frame rail. If the fit is complete, the hoisting operation domain controller (LODC) controls the main arm lock hook electronic control module to fully lock the main arm lock hook and proceed to the next step. If the message indicates that the fit is not complete, the vehicle needs to adjust its position and angle. If the message indicates that the fit is not complete multiple times, the vehicle will suspend the hoisting operation.

[0112] B15. After the vehicle lifting operation is completed normally, the vehicle control unit (VCU) controls the rear transmission control unit (TCUR) to drive the synchronizer, and the lifting operation domain controller (LODC) controls the power take-off (PTO) electronic control module. The two work together to shut down the hydraulic pump PTO and release the hydraulic pressure.

[0113] B16 is replaced by: the vehicle control unit VCU controls the power domain module group to turn off the power supply of the lifting operation hydraulic system; the lifting operation domain controller LODC controls the rear outrigger electronic control module to restore and retract the vehicle's rear outriggers, and sends a message to the smart terminal to complete the lifting operation.

[0114] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0115] (1) This invention uses the SDM stable diffusion model and, through techniques such as graph-to-graph and text-to-graph, expands the training sample data set with high quality, solving the problem of insufficient image learning samples. This allows the intelligent lifting recognition and detection model to be quickly trained, effectively improving generalization ability and reasoning accuracy. Furthermore, by fine-tuning the parameters of the YOLO neural network and integrating it into the controller of a hook-arm garbage truck, it enables recognition and detection based on collected images during lifting operations, meeting the industrial detection level requirements for unmanned lifting operations.

[0116] (2) The present invention designs the overall electronic and electrical architecture of the hook-arm garbage truck, adds an on-board monitoring camera, an automatic parking domain controller, and an intelligent lifting operation domain controller, and coordinates with the intelligent central computing gateway, the chassis power steering system, the braking system, etc., to complete the entire lifting operation process intelligently and automatically without the need for human intervention throughout the process.

[0117] (3) The present invention adds cameras for detection to confirm whether the lifting operation area is safe, that is, there are no pedestrians and other dynamic moving obstacles (vehicle-side AVM surround-view camera, field-side wide-angle camera (second wide-angle camera, third wide-angle camera and fourth wide-angle camera)), the hook of the hook arm truck is aligned with the lifting ring of the garbage compactor and hooked (first wide-angle camera), the hook is firm during the lifting process (first wide-angle camera) and the garbage compactor maintains balance and stability (field-side wide-angle camera), the bottom of the garbage compactor is completely fitted with the rear frame slide rail (field-side wide-angle camera), and the rear support legs of the vehicle are unfolded / retracted (field-side wide-angle camera). The key points of the above visual detection and judgment are trained through image deep learning neural networks and placed in the vehicle-side controller and the field-side controller respectively. During the lifting operation, the field-side controller assists the vehicle-side controller in real-time reasoning, identifies and detects the correctness of the operation, and realizes unmanned lifting operation.

[0118] (4) The present invention establishes a communication connection between the smart terminal, the hook-arm garbage truck, and the site-side MEC, realizes one-key triggering of automatic lifting operations, and automatically sends an early warning to the smart terminal.

[0119] (5) The entire process of lifting the detachable trash bin can be completed with high accuracy and efficiency through the coordination of sensors, controllers, actuators and other components on the vehicle, as well as sensors and controllers on the site, thus avoiding the traditional inefficient manual method (garbage point maintenance personnel observe whether the vehicle hook and the lifting ring of the detachable trash bin are accurately aligned, and let the driver drive the vehicle and hook arm device to make various posture adjustments); the entire lifting operation process is completely unmanned, which improves efficiency and avoids the risk of personal injury in the operation area. BRIEF DESCRIPTION OF THE DRAWINGS

[0120] Figure 1 This is a diagram of the electronic and electrical architecture of the hook-arm garbage truck of the present invention;

[0121] Figure 2 This is a schematic diagram of the communication interaction between the field-side multi-access edge computing MEC system, the hook-arm garbage truck, and the mobile phone APP;

[0122] Figure 3 Schematic diagram for generating high-quality datasets based on the SD stable diffusion AI large model expansion;

[0123] Figure 4 Schematic diagram of fine-tuning training based on the YOLO-v11 model;

[0124] Figure 5 for Figure 4 C2PSA module composition diagram;

[0125] Figure 6 for Figure 4 The structure diagram of the Detect module. DETAILED DESCRIPTION

[0126] The present invention will be further described in detail below with reference to the examples, but the embodiments of the present invention are not limited thereto.

[0127] Before introducing the embodiments of the present invention, the abbreviations used herein are explained as follows:

[0128] 5G-V2X: 5G Vehicle to Everything, communication between vehicles and everything;

[0129] AB: Attention Block, attention block;

[0130] AHD: Analog High Definition, an analog high-definition video transmission technology;

[0131] APDC: Auto Parking Domain Controller, automatic parking domain controller;

[0132] ASN.1: Abstract Syntax Notation One, an ISO / ITU-T standard;

[0133] AVM: Around View Monitor, panoramic imaging system;

[0134] AWS: Amazon Web Service, Amazon Cloud Technology;

[0135] BDC: Body Controller, body domain controller;

[0136] BiFPN-XL: is an extended version of the Bidirectional Feature Pyramid Network (BiFPN), designed to further optimize the feature fusion process;

[0137] C2PSA: Combines the CSP (Cross Stage Partial) structure and the PSA (Pyramid Squeeze Attention) attention mechanism to improve multi-scale feature extraction capabilities;

[0138] CAN: Controller Area Network, controller area network;

[0139] CAN-FD: Controller Area Network Flexible Data-Rate, controller area network with flexible data transmission rate;

[0140] CIoU: Combined Intersection over Union, combined intersection over union ratio;

[0141] CLIP: Constrastive Language-Image Pre-training, a pre-training method or model based on contrastive text-image pairs;

[0142] CLSLoss: Classification Loss, an important loss term in the target detection model, is used to measure the accuracy of the model's prediction of each target object category;

[0143] CNN: Convolutional Neural Network, convolutional neural network;

[0144] Conv2d: Convolution of 2Dimensions, two-dimensional convolution;

[0145] CSD: Central Screen Display, central control screen;

[0146] CSP: Cross Stage Partial Network, cross-stage partial network;

[0147] DWConv: Depthwise Separable Convolution, depthwise separable convolution;

[0148] EC2: Elastic Compute Cloud, elastic computing cloud;

[0149] EEA: Electrical / Electronic Architecture, electronic and electrical architecture;

[0150] EMB: Electro-Mechanical Brake, electronic mechanical brake;

[0151] EPB: Electrical Park Brake, electronic parking;

[0152] EPS: Electrical Power Steering, electronic power steering system;

[0153] Ethernet: In-vehicle Ethernet;

[0154] FP16: Half-Precision Floating-Point, half-precision floating point number;

[0155] FPS: Frames Per Second, frames per second;

[0156] GDPR: General Data Protection Regulation;

[0157] GPU: Graphics Processing Unit, graphics processor;

[0158] GSC: GroupNorm+Swish+Conv joint component, group normalization, Swish smooth activation function, convolution layer

[0159] GT: Ground Truth, ground truth;

[0160] HIPAA: Health Insurance Portability and Accountability Act;

[0161] HSV: Hue / Saturation / Value, hue / saturation / brightness, is a nonlinear model for representing color in color space

[0162] IAM: Identity and Access Management;

[0163] ICDC: Intelligent Cockpit Domain Controller;

[0164] ICG: Intelligent Central Gateway;

[0165] INF2: AWS Inferentia2, the first inference-optimized instance in Amazon EC2;

[0166] IPM: Inverse Perspective Mapping, inverse perspective transformation;

[0167] ITU-T: International Telecommunication Union-Telecommunication Standardization Sector, International Telecommunication Union Telecommunication Standardization Sector;

[0168] KMS: Key Management Service

[0169] LODC: Lifting Operation Domain Controller, lifting operation domain controller;

[0170] LN: Layer Normalization, layer normalization;

[0171] LVDS: Low-Voltage Differential Signaling, low voltage differential signal;

[0172] MCRT: Motor Control Unit Rear Transmission, rear motor controller integrated transmission;

[0173] MEC: Multi-access Edge Computing;

[0174] MLP: Multi-Layer Perceptron, multi-layer perceptron;

[0175] MPC: Model Predictive Control, model predictive control;

[0176] P4D: AWS P4D is an instance based on the AMD EPYC processor, with models such as P4d.24xlarge.

[0177] P5: AWS P5 is a high-performance computing instance based on the NVIDIA H100 Tensor Core GPU.

[0178] PSABlock: Pyramid Squeeze Attention Block, pyramid squeeze attention block;

[0179] RB: ResNet Block, residual network block;

[0180] ReLU: Rectified Linear Unit, linear rectification function;

[0181] RSM: Roadside Safety Message, roadside safety message;

[0182] ResNet:Residual Network, residual network;

[0183] RSU: Road Side Unit, road side unit;

[0184] S3: Amazon Simple Storage Service

[0185] SAE: Society of Automotive Engineers;

[0186] SDM: Stable Diffusion Model, stable diffusion model;

[0187] SILU: Sigmoid Linear Unit, also known as the Swish activation function, is a nonlinear function that multiplies the input by its own sigmoid value;

[0188] Swin Transformer: A computer vision model based on the transformer architecture, designed to address the high computational cost and low efficiency of traditional transformer applications in image processing.

[0189] SW-MSA: Shifted Window Multi-Head Self Attention, multi-head self-attention with window offset;

[0190] TBox: Telematic Box, an intelligent vehicle terminal in the Internet of Vehicles system;

[0191] TCUR: Transmission Control Unit Rear, rear transmission control unit;

[0192] TMC: Thermoplastic Composite Material, self-lubricating oil-containing nylon;

[0193] TPMM: Tire Pressure Monitor System Main, main tire pressure detection module;

[0194] TPMS: Tire Pressure Monitor System Slave, from the tire pressure detection module;

[0195] Transformer: Transformer is a sequence model based on attention mechanism;

[0196] U-NET: is a U-shaped architecture based on convolutional neural networks for image segmentation tasks;

[0197] V2I: Vehicle to Infrastructure, communication technology between vehicles and infrastructure;

[0198] VAE: Variational Autoencoder, variational autoencoder;

[0199] VCU: Vehicle Control Unit, vehicle control unit;

[0200] W-MSA: Window-Based Multi-Head Self Attention, windowed multi-head self-attention;

[0201] YOLO: You Only Look Once, an end-to-end object detection algorithm that can complete recognition in a single scan.

[0202] Example 1:

[0203] A method for automatic lifting detection and recognition of hook-arm garbage trucks first requires training the AI deep learning network built into the vehicle-side control and yard-side controllers based on visual image data of the industrial environment perception scenes involved in the lifting operation. Deep learning models trained based on a small data set are prone to overfitting, and in actual use, the recognition and detection accuracy is not high. This embodiment uses SD stable diffusion AI large model graph generation graph, text generation graph and other technologies to expand the training sample high-quality data set, greatly improving the reasoning accuracy of the vehicle-side and yard-side intelligent lifting recognition and detection models, that is, the built-in AI deep learning network.

[0204] like Figure 3 As shown, based on the SD stable diffusion AI large model, a high-quality data set is generated, including:

[0205] (1) The structure of the AI large model based on SD stable diffusion includes ClipText text image encoder, image information creator (Image Information) based on U-Net and 50 scheduler reconstruction, and VAE model image decoder;

[0206] (2) The visual image and user prompt word text are input to ClipText, the visual image is input to the ResNetEncoder residual network encoder, and the output image embedding is Embedding; the user prompt word text is input to the TransformerEncoder conversion encoder, and the output is 77×768 (77 token embeddings vectors, each token vector has 768 dimensions) text embedding Embedding;

[0207] (3) The SDM stable diffusion model image information creator consists of a U-NET model with an attention mechanism. The Scheduler scheduler is used to complete 50 iterations (running multiple steps) of the U-NET model to generate the predicted noise image tensor.

[0208] (4) U-NET is a conditional diffusion model that is responsible for gradually de-noising in the latent space to generate the target image. Its input is a noisy latent variable (64×64×4 tensor), timestep embedding (Timestep Embedding), and text conditions (ClipText text encoding); its output is the predicted noise residual (same dimension as the input); a U-NET contains 18 ResNet Blocks residual network blocks and 9 Attention blocks; Stable Diffusion U-NET achieves efficient latent space denoising through a fine combination of residual model modules and attention blocks. Its design balances computational efficiency and generation quality;

[0209] (5) The U-NET structure includes a downsampling path encoder, an upsampling path decoder, and a bottleneck layer Bottleneck; the encoder consists of 4 downsampling stages, each of which contains: 2 ResNet Blocks residual network blocks (responsible for feature extraction and temporal condition fusion), 1 Spatial Transformer (Attention) based on the attention mechanism of the spatial deformer (responsible for introducing text conditions and spatial attention), and 1 DownSample downsampling block (reducing resolution through stride convolution or pooling); the decoder consists of 4 upsampling stages, each of which contains: 2 ResNet Blocks residual network blocks (combined with residual connections to restore details), 1 Spatial Transformer (Attention) based on the attention mechanism of the spatial deformer (fusing condition information), and 1 UpSample upsampling block (increasing resolution through interpolation or transposed convolution); the bottleneck layer Bottleneck connects the middle part of the encoder and the decoder, which contains: 2 ResNet Blocks residual network blocks (used to further extract global features), 1 Spatial Transformer (Attention) is a spatial transformer based on the attention mechanism (enhanced long-range dependency modeling);

[0210] (6) The internal components of ResNet Block include: GroupNorm normalization layer to improve training stability; SILU activation to replace traditional ReLU and enhance nonlinearity; 3×3 convolution layer to process spatial features; time step embedding injection to integrate temporal information into features through the fully connected layer; residual connection to retain input features and alleviate gradient disappearance;

[0211] (7) The core mechanisms of Spatial Transformer (Attention) include: Self-Attention, which models the relationship between regions within the image; Cross-Attention, which injects textual conditions (CLIP embedding) into the image generation process; Position Encoding, which preserves spatial position information through sinusoidal encoding;

[0212] (8) In U-NET, an Attention module is added after each ResNet Block that converts latent space data. As a text conditioning mechanism, the input TokenEmbedding is integrated into each processing stage and then sent to the next ResNet Block, forming a serial structure. The input of each Attention module is not all handed over to the next module for processing, but a part of it is sent directly to the final stage of the processing process in the form of an Attention module.

[0213] (9) Latent Seed: Uses Gaussian-distributed noise random probability to initialize latent variables; uses a noise intensity regulator to adjust noise intensity through a proportional variance scaling algorithm, predicts noise residuals based on the timestep, and updates the noise amount embedding noise intensity level latent variable based on the scheduler;

[0214] (10) The VAE model image decoder gradually reconstructs the image (512×512) from the latent variables through upsampling as the output. Its composition structure includes: latent space input layer (64×64×4 tensor); convolution upsampling path decoder, which includes 1 CNN convolution module, 1 MidBlock intermediate block, 3 UpBlock upsampling blocks, 1 ResNet Block residual network block, and 1 GSC (group normalization + Swish activation function + Conv convolution layer) module in the networking order. The MidBlock intermediate module includes: 1 ResNet Block residual network block + 1 SelfAttention self-attention module + 1 ResNet Block residual network block in the networking order; the UpBlock block includes: 1 Interpolate interpolation module (upsamples the image, usually using bilinear interpolation to enlarge the spatial resolution of the image) + 1 Conv convolution layer (performs convolution operation on the upsampled image to restore image details); ResNet The Block residual network introduces skip connections. After the input features are processed through several layers of convolution and activation, the original input features are superimposed, enhancing the network's learning ability and depth, thereby solving the problem of vanishing gradients in deep networks. Each element in the SelfAttention module (such as a pixel or feature in an image) interacts with other elements through attention weights to obtain global features, solving the problem of capturing long-distance dependencies. The GSC module combination (GroupNorm normalizes features, stabilizes the training process, and ensures high-quality normalization even in small batches; the Swish activation function provides smooth and differentiable activation, improving the model's nonlinear expression capabilities, suitable for deep networks and generative models; Conv extracts local features of the image and is the basic operation unit of CNN)

[0215] (11) SDM uses mature AI models and cloud computing resources to generate massive high-quality data sets required for fine-tuning and training YOLO-v11 through the text-to-graph and graph-to-graph mechanisms. You can also consider the trained, mature, and large-scale Amazon stability.ai (Stable Diffusion XL) model, which is accessed through the managed API provided by Amazon Bedrock. AWS uses the Inferentia2 chip and SageMaker inference toolchain to hardware accelerate the model, which can greatly reduce inference latency and cost. The model is deployed on EC2 Inf2 instances (optimized for inference) or GPU instances (such as P4d / P5), and the generated content is stored in S3, supporting version control and lifecycle management. Model access rights are controlled through IAM roles, and static data is encrypted using AWS's KMS key management service. User data is encrypted and transmitted via AWS PrivateLink, complying with GDPR, HIPAA and other standards;

[0216] like Figure 4 As shown, fine-tuning training and inference are performed based on the YOLO-v11 model, specifically including:

[0217] (1) The pre-trained models of YOLO-v8 to YOLO-v11 (such as those trained on the COCO dataset) are mainly targeted at common objects in natural scenes (such as pedestrians, vehicles, animals, etc.). As a general target testing framework, its image features in specialized fields such as industrial quality inspection are significantly different from those of general datasets. Therefore, in applications in specific professional fields, additional fine-tuning and retraining are usually required to solve problems such as the non-existence of COCO datasets for specific operating scenarios and the need to redefine category labels, as well as high-precision positioning of small targets during the operation (alignment of hooks and rings), symmetry detection of medium targets (centering status of guide wheels), contact surface judgment of large targets (detection of fit between box and slide rails), and binary anomaly detection (classification of rear outrigger status).

[0218] (2) Based on the high-quality dataset generated by the stable diffusion AI large model, professional images are annotated using software tools such as LabelImg and Roboflow to generate label files in YOLO format;

[0219] (3) The training set, validation set, and test set are divided into 8:1:1 ratios, and the number of samples in each category is balanced;

[0220] (4) Fine-tune the parameters of the YOLO-v11 model, and the unfreeze and freeze strategies are:

[0221] ① The Backbone model freezes the shallow CSP cross-stage local network stage, because the shallow layer of the Backbone is responsible for extracting basic features (such as edges and textures). The low-level features of industrial equipment and natural images are highly similar, so there is no need for retraining.

[0222] ②Backbone model unfreezes deep C2PSA (including multiple PSA Block pyramid squeeze attention blocks, each PSA contains 1 Attention attention block, such as Figure 5 The PSA Block is a core module in the Swin Transformer. Its primary function is to improve the model's computational efficiency and feature extraction capabilities through W-MSA windowed multi-head self-attention and SW-MSA window-shifted multi-head self-attention. The PSA Block reduces computation by limiting attention computation to within the window. The window-shift mechanism enhances interaction between windows, improving the model's representational capabilities. The PSA Block's working principles include: Patch Partitioning (dividing the input image into non-overlapping patches), Linear Embedding (mapping each patch into a high-dimensional vector space), and the Swin Transformer Block (which includes two main steps: W-MSA and SW-MSA. W-MSA computes multi-head self-attention within each window, while SW-MSA recalculates attention using window-shifting to enhance interaction between windows. Both steps include LN layer normalization and MLP multi-layer perceptron modules). The PSA Block can also be combined with the Patch Merging Layer, which merges features from adjacent patches, reducing resolution while increasing the number of channels, further improving feature extraction. The Swin Transformer module excels at capturing global contextual relationships and needs to be adjusted to recognize structural relationships between mechanical parts (e.g., the topological relationship between hooks and rings). The ultimate goal of this step is to retain general feature extraction capabilities (edges / textures) while adapting high-level semantic features to the morphology of industrial equipment.

[0223] ③ The Neck model unfreezes all BiFPN-XL layers to enhance multi-scale feature fusion capabilities and optimize the collaborative detection of small objects (such as lifting rings that rely on high-resolution feature maps) and large objects (such as boxes that require semantic information from low-resolution feature maps). The dynamic weighting mechanism needs to relearn the importance of features in industrial scenarios.

[0224] ④Head model unfreezes the Detect detector classification branch parameters (Conv2d and CLSLoss), such as Figure 6As shown in the figure, because the vehicle's rear outrigger status (extended / retracted) is a newly defined binary classification task, the classifier needs to be trained from scratch; the initial layers of the regression branch (the first two DWConv and Conv) are frozen. Because the pre-trained model already has general target positioning capabilities, fine-tuning only requires adjusting high-level parameters to adapt to the precise offset of mechanical components. The ultimate goal of this step is to adapt the classification task to the new category (such as the vehicle's rear outrigger status), while the regression task retains the pre-trained positioning capabilities and fine-tunes local parameters;

[0225] (5) Training optimization strategy

[0226] 1) Data enhancement configuration optimization for industrial scenarios of unmanned lifting operations, including but not limited to: enhancing color changes (simulating different lighting conditions), increasing saturation (enhancing rust / stain visibility), controlling brightness (adapting to nighttime operations), increasing rotation range (for variable vehicle entry angles), and enabling MixUp enhancement (improving robustness for small objects).

[0227] 2) For scenarios where the hook is aligned and hooked, the loss function considers the positioning loss CIoULoss and adds a small target weight coefficient, which increases the loss weight:

[0228]

[0229] Among them, CLoU It is a loss function that calculates the distance between two bounding boxes in target monitoring. It is improved based on IoU and can more accurately reflect the similarity between the two bounding boxes. CLoU At the same time, factors such as the position, size, and shape of the bounding box are taken into consideration, which can effectively improve the accuracy of target detection.

[0230] IoU is the intersection over union ratio between two bounding boxes, ρ 2 (b pred ,b gt ) represents the square of the center distance between the prediction box and the GT (GroundTruth) box, c is the diagonal length of the minimum envelope of the two boxes, α is the adjustment coefficient, v represents the aspect ratio of the bounding box, w is the width of the image, h is the height of the image, img_size is the image size, ω small is the small target weight coefficient:

[0231] Dynamic calculation: ω small =1.0+2*(1-(wh) / img_size)

[0232] Fixed setting: For targets with an area smaller than 32*32, ω small Assign a weight of 2.0 to 3.0 times.

[0233] 3) For the case of unbalanced samples of the vehicle's rear outrigger states, the loss function considers the classification focal loss FocalLoss + influence factor to parameterize the focal loss:

[0234]

[0235] Among them, CLS represents the binary cross entropy loss function, α dynamic Represents the weight coefficient, p t represents the predicted probability, γ adaptive is an adaptive adjustment factor, epoch represents the number of cycles of model training, and total_epoch represents the total number of cycles of model training;

[0236] (6) Develop an acceleration strategy, deploy using Tensor RT, convert the fine-tuned YOLO-v11 model to FP16 precision, and optimize the computational graph. Edge device adaptation can use chips such as Nvidia Jetson AGX Orin, with a real-time inference speed of ≥45FPS.

[0237] (7) Model verification and debugging, considering key indicators such as overall accuracy, strict thresholds, and recall rate of detection in lifting operation scenarios (to avoid missed detections that lead to operation failures); visualization tools can use TensorBoard to monitor feature map responses to ensure that small targets such as the lifting ring area are significantly activated;

[0238] Evaluation is performed based on the fine-tuned YOLO-v11 model;

[0239] Perform real-time inference based on the fine-tuned YOLO-v11 model.

[0240] By specifically freezing the shallow Backbone layer, unfreezing the deep layer and Neck, and optimizing the Head classification branch, YOLO-v11 can effectively adapt to the detection requirements of unmanned lifting operations for hook-arm garbage trucks. In actual deployment, the model is deployed to the MEC edge computing system and the vehicle-side LODC lifting operation domain controller. After fine-tuning its parameters, YOLO-v11 feeds real-time detection results (such as coordinate offsets) to the control unit in real time, enabling unmanned and fully automated lifting operations.

[0241] Example 2:

[0242] Based on Example 1, the hook-arm garbage truck includes:

[0243] Image detection module, which is composed of a newly added camera, including:

[0244] A camera installed on the top frame of the garbage room door at a garbage collection point monitors the hook's secure attachment to the lifting ring, the balance and stability of the garbage compactor, and the deployment / retraction of the vehicle's rear outriggers during the lifting operation of a hook-arm garbage truck.

[0245] A camera is deployed on the left and right sides of the middle of the parking space. When the garbage compression box is loaded onto the vehicle, it is used to monitor in real time whether the bottom of the garbage compression box is fully fitted with the rear frame slide rail and whether the rear outriggers of the vehicle are deployed / retracted.

[0246] The three cameras also assist in real-time monitoring of the vehicle's surroundings during lifting operations, ensuring there are no pedestrians or dangerous obstacles within a 5-meter radius of the operating area. They also monitor: the balance and stability of the waste compression bin during lifting; the complete fit of the bottom of the waste compression bin with the rear frame rails; and the deployment / retraction of the vehicle's rear outriggers.

[0247] The hook-arm garbage truck is equipped with four AVM surround-view cameras, which are set up in the front, back, left and right of the vehicle to detect whether the lifting operation area is safe, that is, there are no pedestrians and other dynamic moving obstacles;

[0248] A wide-angle camera is installed at the hook of the hook-arm garbage truck to detect whether the hook of the hook-arm truck is aligned with the lifting ring of the garbage compression box and whether the hook is secure during the lifting process;

[0249] The setting of the line parking space is as follows: starting from 2.5 meters away from the entrance of the garbage collection point, a dedicated line parking space for hook-arm garbage trucks is set up. The width of the line parking space (boundary lines on both sides) is aligned with the inner measure of the garbage room exit, and the length of the line parking space covers the length of the hook-arm garbage truck.

[0250] like Figure 1 As shown in the figure, the electronic and electrical architecture of the existing hook-arm garbage truck is redesigned, mainly including ICG, APDC, chassis system, VCU, power and module group, BDC, ICDC, and LODC, among which:

[0251] (1) The intelligent central computing gateway (ICG) integrates the functions of the vehicle gateway, Tbox, and 5G-V2X, and supports software protocol stack encapsulation and data routing forwarding for vehicle Ethernet, CAN, CAN-FD, wireless cellular networks, V2X communication (this invention mainly reflects the capabilities of V2I), Bluetooth, WIFI, etc.

[0252] (2) The automatic parking domain controller (APDC) uses LVDS to connect to four AVM surround-view cameras to detect and identify parking spaces in the dedicated lanes for hook-arm garbage trucks at fixed urban garbage collection points, as well as the perception and recognition of the surrounding environment (pedestrians, obstacles, etc.) for lifting operations. APDC supports vehicle-mounted Ethernet, CAN-FD, CAN, and other connection methods for external related systems. The deep learning model built into the vehicle-side APDC (using the AVM surround-view fisheye camera) needs to identify and detect scenarios including parking spaces in the lanes, pedestrians in the lifting operation area, and other dynamic obstacles.

[0253] As an optional method, if people are in the vehicle during the lifting operation, the APDC can forward the AVM video stitching image to the intelligent cockpit domain controller (ICDC) via the in-vehicle Ethernet and the ICG gateway for real-time display on the central control screen (CSD).

[0254] (3) The chassis domain CAN bus has electronic parking brake (EPB), brake-by-wire dual-redundancy system (EMB), and electronic power steering system (EPS);

[0255] (4) Because there is no human intervention during the entire process of automatic parking and lifting operations, a certain degree of redundancy needs to be considered for key vehicle-side systems to ensure functional safety. Specifically,

[0256] Dual redundant communication networks: APDC controls chassis system commands via the ICG, forwarding them to the various chassis system modules. The APDC also has a dedicated private CAN network directly connected to the chassis system's EMC and EPS modules. If the network connected to the ICG experiences an anomaly, the APDC can directly send commands to the chassis system via the dedicated private CAN network for control.

[0257] EMB: EMB requires dual redundancy. If EMB#A fails while a person is outside the vehicle, EMB#B can take over and brake normally.

[0258] (5) The power domain has two CAN buses. The power domain CAN-1 has the rear motor controller integrated transmission MCRT and the rear transmission control unit TCUR to realize automatic gear shifting of the vehicle; the power domain CAN-2 has the power domain integrated module group, which mainly includes DC-DC converter, steering DC-AC converter, brake DC-AC converter, high-voltage battery management module, high-voltage power distribution unit, lifting operation hydraulic system power module, etc., which are not listed one by one;

[0259] (6) The vehicle control unit VCU is connected to the chassis domain CAN, power domain CAN-1, and power domain CAN-2 at the same time to realize the control of the vehicle chassis system and power system;

[0260] (7) The body domain controller (BDC) is connected to the ICG gateway via the body domain CAN bus. The master tire pressure monitoring module (TPMM) and the slave tire pressure monitoring module (TPMS) are mounted below the BDC via the CAN bus. The main function of the BDC is to check and issue warnings on the vehicle's tire pressure status before the lifting operation is performed.

[0261] (8) The lifting operation domain controller LODC is connected to the ICG gateway via the CAN bus, and is also connected to the wide-angle camera at the vehicle hook via LVDS. 1) During the lifting operation, the camera collects and monitors the video data of the hook of the hook arm vehicle and the lifting ring of the garbage compactor in real time, and the LODC uses real-time reasoning to confirm whether the hook and the lifting ring are aligned and hooked, and whether the hook is firmly on the lifting ring. 2) The deep learning model built into the vehicle-side LODC (using the wide-angle camera at the vehicle hook) needs to identify and detect the following scenarios: whether the vehicle hook is aligned and hooked with the lifting ring of the garbage compactor, and whether it is firmly;

[0262] (9) LODC connects the detection modules and electronic control modules required for lifting operations through a private CAN, including: hydraulic oil level detection module (checks whether the oil level in the vehicle hydraulic system is normal), hydraulic pipeline detection module (checks whether the lubricating oil in the vehicle hydraulic system is leaking), quick connector detection module (checks whether the quick connector for disconnecting and connecting the oil circuit is damaged), power take-off electronic control module (starts / stops the power take-off in an electronic control mode), rear outrigger electronic control module (lowers / restores the rear outrigger), main arm lock hook electronic control module (locks / releases the main arm lock hook), hook arm extension electronic control module (operates the hook arm extension length);

[0263] As an optional feature, the Intelligent Cockpit Domain Controller (ICDC) can be connected to the central control screen (CSD) via LVDS, allowing the driver to set parameters and monitor the lifting operation status in real time from within the vehicle. The ICDC supports various connectivity options, including in-vehicle Ethernet and CAN-FD, for external systems.

[0264] Combine Figure 2 As shown in the figure, a multi-access edge computing MEC system is used on the field side to communicate and interact with the hook-arm garbage truck and smart terminals such as mobile phones:

[0265] (1) The field-side sensors include a wide-angle camera on the top frame of the garbage room door, a wide-angle camera on the left side of the middle of the line parking space, and a wide-angle camera on the right side of the middle of the line parking space. They are connected to the field-side MEC system through an AHD analog high-definition video transmission link;

[0266] (2) The main components of the field-side MEC system include: image acquisition and preprocessing module, environmental perception real-time reasoning module, logic calculation and judgment warning module, and V2I communication module; among which:

[0267] Image acquisition and preprocessing module: This module processes video images captured by the camera in parallel using a multi-threaded approach. During the preprocessing phase, it optimizes data quality through algorithms such as noise suppression, data standardization, color space transformation, and geometric distortion correction.

[0268] Environmental Perception Real-Time Inference Module: Based on the high-quality image data frames input in the previous step, it recognizes, detects, and tracks dynamic targets such as pedestrians and moving obstacles. Specific algorithms can include the One-Stage detector YOLO-v11, NanoDet, or attention-based optimization models such as MobileViT and EfficientFormer.

[0269] Logical calculation and judgment warning module: Based on the real-time reasoning results of environmental perception, it performs necessary logical calculations and analyzes and judges:

[0270] ① Is the lifting operation area safe (i.e., there are no pedestrians or other dynamic moving obstacles)?

[0271] ②During the lifting process, does the garbage compactor maintain balance and stability? If it deviates from the center, how much is the deviation?

[0272] ③ When the garbage compactor is fully hoisted onto the rear frame, is its bottom completely in contact with the rear frame rail?

[0273] ④ When the lifting operation starts, are the vehicle's rear outriggers deployed? When the lifting operation is completed, are the vehicle's rear outriggers retracted?

[0274] The deep learning model built into the field-side MEC (using the garbage room door top frame camera) requires auxiliary vehicle-side recognition and detection scenarios including pedestrians and other dynamic obstacles in the lifting operation area;

[0275] The deep learning model built into the yard MEC (using the camera on the top frame of the garbage room door) needs to identify and detect scenarios including: pedestrians and other dynamic obstacles in the lifting operation area; the centering / deviation of the garbage compactor and the middle position of the two guide wheels on the rear frame during the lifting process; and the deployment / retraction of the vehicle's rear outriggers when the lifting operation starts / completes.

[0276] The deep learning model built into the yard-side MEC (using cameras on both sides of the center of the yard-side parking space) needs to identify and detect scenarios such as: the bottom of the garbage compression box is completely aligned with the rear frame rail during loading; when the lifting operation is started / completed, the vehicle's rear outriggers are deployed / retracted;

[0277] V2I communication module: Based on the 5G-V2X PC5 interface, it enables direct communication between RSU and vehicles, supports low-latency, high-reliability broadcasting, and has a coverage range of 1km+. Through wireless broadcasting, the above calculation and analysis results can be actively broadcast to the hook-arm garbage truck periodically (low latency <100ms). The vehicle's ICG gateway integrates 5G-V2X functions and can passively receive broadcast V2I messages. The V2I communication protocol can use the SAE J2735 core standard, ASN.1 encoding, and RSM (roadside safety message) data structure;

[0278] The vehicle-side intelligent central computing gateway (ICG) (Gateway+Tbox 5G-V2X) implements V2I communication with the field-side MEC and communicates with the driver's mobile app via Bluetooth. The ICG routes and forwards broadcast messages received from the MEC to the automatic parking domain controller (APDC). The APDC decides whether to brake the vehicle (forwarding vehicle control commands to the chassis system via the ICG) or stop the lifting operation (forwarding vehicle control commands to the lifting operation domain controller (LODC) via the ICG, which then controls various operation modules).

[0279] The driver's mobile app connects to the vehicle via Bluetooth. After exiting the vehicle, the driver can initiate unmanned lifting operations with a single tap on their phone. During the lifting operation, information on the vehicle's operating status, operational information for each lifting module, and hazard warnings can all be transmitted to the mobile app via Bluetooth for real-time display and warning.

[0280] Optionally, the mobile phone APP and the vehicle can also establish a connection through a Wi-Fi hotspot. The vehicle can send the AVM surround-view spliced video stream, the hook and the ring alignment video stream, and other video streams to the mobile phone for real-time monitoring.

[0281] Optionally, if the driver is in the cab, the vehicle can send its own operating status information, operational information of each lifting operation module, hazard warning information, AVM surround view spliced video stream, hook and eye alignment and hooking video streams to the intelligent cockpit domain controller ICDC via the CAN bus and on-board Ethernet for real-time display, monitoring and warning on the central control screen CSD.

[0282] Example 3:

[0283] Based on Example 2, a method for automatically lifting a hook-arm garbage truck includes:

[0284] The automatic parking process of a hook-arm garbage truck includes:

[0285] (A1) The driver drives the hook-arm garbage truck to the city's fixed garbage collection point, stops the truck about 5 meters in front of the dedicated lane parking space, and gets off the truck;

[0286] (A2) The driver uses a mobile phone app to establish a wireless connection with the vehicle, such as Bluetooth or Wi-Fi, and activates automatic parking with one click;

[0287] (A3) AVM surround view camera, which uses a polynomial model to correct distortion by calibrating parameters (intrinsic and extrinsic), and then uses inverse perspective transformation (IPM) to project the four camera images into a unified bird's-eye view coordinate system;

[0288] (A4) Linear parking space detection and recognition: Use the HSV color space to segment the line area (e.g., white / yellow) and Canny to detect edges. Use the Hough transform to detect straight line segments and fit the parking space boundaries. Use a semantic segmentation model (e.g., UNet) to directly predict the parking space mask.

[0289] (A5) Obstacle Identification and Localization: Use YOLO-v11 or EfficientDet to detect pedestrians and dynamic obstacles in the lifting operation area, estimate the relative distance between the vehicle and the obstacles, and perform real-time tracking.

[0290] (A6) Generate a parking path based on the parking space coordinates, obstacle list, and vehicle's current position using a hybrid A* algorithm, taking into account vehicle kinematic constraints. Expand the vehicle's outline by the obstacle boundary, taking equally spaced points along the path to determine if they intersect with the expanded obstacle.

[0291] (A7) Use the Model Predictive Control (MPC) algorithm to optimize control inputs, minimizing path tracking error and the rate of change of the controlled variable. Send control commands to the EPS to manipulate the steering wheel angle. Send control commands to the EMB to control vehicle speed, maintaining a constant low speed (<5 km / h) for the drive-by-wire throttle / brake. Predict parking space locations based on historical data. If parking space information is lost and an anomaly occurs, trigger emergency braking.

[0292] (A8) The vehicle stops precisely in the center of the parking space, completing the automatic parking maneuver and exiting the intelligent driving mode.

[0293] The status inspection process before unmanned lifting operation includes:

[0294] (B1) Maintenance personnel at municipal waste collection points confirm whether the condition of the waste compression bin meets the requirements for unmanned lifting operations, including whether it is placed in a garbage room on flat ground, whether the bin is deformed, whether the lifting ring is damaged, and whether the garbage is overloaded (weight sensors can be installed in the garbage room to weigh the waste compression bin, or the metal garbage cover can be checked to ensure it is tightly closed to prevent garbage from spilling).

[0295] (B2) The maintenance personnel of the urban garbage collection point shall confirm that the ground in the lifting operation area must be solid and flat to avoid soft or sloped ground that may cause the vehicle to tilt;

[0296] (B3) After the status of the garbage compression box and the ground conditions of the working environment are confirmed, all garbage point operation and maintenance personnel leave the site and then notify the driver that unmanned lifting operations can be carried out. The driver uses the mobile phone APP outside the vehicle to start the status check function before the lifting operation with one click;

[0297] (B4) The TBox 5G-V2X module integrated in the vehicle ICG gateway establishes a V2I short-range wireless connection with the field-side MEC, allowing the field-side MEC to periodically send environmental perception results to the vehicle for decision-making;

[0298] (B5) The vehicle-side LODC receives a status check instruction from the ICG gateway before the lifting operation, and instructs the hydraulic oil level detection module, hydraulic pipeline detection module, and quick connector detection module to report the detection results to itself respectively;

[0299] (B6) LODC determines whether the hydraulic lubricating oil level is normal, whether the hydraulic pipeline is intact, whether the quick connector is damaged, and whether there is any lubricating oil leakage;

[0300] (B7) In addition, the LODC asks the power take-off electronic control module, rear outrigger electronic control module, main arm hook lock electronic control module, and hook arm extension electronic control module to report their status to itself respectively;

[0301] (B8) LODC determines whether the status of the power take-off, rear outrigger, main arm lock hook, and hook arm extension and retraction electronic control modules are normal;

[0302] (B9) The vehicle-side body domain controller (BDC) receives a status check instruction from the ICG gateway before the lifting operation, instructs the TPMM and TPMS to report the tire pressure test data to itself, and analyzes and determines whether the vehicle's main tire pressure and auxiliary tire pressure are normal;

[0303] (B10) Chassis EPS and EMB, check whether their working status is normal and report the status;

[0304] (B11) The vehicle-side AVM surround view camera and APDC confirm that there are no personnel or dangerous objects within a 5-meter radius of the lifting operation area;

[0305] (B12) The on-site camera and MEC assist in confirming that there are no personnel or dangerous objects within a 5-meter radius of the lifting operation area, and the confirmation result message is sent to the vehicle via V2I;

[0306] (B13) All status confirmation information before the aforementioned lifting operations is reported by the LODC, BDC, APDC, EPS, and EMB to the vehicle's ICG gateway, and then sent to the driver's mobile app via the ICG's integrated TBox Bluetooth wireless protocol. The mobile app will visually list the detailed results of each status check, and any abnormal status will require an alert on the mobile phone;

[0307] The unmanned lifting operation workflow includes:

[0308] (C1) The driver can activate the unmanned lifting operation function with one click through the mobile phone APP;

[0309] (C2) Operation instructions are sent to the TBox integrated into the vehicle's ICG via Bluetooth. The ICG forwards the operation instructions to the VCU. The VCU controls the EPB, which applies the vehicle's parking brake. The VCU controls the TCUR, which shifts the vehicle into neutral.

[0310] (C3) The ICG forwards the operation instruction to the LODC, which controls the rear outrigger electronic control module to lower the vehicle's rear outriggers.

[0311] (C4) The VCU controls the power module (the lifting hydraulic system power supply is already built in), starts the vehicle engine, and turns on the hydraulic system power. The VCU also controls the MCRT, maintaining the speed between 1500 and 2000 rpm to provide sufficient power.

[0312] (C5) The VCU controls the TCUR drive synchronizer, and the LODC controls the power take-off electronic control module. The two work together to complete the hydraulic pump power take-off action;

[0313] (C6) The LODC controls the main arm hook electronic control module to set the hook to the "release" position and hold it for 5 to 8 seconds until the main arm hook is fully released;

[0314] (C7) The LODC controls the hook arm extension and retraction electronic control module, allowing the main arm cylinder to continue extending until the auxiliary arm hook flips over to a position slightly lower than the center of the garbage compression box hook;

[0315] (C8) The VCU controls the TCUR drive synchronizer, and the LODC controls the power take-off electronic control module. The two work together to shut down the hydraulic pump power take-off.

[0316] (C9) The VCU controls the TCUR to drive the synchronizer, shifting the vehicle into reverse gear. The VCU then controls the MCRT to reverse the vehicle to the center of the hook and the trash compactor's lifting ring. A camera at the vehicle's hook monitors the connection between the hook and the lifting ring in real time. Once the LODC confirms a secure connection, the VCU controls the EMB to stop the vehicle.

[0317] (C10) Hydraulic power take-off is performed again: the VCU controls the TCUR drive synchronizer, and the LODC controls the power take-off electronic control module, and the two cooperate to complete the hydraulic pump power take-off action;

[0318] (C11) The LODC controls the hook arm extension and retraction electronic control module, causing the main arm cylinder to continuously retract, pulling the garbage compactor onto the slide rails of the rear frame. The compactor is continuously lifted until the compactor is completely suspended in the air and parallel to the frame, until the compactor is completely pulled onto the rear frame.

[0319] (C12) The camera on the top frame of the garbage room door monitors the centering between the garbage bin and the two guide wheels on the rear frame, and the MEC performs real-time analysis and judgment. If it is determined that the garbage bin has deviated from the centering position, an early warning message is immediately sent via V2I to notify the vehicle to adjust the position and angle.

[0320] (C13) During the loading process, cameras on the left and right sides of the middle parking space at the end of the yard monitor whether the bottom of the garbage compression box is fully aligned with the rear frame rails, and the MEC performs real-time analysis and judgment. The MEC sends the alignment status message to the vehicle via V2I.

[0321] (C14) The vehicle receives the fit status message between the box and the rear frame rail through V2I. If it is fully fitted, the LODC controls the main arm lock hook electronic control module to set the lock hook to the "locked" position and maintain it for 5 to 8 seconds until the main arm lock hook is fully locked. If the V2I message indicates that it cannot fit, the vehicle needs to adjust its position and angle. After receiving the message that it cannot fit completely several times, the vehicle will suspend the lifting operation and warn the driver's mobile phone APP through wireless Bluetooth for manual intervention;

[0322] (C15) After the vehicle lifting operation is completed normally, the VCU controls the TCUR drive synchronizer and the LODC controls the power take-off electronic control module. The two work together to close the hydraulic pump power take-off and release the hydraulic pressure.

[0323] (C16) The VCU controls the power domain module group to shut down the power supply of the lifting operation hydraulic system; the LODC controls the rear outrigger electronic control module to restore and retract the vehicle's rear outriggers; the vehicle TBox notifies the driver's mobile phone app via wireless Bluetooth that the lifting operation is complete;

[0324] (C17) After the driver gets on the vehicle, he manually releases the EPB handbrake. After receiving the signal, the vehicle ICG disconnects the V2I connection with the field-side MEC, and the driver drives away.

[0325] Although the present invention is described herein with reference to illustrative embodiments of the present invention, the above embodiments are merely preferred embodiments of the present invention, and the embodiments of the present invention are not limited to the above embodiments. It should be understood that those skilled in the art can design many other modifications and implementations, which will fall within the scope and spirit of the principles disclosed in this application.

Claims

1. A method for automatically lifting and identifying a hook-arm garbage truck, characterized in that: include: Step S1: Expand the image samples of the automatic lifting operation process of the hook-arm garbage truck based on the SD stable diffusion AI large model to obtain a data set; Step S2: fine-tune the deep learning model using the expanded data set; Step S3: The hook-arm garbage truck uses the obtained deep learning model to automatically detect and identify images collected during the lifting operation.

2. The method for automatically lifting and identifying a hook-arm type garbage truck according to claim 1 is characterized in that: In step S1, the AI large model based on SD stable diffusion includes a text image encoder, a noise intensity variance regulator, an image information creator, and a VAE model image decoder, wherein: Text-image encoder, used to encode the input visual image and user prompt text, and output image embedding and text embedding; The noise intensity variance regulator is used to initialize the input noise into latent variables, adjust the noise intensity, predict the noise residual according to the time step, and update the latent variables; An image information creator, configured to use a scheduler to iteratively schedule a computation on a U-NET model with an attention mechanism to generate a predicted noisy image tensor; the input of the U-NET model with the attention mechanism is the updated latent variable and the image embedding and text embedding; The VAE model image decoder is used to reconstruct the predicted noise image tensor to generate a target image, i.e., a data set.

3. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 2, characterized in that: The text image encoder includes a residual network encoder and a conversion encoder. The residual network encoder is used to receive an input visual image and output an image embedding, and the conversion encoder is used to receive an input user prompt word text and output a text embedding.

4. The method for automatically lifting and identifying a hook-arm type garbage truck according to claim 2, characterized in that: The U-NET model includes a downsampling path encoder, an upsampling path decoder, and a bottleneck layer, where: The downsampling path encoder consists of four downsampling stages, each of which contains: two residual network blocks for feature extraction and temporal conditioning, a spatial deformer based on the attention mechanism for introducing context conditioning and spatial attention, and a downsampling block for reducing the resolution through strided convolution or pooling; The upsampling path decoder consists of four upsampling stages, each of which contains: two residual network blocks for recovering details in combination with residual connections, an attention-based spatial deformer for fusing conditional information, and an upsampling block for improving resolution through interpolation or transposed convolution; The bottleneck layer is used to connect the middle part of the encoder and decoder. It contains two residual network blocks for further extracting global features and an attention-based spatial deformer for enhancing long-range dependency modeling. In the U-NET model, an attention module is added after each residual network block. The input token embedding is incorporated into each processing stage and then sent to the next residual network block, forming a serial structure. A portion of the input of each attention module is directly sent to the final stage of the processing process in the form of an attention module.

5. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 4, characterized in that: The residual network block includes: a normalization layer for prompting training stability, an activation function SILU for enhancing nonlinearity, a 3×3 convolutional layer for processing spatial features, a time step embedding injection for integrating temporal information into features through a fully connected layer, and a residual connection for retaining input features and alleviating gradient disappearance.

6. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 4, characterized in that: The core mechanisms of the attention-based spatial deformer include: Self-attention, modeling the relationship between regions within the image; Cross-Attention, injecting text conditions into the image generation process; Position encoding, which preserves spatial position information through sinusoidal encoding.

7. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 2, characterized in that: The VAE model image decoder is used to gradually reconstruct an image from the latent variable as an output through upsampling. The VAE model image decoder includes a latent space input layer and a convolutional upsampling path decoder. The convolutional upsampling path decoder includes, in networking order, 1 convolution module, 1 intermediate module including 1 residual network block + 1 self-attention module + 1 residual network block, 3 interpolation modules including, in networking order, 1 for upsampling the image and amplifying the spatial resolution of the image, 1 upsampling block of the convolution layer for performing a convolution operation on the upsampled image and restoring image details, 1 residual network block for superimposing the original input features after several layers of convolution and activation processing by introducing a jump connection, and 1 GSC module combination. The GSC module combination is group normalization + activation function Swish + convolution layer.

8. The method for automatically lifting and identifying a hook-arm type garbage truck according to any one of claims 1 to 7, characterized in that: The step S2 specifically includes: Define category labels; The operation process defines small target high-precision positioning, medium target symmetry detection, large target contact surface judgment, and binary abnormality detection. The small target high-precision positioning includes hook and ring alignment; the medium target symmetry detection includes guide wheel centering status; the large target contact surface judgment includes box and slide rail fit detection; the binary abnormality detection includes rear leg status classification. Label the data set and generate a label file in YOLO format; Divide the dataset into training set, validation set and test set; Use the training set to fine-tune the parameters of the pre-trained model, including: A1, Backbone model freezes the shallow cross-stage local network CSP stage; A2,Backbone model unfreezes the deep C2PSA stage, which adjusts high-level semantic features to identify the structural associations of mechanical parts; A3. Neck model unfreezes all BiFPN-XL layers to enhance multi-scale feature fusion capabilities and optimize the collaborative detection of small and large objects; A4. In the Head model, the Detect detector classification branch parameters Conv2d and CLSLoss are unfrozen and trained from scratch to adapt to the new category; the initial layer of the regression branch is frozen to retain the pre-trained positioning ability and fine-tune the local parameters to adapt to the precise offset of the mechanical parts; Use the training set to train and optimize the pre-trained model, including: B1. Optimize the sample enhancement configuration in the training set, including one or more of enhancing hue changes, increasing saturation, controlling brightness, increasing rotation range, and enabling MixUp enhancement; B2. For the scene where the hook is aligned with the ring, the loss function for calculating the distance between the two bounding boxes in the target monitoring is in, Used to reflect the similarity between two bounding boxes; IoU is the intersection over union ratio between two bounding boxes, ρ 2 (b pred ,b gt ) represents the square of the center distance between the predicted box and the ground truth GT box, c is the diagonal length of the minimum envelope of the two boxes, α is the adjustment coefficient, v represents the aspect ratio of the bounding box, w is the width of the image, h is the height of the image, img_size is the image size, ω small is the small target weight coefficient; Dynamic calculation: ω small =1.0+2*(1-(wh) / img_size); Fixed setting: For targets with an area smaller than 32*32, ω small Assign 2.0 to 3.0 times weight; B3. For the case of unbalanced samples of vehicle rear outrigger status, binary cross entropy loss function in, represents the binary cross entropy loss function, α dynamic Represents the weight coefficient, p t represents the predicted probability, γ adaptive is an adaptive adjustment factor, epoch represents the number of cycles of model training, and total_epoch represents the total number of cycles of model training; B4. Use Tensor RT deployment to convert the fine-tuned pre-trained model to FP16 precision, optimize the computational graph, and obtain a deep learning model. B5. Use the validation set and test set to evaluate the deep learning model of B4 respectively. After the evaluation is passed, the deep learning model will be deployed to the edge computing MEC system of the hook-arm garbage truck, the vehicle-side lifting operation domain controller LODC, and the vehicle-side automatic parking domain controller APDC respectively.

9. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 8, characterized in that: The step S3 specifically includes: The hook-arm garbage truck uses the obtained deep learning model to automatically detect and identify images collected during the lifting operation; The deep learning model built into the hook-arm garbage truck receives input lifting operation video data, performs image detection and recognition, and feeds back the real-time detection results to the vehicle control unit VCU of the hook-arm garbage truck in real time, realizing unmanned and fully automatic lifting operations.

10. The method for automatic lifting identification and detection of a hook-arm type garbage truck according to claim 9, characterized in that: The hook-arm garbage truck includes an image detection module, an intelligent central computing gateway (ICG), an automatic parking domain controller (APDC), a chassis system, a vehicle control unit (VCU), a power system, a body domain controller (BDC), an intelligent cockpit domain controller (ICDC), and a lifting operation domain controller connected to the intelligent central computing gateway (ICG). The image detection module is used to obtain video data of the hook-arm garbage truck's parking space and the surrounding environment of the lifting operation, and input it into the automatic parking domain controller APDC. It is also used to collect video data of the hook of the hook-arm garbage truck and the lifting ring of the garbage compactor, and input it into the lifting operation domain controller LODC. The intelligent central computing gateway (ICG) is used to implement protocol stack encapsulation and data routing and forwarding between the automatic parking domain controller (APDC), chassis system, vehicle control unit (VCU), power system, body domain controller (BDC), intelligent cockpit domain controller (ICDC), and lifting operation domain controller (LODC); The automatic parking domain controller (APDC) is used to recognize images input by the image detection module and communicate with the vehicle control unit (VCU). The VCU chassis system can then implement automatic parking, or the APDC can directly issue control commands to the chassis system to implement automatic parking. The APDC has a built-in AI deep learning neural network that trains and learns from massive amounts of images to detect, judge, and control the automatic parking process. Power system, used to provide transmission control, speed control and power; Vehicle control unit VCU, used to control the chassis system and power system; Body domain controller (BDC), used to check and issue warnings on vehicle tire pressure status; The lifting operation domain controller LODC is communicated with the image detection module and is used to judge whether the hook and the lifting ring are aligned and hooked, and whether the hook is firmly on the lifting ring based on the input video data; the lifting operation domain controller LODC has a built-in AI deep learning neural network, which trains and learns massive images to realize the detection, judgment and control of the automatic lifting operation process; it also connects the detection module and the electronic control module through a private CAN. The detection module is used to detect the hydraulic oil volume, hydraulic pipeline and quick connector of the hook-arm garbage truck, and the electronic control module is used to control the power take-off, rear support legs, main arm lock hook and hook arm extension of the hook-arm garbage truck, which is used to realize status detection before lifting operation and start lifting operation.

Images