A heavy truck onboard reversing imaging system

Abstract

The present invention relates to the field of reversing image systems, and more specifically, to a heavy truck vehicle-mounted reversing image system, a vehicle-mounted terminal, an edge computing unit, and a cloud service center; wherein the vehicle-mounted terminal is used to collect image data of the vehicle's surrounding environment and display the image processed by the edge computing unit; the edge computing unit is used to receive the image data collected by the vehicle-mounted terminal, and perform layered circular filtering enhancement processing on it, generate enhanced images and transmit them to the vehicle-mounted terminal for display; the cloud service center is used to receive the image data and processing logs uploaded by the edge computing unit, and generate image enhancement strategies based on global data, which are sent to the edge computing unit, and utilize the spatial scale characteristics of the image, and adaptively extract the feature layers of the image at different spatial frequencies through a set of circular Gaussian filters of different sizes. The nonlinear mapping method of each layer is designed in a targeted manner, and differentiated segmentation strategies are adopted in dark areas, bright areas, and transition areas.

Description

Technical Field

[0001] The present invention relates to the field of reversing imaging systems, and more particularly to a reversing imaging system mounted on a heavy truck. Background Art

[0002] With the rapid development of the transportation industry, heavy-duty trucks have become the backbone of freight transport. However, due to their enormous size and wide rear blind spots, reversing is difficult, leading to frequent accidents and posing safety risks to drivers and pedestrians. Therefore, there is an urgent need to develop advanced on-board surround-view systems that can provide heavy-duty truck drivers with all-round, high-definition reversing images, enhance their situational awareness in complex scenarios, and reduce the risk of accidents.

[0003] Currently, in-vehicle camera technology is widely used in both passenger cars and commercial vehicles. The mainstream solution involves installing a fisheye camera on the rear of the vehicle. This captures rearward images through a wide-angle lens, then applies algorithms like image distortion correction and contrast enhancement to improve visual quality and assist the driver in reversing. However, due to the computing power and power consumption limitations of automotive-grade hardware platforms, these algorithms are often relatively simple and lack adaptability to harsh environments such as low light levels, strong backlight, and rain and fog. This makes them less effective in the complex operating conditions faced by heavy-duty trucks. This is particularly true when reversing in narrow lanes or construction sites, where images often suffer from underexposure, overexposure, and incomplete field of view, making it difficult for the driver to make informed decisions.

[0004] To overcome the impact of harsh environments, the industry has begun introducing multi-camera solutions and intelligent enhancement algorithms. For example, multiple cameras, including front, rear, left, and right cameras, collaborate to capture panoramic images, combined with multi-frame noise reduction and intelligent contrast enhancement techniques to improve visual quality under complex lighting conditions. Existing surround view systems use the CLAHE algorithm to adaptively adjust pixel brightness and reduce noise through multi-frame weighted averaging, which improves imaging in extreme conditions such as backlighting. However, due to limitations in single-vehicle data and computing power, this solution cannot easily handle long-tail cases such as drastic lighting changes and partial occlusions. Furthermore, algorithm parameter tuning is time-consuming and labor-intensive, making it difficult to adapt to varying vehicle operating conditions and driving habits.

[0005] With the development of connected vehicles and autonomous driving technologies, academia has begun exploring a new paradigm for cloud-based collaborative perception. The basic idea is to utilize onboard communication modules to upload image data collected by the vehicle to the cloud, leveraging the cloud's massive data and computing power to optimize perception models. The results are then fed back to the vehicle to guide perception decisions. This cloud-based learning-at-the-edge execution architecture enables algorithms to scale from single vehicles to fleets, significantly expanding the optimization space for perception algorithms based on data evolution. However, existing research has largely focused on autonomous driving for passenger vehicles and has not fully considered the specificities of heavy-duty truck reversing scenarios, such as cargo compartment obstructions, narrow aisles, and harsh operating conditions, making direct application to passenger vehicle solutions difficult. Furthermore, the latency and bandwidth limitations of cloud-edge data exchange make it difficult for such solutions to provide real-time guidance on reversing commands.

[0006] In summary, existing heavy truck reversing assistance systems still cannot meet actual usage needs in terms of perception accuracy, scene adaptability, and cloud-based empowerment. The root cause lies in:

[0007] 1. Existing image enhancement algorithms fail to fully consider the characteristics of heavy truck reversing images, which are characterized by large near objects and small far objects, and rich details. They use a single-scale filter and focus only on overall brightness adjustment, while ignoring the refined enhancement of features at different levels. This makes it difficult to balance detailed textures with overall contours.

[0008] 2. The existing adaptive enhancement strategy has single parameters, rigid rules, and lacks a flexible mapping mechanism. It is difficult to cope with extreme lighting conditions such as backlight and dark night, making it easy for key targets to be submerged in underexposed or overexposed areas, posing a safety hazard.

[0009] 3. The existing fusion method uses simple weighting, which fails to give full play to the complementary role of features at different levels and is struggling to balance the sharpness of details and overall naturalness.

[0010] 4. The existing system lacks cloud-based data aggregation and strategy evolution mechanisms. Its perception algorithms are limited to single-vehicle learning, failing to fully exploit the common patterns across data from different vehicle types and road conditions. Algorithm generalization and adaptability to long-tail scenarios need improvement. Furthermore, cloud-based capabilities fail to reach the front lines in real time, leaving vehicle-side perception disconnected from cloud-based learning and failing to form a closed-loop evolutionary loop.

[0011] Therefore, there is an urgent need for a new on-board image enhancement solution for heavy truck reversing. Summary of the Invention

[0012] The objective of this invention is to address the aforementioned technical issues by proposing a reversing image system for heavy-duty trucks. At the algorithmic level, this system fully utilizes the circular imaging characteristics of heavy-duty truck reversing images to refine and differentiate features at different spatial scales, achieving seamless integration from the local to the overall image, and finding the optimal balance between detail sharpness and overall coordination. Regarding brightness adjustment, this system implements adaptive, piecewise nonlinear enhancement for varying lighting and road conditions, maximizing the dynamic range of imaging and ensuring that key targets remain visible even under extreme conditions. At the architectural level, this system breaks the shackles of single-vehicle perception and fully leverages fleet data to enhance perception capabilities. This system forms a continuous evolutionary path across vehicle types and scenarios in the cloud, achieving deep collaboration between cloud-side learning and vehicle-side execution, harmoniously unifying advanced algorithms with automotive-grade constraints and seamlessly integrating the perception experience with driving habits. Through collaborative innovation of cloud-edge-end algorithms, this system achieves comprehensive improvements in the accuracy, real-time performance, adaptability, and flexibility of heavy-duty truck reversing image enhancement, building a safe, reliable, and continuously evolving intelligent reversing assistance system that will help the heavy-duty truck industry reduce costs and increase efficiency, while serving the national economy and people's livelihood.

[0013] The present invention provides a heavy truck vehicle-mounted reversing imaging system, comprising: a vehicle-mounted terminal, an edge computing unit, and a cloud service center; wherein,

[0014] The vehicle-mounted terminal is used to collect image data of the vehicle's surrounding environment and display the image processed by the edge computing unit;

[0015] The edge computing unit is used to receive the image data collected by the vehicle terminal, perform layered circular filtering enhancement processing on the image data, and generate an enhanced image to be transmitted to the vehicle terminal for display;

[0016] The cloud service center is used to receive the image data and processing logs uploaded by the edge computing unit, and generate an image enhancement strategy based on global data, and send it to the edge computing unit.

[0017] Specifically, the vehicle-mounted terminal includes:

[0018] Multiple cameras are installed at different positions on the rear of the vehicle to collect image data of the vehicle's surroundings from multiple angles;

[0019] A display screen installed in the vehicle cab, for displaying the enhanced image transmitted by the edge computing unit;

[0020] The camera is connected to the edge computing unit via a CAN bus, and the display screen is connected to the edge computing unit via a USB.

[0021] Specifically, the functions of the cloud service center include:

[0022] Global data management, receiving, storing and analyzing the image data and processing logs uploaded by the edge computing unit;

[0023] Adaptive strategy learning: Based on global data, it trains and generates image enhancement strategies suitable for different scenarios;

[0024] Fleet management and monitoring: real-time monitoring of the status of all connected vehicles, and provision of remote management and upgrade services.

[0025] Specifically, the edge computing unit includes: a layered circular filtering enhancement module, a cloud collaboration module and a lightweight inference engine;

[0026] The layered circular filtering enhancement module is used to receive the image captured by the vehicle-mounted terminal, perform adaptive layered enhancement, and output an enhanced image;

[0027] The cloud collaboration module is used to exchange data with the cloud service center and receive the layered enhancement parameters issued by the cloud service center;

[0028] The lightweight inference engine is used to load the end-to-end hierarchical enhancement model trained by the cloud service center to process images in real time.

[0029] Specifically, the layered circular filtering enhancement module includes: a layered module, an adaptive enhancement module and a layered fusion module;

[0030] The hierarchical module is used to use N circular Gaussian filters {H i |H i ∈R K×K , i=1,2,...,N} for the input image I∈R H×W Perform convolution layer by layer to obtain N-layer scale images {I i |I i ∈R, i=1, 2, ..., N}, the convolution process is:

[0031] I i =I*H i , i=1,2,...,N

[0032] Among them, * represents the convolution operation; the filter radius increases in geometric progression:

[0033] r i =r0·q i-1 , i=1,2,...,N

[0034] Among them, r0 and q are the initial radius and geometric coefficient respectively;

[0035] The adaptive enhancement module is used to enhance the layered image {I i}Perform adaptive nonlinear mapping and output enhanced image {I′ i}:

[0036] I′ i =f i (I i ; α i , β i , γ i ), i = 1, 2, ..., N

[0037] Among them, f i is the piecewise mapping function:

[0038]

[0039] Among them I i represents some strength or value of the input, and the parameter α i , β i , γ i , δ i Determined by the enhancement policy received by the cloud collaboration module;

[0040] The layered fusion module is used to enhance the image I' i Perform weighted summation to obtain the final enhanced result I′:

[0041]

[0042] Among them, w i is the weight, N is the total number of images;

[0043]

[0044] Among them, a i is the i-th layer score issued by the cloud service center, and norm represents the normalization function.

[0045] Specifically, the functions of the cloud collaboration module include:

[0046] Environmental perception: extracting statistical features of images and uploading them to the cloud service center, and receiving environmental categories returned by the cloud service center;

[0047] Data transmission: uploading the hierarchical features and enhancement results to the cloud service center for enhancement strategy learning;

[0048] Strategy update, receiving the latest layered enhancement parameters issued by the cloud service center, including filtering scale, mapping coefficient and fusion weight.

[0049] Specifically, the lightweight inference engine is derived based on cloud-based knowledge distillation and is used to perform real-time layered enhancement on the vehicle side. Its structure is as follows:

[0050]

[0051] in, is an end-to-end mapping network trained in the cloud, e represents the environment parameters, and Θ is the model weight;

[0052] The lightweight inference engine uses TVM for on-device acceleration optimization, uses heterogeneous computational graphs to convert models into highly optimized hardware-related programs, and utilizes on-board GPUs and NVDLA for low-latency inference.

[0053] Specifically, the edge computing unit uses the NVIDIA Jetson AGX Xavier embedded system module, including an 8-core ARM CPU, a 512-core Volta GPU, 64 tensor cores and 2 NVDLA accelerators, equipped with 16GB LPDDR4 memory and 32GB eMMC storage.

[0054] Specifically, the camera in the vehicle terminal is a fisheye camera, using a Sony IMX586 sensor, with a resolution of 8000×6000, a pixel size of 0.8μm, and a field of view of 200°; the display screen is a 12.3-inch LCD screen with a resolution of 1920×720, a brightness of 1000nit, and supports touch function.

[0055] The present invention has the following beneficial effects:

[0056] 1. Efficient collaboration between multi-scale hierarchical feature extraction and adaptive enhancement. Traditional image enhancement methods usually adopt fixed scales and unified strategies, which makes it difficult to take into account both the overall structure and local details of heavy truck reversing images. The present invention cleverly utilizes the spatial scale characteristics of the image, and adaptively extracts the feature layers of the image at different spatial frequencies through a set of circular Gaussian filters of different sizes. The nonlinear mapping method of each layer is designed in a targeted manner, and differentiated segmentation strategies are adopted in dark areas, bright areas, and transition areas to stretch and compress the dynamic range in a targeted manner. The two links complement each other. First, the complex original image data is losslessly decoupled and broken down into parts. Then, each component is precisely controlled to give full play to its strengths. Finally, they are fused and reconstructed to turn the parts into a whole. In this way, a perfect balance is achieved between the whole and the part, light and dark, and primary and secondary. The image layers are rich but not messy, and the details are clear but not abrupt.

[0057] 2. Adaptive adjustment of fusion weights and the dynamic balance between hierarchical features. When fusing multi-scale hierarchical features, traditional methods often use fixed weight summation or simply assign greater weight to the high-frequency layer, failing to fully realize the potential of each layer's features. This invention, however, takes a different approach, dynamically adjusting the weights of each layer based on scene semantic information transmitted in real time from the cloud. When the reversing scene is primarily close-up and detail is paramount, the weights of the mid- and high-frequency layers increase; when the scene is vast and overall coordination is more important, the weight of the low-frequency base layer increases. Maintaining a dynamic balance between high and low frequency layers avoids both detail loss due to oversmoothing and noise interference due to oversharpening. Like an automatic gain control system, it constantly monitors environmental feedback and dynamically adjusts the gain of each frequency band, ensuring an optimal balance between overall coordination and rich detail. Image quality varies with scene and time.

[0058] 3. Closed-loop iteration of cloud-side data aggregation and on-device policy updates. Single-vehicle intelligence has always been plagued by limited training data and insufficient scenario coverage. Cloud-based learning, by backhauling and aggregating data to form massive scenario datasets, provides a solid foundation for improving algorithm performance. Simultaneously, the cloud leverages its powerful computing power to continuously refine essential strategies, which are then distributed to the edge to guide parameter updates. This repetitive cycle of continuous data supply and policy evolution enables the system to break through its perception ceiling, achieving a qualitative leap in scenario adaptability and robustness. The edge, rather than passively accepting input, actively adapts to hardware constraints, maximizing the potential of automotive-grade computing and contributing to improvements in real-time performance and deployment efficiency. This bidirectional architecture, where learning is achieved through the cloud and enlightenment is achieved through the cloud, not only extends the benefits of the cloud to the edge but also enables synergy between the edge and the cloud, delivering comprehensive benefits far exceeding those of single-vehicle intelligence.

[0059] 4. Efficient and lightweight edge-side inference and a real-time, high-definition visual experience. The ultimate goal of image enhancement is to provide the driver with a lossless visual experience, which places stringent demands on computing latency. This invention cleverly utilizes knowledge distillation to share cloud knowledge and design a small, sophisticated network specifically for the edge. Through software-hardware collaborative optimization, communication optimization, and codec acceleration, it maximizes cloud-side performance within limited computing power. This creates an efficient and convenient bridge between complex algorithms and in-vehicle hardware, allowing advanced cloud-side achievements to be embedded in the edge while maximizing the use of limited edge-side resources. This eliminates the conflict between high latency, high energy consumption, and high performance, resulting in a seamless, high-definition, smooth, and real-time experience.

[0060] In summary, this invention cleverly combines hierarchical feature extraction with adaptive enhancement, dynamic weight adjustment with multi-scale fusion, end-cloud collaboration with closed-loop evolution, and lightweight inference with real-time high-definition display. This multi-dimensional approach synergizes and mutually promotes each other, systematically addressing the shortcomings of previous solutions in scene adaptability, overall detail balance, and end-cloud collaboration. This has resulted in comprehensive breakthroughs in heavy-duty truck reversing image enhancement. We believe that intelligent, high-definition, and robust visual assistance will become the new standard for the heavy-duty truck industry, safeguarding driving safety. The innovative concepts of this invention will provide a valuable design paradigm for the field of in-vehicle vision and fuel the further development of end-cloud integration. BRIEF DESCRIPTION OF THE DRAWINGS

[0061] Figure 1 This is a framework diagram of a heavy truck onboard reversing imaging system of the present invention; DETAILED DESCRIPTION

[0062] To further illustrate the technical means and effectiveness of the present invention in achieving its intended objectives, the following, in conjunction with the accompanying drawings and preferred embodiments, provides a detailed description of the specific implementation, structure, features, and effectiveness of a heavy-duty truck reversing camera system according to the present invention. In the following description, references to "one embodiment" or "another embodiment" do not necessarily refer to the same embodiment. Furthermore, specific features, structures, or characteristics of one or more embodiments may be combined in any suitable manner.

[0063] Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention belongs.

[0064] The specific solution of the heavy truck onboard reversing imaging system provided by the present invention will be described in detail below with reference to the accompanying drawings.

[0065] Example 1

[0066] See also Figure 1 The present invention discloses a heavy-duty truck on-board reversing imaging system, which includes an on-board terminal 1, an edge computing unit 2, and a cloud service center 3. The three work together to achieve adaptive enhancement of the environmental image during the vehicle reversing process.

[0067] The vehicle-mounted terminal 1 is installed in the vehicle and primarily consists of a camera and a display. The camera uses a fisheye camera equipped with a Sony IMX586 sensor, boasting an ultra-high resolution of 8000×6000 pixels and a wide 200° field of view, enabling comprehensive image capture of the vehicle's rearward environment. Multiple cameras are installed at different locations on the rear of the vehicle and connected to the edge computing unit 2 via the CAN bus, transmitting captured image data in real time. The display is a 12.3-inch LCD with a resolution of 1920×720 pixels, a brightness of up to 1000 nits, and touch support. Installed in the cab, it displays clear, real-time reversing images processed by the edge computing unit 2, assisting the driver in assessing the reversing environment.

[0068] Edge Computing Unit 2 is the core of the system, responsible for adaptively enhancing images captured by the vehicle terminal 1. It utilizes the NVIDIA Jetson AGX Xavier embedded system module, boasting powerful heterogeneous computing capabilities, including an 8-core ARM CPU, a 512-core Volta GPU, 64 Tensor Cores, and two NVDLA accelerators. It also features 16GB of LPDDR4 memory and 32GB of eMMC storage, enabling efficient execution of complex image enhancement algorithms. The Edge Computing Unit 2 software primarily consists of three components: a layered circular filtering enhancement module, a cloud collaboration module, and a lightweight inference engine.

[0069] The layered circular filtering enhancement module includes: a layered module, an adaptive enhancement module and a layered fusion module;

[0070] The hierarchical module is used to use N circular Gaussian filters {H i |H i ∈R K×K , i=1,2,...,N} for the input image I∈R H×W Perform convolution layer by layer to obtain N-layer scale images {I i |I i ∈R, i=1, 2, ..., N}, the convolution process is:

[0071] I i =I*H i , i=1,2,...,N

[0072] Among them, * represents the convolution operation; the filter radius increases in geometric progression:

[0073] r i =r0·q i-1 , i=1,2,...,N

[0074] Among them, r0 and q are the initial radius and geometric coefficient respectively;

[0075] The adaptive enhancement module is used to enhance the layered image {I i}Perform adaptive nonlinear mapping and output enhanced image {I′ i}:

[0076] I′ i =f i (I i ; α i , β i , γ i ), i = 1, 2, ..., N

[0077] Among them, f i is the piecewise mapping function:

[0078]

[0079] Among them I i represents some strength or value of the input, and the parameter α i , β i , γ i , δ i Determined by the enhancement policy received by the cloud collaboration module;

[0080] The layered fusion module is used to enhance the image I' i Perform weighted summation to obtain the final enhanced result I′:

[0081]

[0082] Among them, w i is the weight, N is the total number of images;

[0083]

[0084] Among them, a i is the i-th layer score issued by the cloud service center 3, and norm represents the normalization function.

[0085] The layered circular filter enhancement module is one of the core components of the edge computing unit 2, which is used to receive the original image data from the vehicle terminal 1 and perform adaptive and layered image enhancement. It first uses a set of circular Gaussian filters of different sizes to perform layer-by-layer convolution on the original image, decomposing the image into N scale layers {I i |I i ∈R H×W , i = 1, 2, ..., N}. The radius r of these Gaussian filters iThe filter increases in a geometric progression, with the initial radius r0 and common ratio q determined in real time by the cloud collaboration module based on the current scene characteristics. Smaller-radius filters extract high-frequency components such as image details and textures, while larger-radius filters extract low-frequency components such as background and contours. This layered strategy fully utilizes the circular imaging characteristics of fisheye cameras, preserving the overall structure while highlighting details at different levels.

[0086] Layered enhancement module obtains layered images {I i}, adaptive nonlinear enhancement is performed on each layer, and the mapping function is f i (I i ; α i , β i , γ i ), output the corresponding enhancement layer I′ i The mapping function adopts a piecewise nonlinear design, dividing the input brightness into three intervals: low illumination, medium illumination, and high illumination, and adopting different strategies for each interval: the low illumination interval adopts the γ power function to stretch and enhance the dark details; the high illumination interval adopts the γ root function to compress and prevent overexposure of bright areas; the medium illumination interval adopts linear transformation to expand the brightness range to δ i , increase the contrast. Among them, the segmentation point α i , β i and the mapping coefficient γ i , δ i All of these are issued in real time by the cloud collaboration module based on environmental conditions. This adaptive segmented mapping intelligently adjusts the gain of each brightness range, enhancing the overall visual effect while protecting key information at different levels. Taking a daytime backlit environment as an example, the layering module appropriately reduces the brightness of the low-frequency background layer to avoid overexposure, while simultaneously stretching the details of the high-frequency foreground layer to avoid underexposure. In low-light environments at night, each frequency band is evenly stretched for overall brightening. The parameter γ typically ranges from 1.5 to 3.0; larger values ​​increase the nonlinearity of the mapping curve. δ is typically 90% to 98%, meaning the target brightness range is 90% to 98% of the original dynamic range.

[0087] Next, the layered fusion module adopts a weighted fusion strategy to combine the enhanced layers {I′ i} and the corresponding weight {w i} to generate the final enhanced result I′. iA normalized Softmax approach is employed, generated in real time by the cloud collaboration module based on the current scene characteristics, automatically highlighting the contribution of key layers. When the reversing scene is dominated by details, the fusion module appropriately increases the weight of mid- and high-frequency layers. Conversely, when the scene has a greater depth of field and the overall sense is more important, the weight of the low-frequency layers is appropriately increased. This maintains a dynamic balance, ensuring a harmonious unity between details and the overall picture. The Softmax scaling factor a typically ranges from 5 to 10; a larger value results in a greater concentration of weight on key layers.

[0088] The cloud collaboration module is the bridge for data exchange between the vehicle-side system and the cloud. On the one hand, it is responsible for transmitting the image data, layered features, enhancement results and other information collected by the vehicle-side back to the cloud for global data analysis and strategy learning by the cloud; on the other hand, it is responsible for receiving the latest environmental parameters and layered fusion weights sent by the cloud, guiding the real-time adjustment of the vehicle-side algorithm, thereby achieving the continuous evolution of the vehicle-side system. In addition, the cloud collaboration module also assumes the responsibility of environmental perception. It will extract the statistical features of the vehicle-side image (such as average brightness, contrast, etc.), and combine it with vehicle positioning, time and other information, and upload it to the cloud environment classifier. After the classifier infers the scene category of the vehicle (such as day, night, rain and fog, etc.), it will be sent to the vehicle side, triggering the corresponding enhancement strategy switch, so that the system can adapt to different environmental conditions.

[0089] The lightweight inference engine is derived from cloud-based knowledge distillation and is used to perform real-time layered enhancement on the vehicle side. Its structure is as follows:

[0090]

[0091] in, is an end-to-end mapping network trained in the cloud, e represents the environment parameters, and Θ is the model weight;

[0092] The entire layered enhancement process is executed in a pipeline on the GPU and NVDLA of the edge computing unit 2, with high efficiency and parallelism. To further improve computing performance, the lightweight inference engine adopts a highly optimized end-to-end layered enhancement network. One forward propagation can directly predict the segmentation parameters α of each layer i , β i , γ i , δ i and fusion weight w iThe network adopts a backbone network + multi-head regression layer structure and uses TVM for deep heterogeneous optimization, converting the model into an efficient, hardware-friendly tensor program. The backbone network can use lightweight feature extractors such as MobileNet and ShuffleNet to achieve millisecond-level real-time inference on edge devices. The model parameters Θ of the inference engine are pre-trained by the cloud service center 3 based on massive data and knowledge distillation. They can be updated regularly based on factors such as vehicle model and environment to follow the cloud algorithm iteration.

[0093] The lightweight inference engine takes on the heavy responsibility of real-time computing on the vehicle side. It adopts an end-to-end architecture of backbone network + multi-head regression layer, loads the parameters obtained by cloud-based knowledge distillation, and can predict all layered enhancement parameters at one time, including the division points and slopes of the piecewise function, the layered filter radius, the fusion weights, etc. Compared with traditional step-by-step calculations, end-to-end inference has lower latency and higher parallelism. In addition, the inference engine is also specially optimized using the TVM tool to generate customized calculation graphs for the heterogeneous computing resources on the Xavier board (such as GPU, NVDLA, etc.), and automatically complete the underlying optimizations such as operator scheduling, memory layout, and data prefetching, so that the algorithm can run smoothly on the automotive-grade embedded platform.

[0094] Cloud Service Center 3 is the "brain" of the system, providing data support and decision-making guidance for the vehicle-side system. It comprises three functional modules: global data management, adaptive policy learning, and fleet management. Global data management is responsible for receiving, cleaning, storing, and managing environmental data uploaded by all vehicles, such as raw images, hierarchical features, enhancement results, and driver feedback. It also annotates and categorizes these data, building a massive heavy-truck reversing dataset to serve as input for policy learning.

[0095] Adaptive policy learning is the core of Cloud Service Center 3. It leverages machine learning algorithms to automatically analyze patterns from global data and generate optimal image enhancement strategies for different scenarios. It employs an active learning paradigm, selectively sampling and annotating a small number of samples from a massive dataset to initialize an adaptive layered enhancement model. This model is then continuously fine-tuned and updated with new data, enabling continuous model evolution and the gradual acquisition of image enhancement capabilities applicable to all weather conditions and road conditions. Cloud Service Center 3 also provides fleet management capabilities, including real-time monitoring of all vehicle locations, speeds, battery levels, and abnormal conditions, as well as software updates and remote diagnostics, to facilitate unified scheduling.

[0096] The present invention's workflow is as follows: When the driver shifts the vehicle into reverse gear, multiple fisheye cameras in the vehicle-mounted terminal 1 begin operating, capturing real-time images of the environment behind and to the sides of the vehicle at a rate of 30 frames per second. These images are then transmitted to the edge computing unit 2 via the CAN bus. Edge computing unit 2 first applies Gaussian filtering to the raw image using a layered circular filtering module. It then enhances each layer using piecewise nonlinear mapping, and finally, through weighted fusion, produces the clearest and most stable image of the reversing environment. Simultaneously, the cloud collaboration module in edge computing unit 2 transmits data such as environmental characteristics and stratification results back to the cloud service center 3 for big data analysis. It also receives real-time updates to the enhancement strategy from the cloud, such as adjusting the inflection point and slope of the mapping curve based on weather conditions or adjusting the fusion ratio of different layers based on driving habits. Cloud service center 3 uses the transmitted data to continuously learn and optimize the layered enhancement strategy, forming a closed loop. Finally, edge computing unit 2 transmits the real-time enhanced, high-definition reversing image via USB to the vehicle-mounted terminal 1's central control screen for display, assisting the driver in accurately assessing the environment behind the vehicle and driving safely.

[0097] The layered circular filtering enhancement algorithm proposed in this paper fully considers the circular characteristics of fisheye camera imaging, uses circular rather than rectangular filters to decompose the image at multiple scales, and adopts an adaptive enhancement strategy for different scale layers. This algorithm is more consistent with the physical model of fisheye imaging and can improve the overall visual effect while reducing image distortion.

[0098] The adaptive segmented mapping function can automatically adjust the segmentation points and slope according to the ambient brightness, and adopt different nonlinear stretching or compression strategies for dark areas, bright areas, and normal areas. While improving the overall brightness and contrast of the image, it avoids overexposure and underexposure, making the reversing image still clear under complex lighting conditions.

[0099] The weighted fusion process removes image noise while seamlessly integrating features from different layers, preserving the overall outline while highlighting local details, improving the visual quality of both near and far views. The softmax-normalized fusion weights ensure a smooth transition between layered features, enhancing the aesthetics of the image.

[0100] Furthermore, the system's cloud-edge-device collaborative architecture enables continuous evolution. The massive amount of vehicle-side data transmitted back provides a solid foundation for policy learning, enabling the system to automatically extract the inherent laws of image enhancement in complex scenarios from this data. Enhancement policies are optimized in a closed-loop manner in the cloud and then distributed to all vehicles, simultaneously improving the overall intelligence of the fleet. The active learning paradigm further reduces the reliance on human effort for model training and accelerates policy iteration.

[0101] End-to-end inference and TVM hardware-software collaborative optimization address the efficiency challenges of algorithm-side deployment. Traditional step-by-step processing methods not only suffer from high computational redundancy, but also make it difficult for general-purpose processors to achieve optimal algorithm performance. The integrated inference model proposed in this system generates all layered enhancement parameters end-to-end, eliminating the overhead of intermediate data exchange. The perfect combination of the dedicated Xavier computing platform and TVM deep optimization software maximizes the potential of onboard heterogeneous computing resources, achieving server-like inference performance on automotive-grade chips with ultra-low millisecond latency, meeting the demanding requirements of real-time reverse video processing.

[0102] It should be noted that the above is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the principles of the present invention should be included in the scope of protection of the present invention.

Claims

1. A heavy truck onboard reversing image system, characterized in that: include: Vehicle terminal, edge computing unit and cloud service center; among them, The vehicle-mounted terminal is used to collect image data of the vehicle's surrounding environment and display the image processed by the edge computing unit; The edge computing unit is used to receive the image data collected by the vehicle-mounted terminal, perform layered circular filtering enhancement processing on the image data, and generate an enhanced image to be transmitted to the vehicle-mounted terminal for display; The cloud service center is used to receive the image data and processing logs uploaded by the edge computing unit, and generate an image enhancement strategy based on the global data, and send it to the edge computing unit; The edge computing unit includes: a layered circular filtering enhancement module, a cloud collaboration module and a lightweight inference engine; The layered circular filtering enhancement module is used to receive the image collected by the vehicle-mounted terminal, perform adaptive layered enhancement, and output an enhanced image; The cloud collaboration module is used to exchange data with the cloud service center and receive the layered enhancement parameters issued by the cloud service center; The lightweight inference engine is used to load the end-to-end hierarchical enhancement model trained by the cloud service center to process images in real time; The layered circular filtering enhancement module includes: a layered module, an adaptive enhancement module and a layered fusion module; The hierarchical module is used to use N circular Gaussian filters {H i |H i ∈R K×K , i = 1, 2, ..., N} for the input image I∈R H×W Perform convolution layer by layer to obtain N-layer scale images {I i |I i ∈R, i=1, 2, ..., N}, the convolution process is: I i =I*H i ,i=1,2,...,N Among them, * indicates the convolution operation; the filter radius increases in geometric progression: r i =r0·q i-1 ,i=1,2,...,N Among them, r0 and q are the initial radius and geometric coefficient respectively; The adaptive enhancement module is used to enhance the layered image {I i }Perform adaptive nonlinear mapping and output enhanced image {I′ i }: I' i =f i (I i ;a i ,b i ,c i ), i=1, 2,…, N Among them, f i is the piecewise mapping function: Among them I i represents some kind of strength or value of the input, and the parameter α i , β i , γ i , δ i Determined by the enhancement policy received by the cloud collaboration module; The layered fusion module is used to enhance the image I' i Perform weighted summation to obtain the final enhanced result I′: Among them, w i is the weight, N is the total number of images; Among them, a i is the i-th layer score issued by the cloud service center, and norm represents the normalization function.

2. The heavy truck vehicle-mounted reversing imaging system as claimed in claim 1, characterized in that: The vehicle-mounted terminal comprises: Multiple cameras are installed at different positions on the rear of the vehicle to collect image data of the vehicle's surroundings from multiple angles; A display screen, installed in the vehicle cab, for displaying the enhanced image transmitted by the edge computing unit; The camera is connected to the edge computing unit via a CAN bus, and the display screen is connected to the edge computing unit via a USB.

3. The heavy truck vehicle reversing imaging system as claimed in claim 1, characterized in that ,The functions of the cloud service center include: Global data management, receiving, storing and analyzing the image data and processing logs uploaded by the edge computing unit; Adaptive strategy learning: Based on global data, train and generate image enhancement strategies suitable for different scenarios; Fleet management and monitoring: real-time monitoring of the status of all connected vehicles, and provision of remote management and upgrade services.

4. The heavy truck vehicle reversing imaging system as claimed in claim 1, characterized in that ,The functions of the cloud collaboration module include: Environmental perception, extracting statistical features of images and uploading them to the cloud service center, and receiving environmental categories returned by the cloud service center; Data transmission: uploading the hierarchical features and enhancement results to the cloud service center for enhancement strategy learning; Strategy update, receiving the latest hierarchical enhancement parameters from the cloud service center, including filtering scale, mapping coefficient and fusion weight.

5. The heavy truck vehicle reversing imaging system as claimed in claim 1, characterized in that ,The lightweight reasoning engine is obtained based on the cloud knowledge distillation and is used to perform real-time hierarchical enhancement on the vehicle side. Its structure is: in, is an end-to-end mapping network trained in the cloud, e represents the environment parameters, and Θ is the model weight; The lightweight inference engine uses TVM for on-end acceleration optimization, uses heterogeneous computational graphs to convert models into highly optimized hardware-related programs, and uses on-board GPUs and NVDLA for low-latency inference.

6. The heavy truck vehicle-mounted reversing imaging system according to any one of claims 1 to 5, characterized in that ,The edge computing unit uses NVIDIA Jetson AGX Xavier embedded system module, including 8-core ARM CPU, 512-core Volta GPU, 64 tensor cores and 2 NVDLA accelerators, equipped with 16GB LPDDR4 memory and 32GB eMMC storage.

7. The heavy truck vehicle reversing imaging system as claimed in claim 2, characterized in that The camera in the vehicle terminal is a fisheye camera, using a Sony IMX586 sensor, with a resolution of 8000×6000, a pixel size of 0.8μm, and a field of view of 200°; the display screen is a 12.3-inch LCD screen with a resolution of 1920×720, a brightness of 1000nit, and supports touch function.

Images