A vehicle-mounted double-sided LED advertising display system and method with edge AI computing power

Abstract

The application discloses a vehicle-mounted double-sided LED advertisement display system and method with edge AI computing power, and relates to the field of intelligent vehicle-mounted advertisements. In view of the problem that the prior art lacks real-time environment sensing and content linkage and cannot realize double-sided differentiated intelligent display, the application locally deploys a lightweight scene understanding AI model in a vehicle-grade edge AI computing unit, fuses multi-source sensing data of a vehicle-mounted camera, a millimeter wave radar and an infrared sensor, obtains the attributes of pedestrians outside the vehicle, the environmental conditions and the state of passengers inside the vehicle in real time, generates differentiated advertisement strategies for the outside and the inside of the vehicle respectively, and synchronously distributes two-way content data packets to double-sided LED display screen driving through a cooperative control unit, so that precise and synchronous display is realized. The application realizes the deep integration of edge sensing and double-sided display, supports precise audience delivery outside the vehicle and personalized experience inside the vehicle, and effectively improves the intelligent level and display efficiency of vehicle-mounted advertisements.

Description

Technical Field

[0001] This invention relates to the field of intelligent vehicle advertising, and specifically discloses a vehicle-mounted dual-sided LED advertising display system and method with edge AI computing power. Background Technology

[0002] As an important form of outdoor media, the effectiveness of in-vehicle mobile advertising depends on an accurate understanding of the target audience's characteristics and the environmental context. Dual-sided LED displays have attracted attention because they can simultaneously target both pedestrians outside the vehicle and passengers inside. However, with the increasing complexity of urban road environments and the diversification of passenger needs, traditional in-vehicle advertising solutions are no longer sufficient to meet the requirements for precise and intelligent delivery.

[0003] Current in-vehicle advertising solutions include those that use dual-sided displays on the vehicle's roof or body to play advertisements, but these typically rely solely on a pre-set content library for periodic rotation or remote content updates. Other solutions trigger ad content switching based on the vehicle's location or route. These existing solutions share a common technical approach: they all employ a "pre-set content + timed / conditional switching" model, lacking real-time perception and understanding of the external environment and passenger status. Therefore, existing solutions struggle to dynamically adjust advertising content strategies based on road conditions, pedestrian characteristics, surrounding environment, and passenger status. This results in the external display failing to accurately target potential audiences, and the internal display failing to provide a personalized experience based on passenger characteristics. Furthermore, existing solutions lack feedback and optimization mechanisms based on display performance data, hindering continuous adjustment and optimization of ad content to improve display effectiveness.

[0004] In summary, in existing technologies, edge AI perception capabilities and dual-sided LED display systems are independent of each other and lack coordination. The displayed content cannot be effectively linked with the real-time environmental perception results, making it difficult to achieve differentiated intelligent advertising displays on both sides. These technical bottlenecks have become key issues restricting the development of in-vehicle intelligent advertising display technology and urgently need to be resolved. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a vehicle-mounted dual-sided LED advertising display method with edge AI computing power, which can effectively solve the problems mentioned in the background technology.

[0006] To achieve the above objectives, the technical solution adopted by the present invention is as follows: In a first aspect, the present invention provides a method for displaying vehicle-mounted dual-sided LED advertisements with edge AI computing power, the method comprising the following steps: Establish a real-time communication link between an automotive-grade edge AI computing unit and a dual-sided LED display screen; Deploy vehicle-mounted cameras, millimeter-wave radar, and infrared sensors to construct a three-dimensional perception network covering the external environment and the interior space of the vehicle, and synchronize the data collected by the vehicle-mounted cameras, millimeter-wave radar, and infrared sensors through a unified clock source; A scene understanding engine is deployed locally in the automotive-grade edge AI computing unit. The scene understanding engine includes a lightweight compressed AI model of the scene understanding engine. Multi-source perception data that has been synchronized with time is input into the AI ​​model of the scene understanding engine to perform inference and obtain scene understanding results. Based on the scenario understanding results, differentiated advertising content strategies are generated for the exterior and interior display surfaces of the vehicle, respectively. The collaborative control unit synchronously distributes the advertising content data packets corresponding to the differentiated advertising content strategy to the exterior display control driver and the interior display control driver. The exterior display control driver and the interior display control driver then drive the exterior LED display and the interior LED display to display the advertising content, respectively.

[0007] On the other hand, the present invention provides a vehicle-mounted dual-sided LED advertising display system with edge AI computing power, comprising: The automotive-grade edge AI computing unit has a built-in neural network accelerator, general-purpose computing processor and high-speed communication interface, which is used to deploy scene understanding engine and perform AI inference on multi-source perception data. The perception module includes onboard cameras, millimeter-wave radar, and infrared sensors deployed on the exterior and interior of the vehicle body to collect data on the external environment and passenger data. The time synchronization module is used to synchronize the data collected by the vehicle camera, the millimeter-wave radar and the infrared sensor using a unified clock source. The scene understanding engine, deployed locally on the automotive-grade edge AI computing unit, includes a lightweight and compressed AI model of the scene understanding engine, which is used to perform inference on multi-source perception data after time synchronization and output scene understanding results. The vehicle exterior advertising content strategy generation unit is used to receive the vehicle exterior scene analysis results output by the scene understanding engine, and generate the advertising content strategy for the vehicle exterior display surface based on pedestrian detection results and attribute characteristics, environmental weather conditions, and road congestion. The in-vehicle advertising content strategy generation unit is used to receive the in-vehicle scene analysis results output by the scene understanding engine, and generate advertising content strategies for the in-vehicle display surface based on the statistics of the number of passengers in the vehicle, age distribution, emotional state assessment and gaze behavior detection. The collaborative control unit has a built-in dual display control interface and a multi-channel high-speed data bus interface. It is used to receive advertising content data packets output by the AI ​​decision module, align the delivery time windows of the exterior display and the interior display to the same synchronous start time, and send display update instructions and frame buffer data to the exterior display control driver and the interior display control driver synchronously through the dual display control interface respectively. A dual-sided LED display module includes an exterior LED display and an interior LED display. The exterior LED display is controlled by the exterior display surface to display advertising content, and the interior LED display is controlled by the interior display surface to display advertising content. The feedback optimization module includes an external passenger flow statistics camera and an in-vehicle interactive sensor, which are used to collect display effect data and generate training samples through the data processing and analysis subsystem. These samples are then used by the online learning module of the AI ​​model of the scene understanding engine to perform incremental learning to update the model parameters.

[0008] In summary, this application includes at least one of the following beneficial technical effects: 1. This invention constructs a three-dimensional perception network covering both the external environment and the internal space of the vehicle, and deploys a lightweight and compressed scene understanding engine locally on an automotive-grade edge AI computing unit. This enables the advertising display system to possess, for the first time, multi-dimensional real-time perception capabilities of surrounding pedestrian attributes, road scenes, weather lighting, and the status of passengers inside the vehicle. This edge AI architecture accurately synchronizes the data collected by multi-source heterogeneous sensors in the time dimension and inputs it into the AI ​​model to perform inference, allowing the perception results to directly drive advertising content decisions. This fundamentally changes the traditional passive playback mode of in-vehicle advertising, which relies solely on a preset content library for timed rotation or remote control, achieving a technological leap from fixed content playback to real-time scene adaptation.

[0009] 2. This invention establishes an independent and parallel operating system for generating exterior and interior advertising content strategies. A collaborative control unit simultaneously distributes the two differentiated advertising content data packets to both the exterior and interior display control drivers at a unified, synchronized start time. This allows the exterior display to push the most relevant advertising material based on pedestrian group characteristics in real time, while the interior display pushes personalized content based on individual passenger status. This collaborative control architecture keeps the time deviation within the refresh interval of two adjacent frames, achieving precise synchronized output of the image while ensuring independent decision-making on both sides. This effectively solves the technical challenge of existing technologies where edge AI perception capabilities and dual-sided LED display systems are independent and lack collaboration.

[0010] 3. This invention establishes a feedback optimization closed loop based on display effect data by deploying external passenger flow statistics cameras and in-vehicle interaction sensors. It also introduces an incremental learning strategy and an elastic weight consolidation mechanism, enabling the AI ​​model of the scene understanding engine to continuously update its parameters based on quantitative statistical results of real advertising exposure events and passenger interaction behaviors. This feedback mechanism combines a conservative update strategy with historical importance constraints, preventing drastic drift in key parameters while absorbing new knowledge. This allows the system's advertising content selection capability to continuously optimize and improve over time, forming a continuous iterative cycle of perception, decision-making, display, feedback, and optimization. This provides a long-term, sustainable path for improving the efficiency of in-vehicle intelligent advertising systems. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall technical solution architecture for a vehicle-mounted dual-sided LED advertising display method with edge AI computing power.

[0012] Figure 2 This is a schematic diagram illustrating the core principle framework of the collaborative control architecture between the edge AI sensing module and the double-sided LED display module.

[0013] Figure 3 This is a logical flow framework diagram of the scene understanding and feature extraction engine that integrates and collects multi-source perception data from inside and outside the vehicle and deploys AI models locally.

[0014] Figure 4 This is a schematic diagram of the multi-level interaction relationship and data flow between the intelligent generation strategy for dual-sided differentiated advertising content and the synchronous distribution mechanism of the collaborative control unit. Detailed Implementation

[0015] To make the objectives, technical solutions, and advantages of this invention clearer, the following description is provided in conjunction with the appendix. Figure 1 To be continued Figure 4 The present invention will be further described in detail below with reference to specific embodiments.

[0016] Firstly, this invention provides a method for displaying dual-sided LED advertisements in vehicles with edge AI computing power. This method achieves intelligent advertising content generation and differentiated dual-sided display based on real-time perception of the vehicle's internal and external environment by constructing a collaborative control architecture between an edge AI perception module and a dual-sided LED display module. The following details each technical aspect.

[0017] The present invention provides a vehicle-mounted dual-sided LED advertising display method with edge AI computing power, comprising the following core steps: constructing a real-time communication link between an automotive-grade edge AI computing unit and a dual-sided LED display screen; deploying multi-sensing devices such as vehicle-mounted cameras, millimeter-wave radar, and infrared sensors; deploying a lightweight deep learning model; generating a dual-sided differentiated advertising content strategy based on AI analysis results; synchronously distributing advertising content through a collaborative control unit; and establishing a feedback optimization mechanism.

[0018] The first step, S1, involves establishing a real-time communication link between the automotive-grade edge AI computing unit and the dual-sided LED display screen, which includes the following sub-steps.

[0019] Step S101: Configure the hardware resources of the automotive-grade edge AI computing unit. The computing unit is the core computing node of the entire advertising display system, which integrates a high-performance neural network accelerator, a general-purpose computing processor, a memory controller, and a high-speed communication interface.

[0020] Neural network accelerators are used to accelerate inference operations for AI models, general-purpose processors handle system scheduling and logic control, and high-speed communication interfaces handle data exchange with external devices. The computing power of the computing units must meet the requirements for real-time analysis and processing of multiple video streams.

[0021] Step S102: Establish a physical communication link between the automotive-grade edge AI computing unit and the dual-sided LED display. The link is constructed using an automotive-grade high-speed serial bus protocol, specifically the automotive Ethernet protocol or the CAN-FD protocol.

[0022] The in-vehicle Ethernet protocol or CAN-FD protocol supports high-speed point-to-point data transmission, featuring deterministic latency and high reliability, and can adapt to operating conditions such as vibration and temperature changes in the vehicle environment. The edge AI computing unit achieves physical connection with the control and drive circuit of the dual-sided LED display through its own high-speed communication interface.

[0023] Step S103: Establish an end-to-end transmission mechanism for bidirectional data and control flows. Based on the existing high-speed physical link, establish a bidirectional transmission channel between the edge AI computing unit and the dual-sided LED display driver circuit. The bidirectional transmission channel supports the output of perceived data from the computing unit to the display driver, and also supports the return of display status feedback from the driver to the computing unit, forming a complete data loop.

[0024] The transmission rate of the high-speed bus must ensure that the end-to-end latency is controlled within 100 milliseconds. The end-to-end latency consists of the cumulative time consumed by five stages: the processing time for the edge AI computing unit to complete AI inference and generate advertising content decisions; the encapsulation time for packaging decision data into bus data frames; the transmission time of the data frames on the bus; the parsing time for the display driver circuit to receive and parse the data frames; and the refresh time for the driver circuit to control the LED pixels to complete the image rendering update. The sum of the times consumed by these five stages must be less than 100 milliseconds.

[0025] The actual measurement method for end-to-end latency is as follows. A timer is maintained inside the edge AI computing unit. The timer starts when the AI ​​inference is completed and outputs decision data, and stops when the display driver circuit completes screen refresh and returns a refresh completion confirmation frame to the computing unit. The timer value at this point is the measured end-to-end latency. The measured value is compared with a 100-millisecond threshold. If the measured value is less than or equal to the threshold, the current communication link is determined to meet the real-time requirements.

[0026] If the measured value is greater than the threshold, this method performs the following delay adjustment operation.

[0027] First, the bus communication rate can be increased to the next available rate level by adjusting the bus communication clock frequency. When using the automotive Ethernet protocol, the physical layer transmission rate is defined as 100BASE-T1 in the IEEE 802.3bw standard and 1000BASE-T1 in the IEEE 802.3bp standard. Both rate levels are implemented through the same physical interface. 100BASE-T1 is the base rate level, and 1000BASE-T1 is the adjacent higher rate level. There is no intermediate rate level between them. Therefore, switching the communication rate from 100BASE-T1 to 1000BASE-T1 is the only path to increase the next available rate.

[0028] When using the CAN-FD protocol, the data segment baud rate follows the ISO 11898-1 standard. Commonly available selectable rate levels include 2Mbps, 5Mbps, and 8Mbps, each corresponding to a different maximum operating frequency supported by the physical layer transceiver chip. 2Mbps is the base rate level, 5Mbps is the intermediate rate level, and 8Mbps is the highest rate level. The actual selection is determined based on the bus load capacity: In this embodiment, when the bus load rate is below 30%, the data segment baud rate is increased to 8Mbps; when the bus load rate is between 30% and 60%, the data segment baud rate is increased to 5Mbps; when the bus load rate exceeds 60%, the baud rate is no longer increased, and the second delay adjustment operation is performed instead.

[0029] Second, the transmission time of data frames on the bus is shortened by reducing the amount of image data per frame. Specifically, while maintaining the same display resolution, the image encoding is converted from RGB888 color format to RGB565 format, reducing the amount of data per pixel from 24 bits to 16 bits.

[0030] Both of these color formats are standard formats in the display driver field. After conversion, the color depth is reduced, but the visual difference is within the acceptable range for the human eye. If the displayed content is text or graphics, run-length encoding is used for compression before encapsulation and transmission. Run-length encoding encodes consecutively repeating pixel values ​​into numerical pairs, that is, the number of consecutive repetitions and the pixel value, which can effectively compress the amount of image data containing large areas of solid color.

[0031] The two adjustment methods mentioned above can be implemented individually or in combination until the measured end-to-end latency falls back to within the 100-millisecond threshold.

[0032] Step S104: Deploy a real-time operating system kernel and a communication protocol stack inside the edge AI computing unit. The real-time operating system kernel provides task scheduling, resource management, and priority control services to ensure that AI inference tasks and communication transmission tasks are executed in an orderly manner according to preset time constraints. The communication protocol stack is used to implement data encapsulation, routing, flow control, error detection, and retransmission of the high-speed bus protocol.

[0033] Error detection employs a Cyclic Redundancy Check (CRC) algorithm. The sending end calculates and appends a CRC checksum when encapsulating the data frame. Upon receiving the data frame, the receiving end recalculates the checksum and compares it with the received checksum. If they match, the data frame is considered to have been transmitted intact; if they do not match, the data frame is considered corrupted, and the receiving end requests a retransmission from the sending end.

[0034] Flow control is managed based on the state of the data buffer queue. The sender maintains a transmission buffer. When the used space of the buffer exceeds 80% of its total capacity, the sender reduces the data frame transmission rate; when the used space falls back to below 50% of the total buffer capacity, the normal transmission rate is restored. These two thresholds are determined with reference to general engineering practices in the field of data flow control: 80% is the high-water mark, used to trigger flow control to avoid buffer overflow and frame dropping; 50% is the low-water mark, used to release flow control to maintain normal throughput.

[0035] Step S105 achieves deep integration of the edge AI perception module and the dual-sided LED display module. Through the aforementioned collaborative control architecture, the perception data stream and display command stream converge within the automotive-grade edge AI computing unit. The AI ​​inference results are directly converted into display control commands and sent to the corresponding display driver unit without needing an additional intermediate conversion layer.

[0036] After implementing step S1, the vehicle-mounted dual-sided LED advertising display system establishes a complete closed-loop data channel from environmental perception to content generation and real-time updates of the dual-sided displays. This communication link enables subsequent multi-source perception data fusion processing, local inference of AI models, and synchronous distribution of differentiated advertising content to be completed on this channel.

[0037] The next step, S2, involves deploying various sensing devices such as vehicle cameras, millimeter-wave radar, and infrared sensors to build a three-dimensional sensing network covering both the external environment and the interior space of the vehicle. This includes the following sub-steps.

[0038] Step S201: Deploy external visual perception equipment by installing a forward-facing camera on the inside of the windshield or in the front bumper of the vehicle. The forward-facing camera's field of view covers the road ahead and the surrounding area, and is used to collect information on road conditions, vehicles, and pedestrians ahead.

[0039] A rearview camera is installed at the rear of the vehicle. The rearview camera's field of view covers the area behind the vehicle and is used to collect information about following vehicles.

[0040] Side-view cameras are installed near the side mirrors or on the doors of the vehicle. The field of view of the side-view cameras covers the blind spots to the sides of the vehicle and information about adjacent lanes.

[0041] Each camera uses an automotive-grade image sensor, featuring wide dynamic range and low-light imaging capabilities, enabling stable output of high-definition video images under various lighting conditions and weather conditions. The video output resolution of each camera is 1280×720 pixels, and the frame rate is 30 frames per second.

[0042] The frame rate of 30 frames per second is based on the following: When a vehicle is traveling on an urban road, the typical speed of a pedestrian crossing the road is approximately 1.5 meters per second. At a sampling rate of 30 frames per second, the displacement of a pedestrian between two adjacent frames is approximately 5 centimeters. This spatial resolution is sufficient to meet the requirements of subsequent object detection and tracking algorithms for inter-frame correlation. The video data thus possesses both timeliness and continuity.

[0043] Step S202: Deploy external radar sensing equipment by installing a forward-facing millimeter-wave radar inside the front bumper or behind the windshield. The forward-facing millimeter-wave radar transmits millimeter-wave signals forward and receives reflected echoes from targets to detect the distance, speed, and angle information of vehicles, pedestrians, and obstacles ahead.

[0044] A rear-facing millimeter-wave radar is installed inside the rear bumper of the vehicle. This radar is used to detect vehicles approaching from behind and to determine following distance.

[0045] Side-mounted millimeter-wave radar is installed on the side of the vehicle. This radar is used to detect moving targets in adjacent lanes and in the blind spots to the sides of the vehicle.

[0046] Each millimeter-wave radar is capable of penetrating adverse weather conditions such as rain, snow, fog, and haze, and is unaffected by lighting conditions, thus providing reliable target detection in all weather conditions. Each millimeter-wave radar outputs a target list data, which includes target identifiers, relative distance, relative velocity, relative angle, and radar cross-section attributes. The data update rate for each millimeter-wave radar is 20 times per second.

[0047] The data update rate of 20 times per second is based on the following: When a vehicle is traveling at high speed, the relative speed between the target vehicle ahead and the vehicle itself can reach 30 meters per second. At an update rate of 20 times per second, the relative displacement between two adjacent detections is approximately 1.5 meters. This spatial resolution is sufficient to meet the needs of collision warning and following distance monitoring. This update rate meets the requirements of real-time tracking applications.

[0048] Step S203: Deploy external auxiliary sensing equipment by installing infrared supplementary lighting equipment on the exterior of the vehicle body. In low-light environments, the infrared supplementary lighting equipment emits infrared light into the camera's field of view for auxiliary illumination, enhancing imaging performance at night and in low-light conditions.

[0049] An ambient light sensor is installed on the exterior of the vehicle. This sensor detects the intensity of ambient light outside the vehicle, providing input for adaptive adjustment of exposure parameters by the visual perception equipment.

[0050] Step S204: Deploy in-vehicle space sensing equipment by installing an in-vehicle infrared camera above the rearview mirror or on the roof lining. The in-vehicle infrared camera's field of view covers the front passenger area and is used to collect facial and body posture images of the passengers. The in-vehicle infrared camera uses near-infrared imaging principles and has a built-in active infrared light source, enabling clear imaging in low-light or no-light environments.

[0051] An in-cabin millimeter-wave radar is installed under the roof lining or behind the seat backs. This radar is used to detect the heart rate, respiratory vital signs, presence, and seating position of passengers inside the vehicle.

[0052] Step S205: Establish a time synchronization mechanism for multi-source sensing devices, configuring a unified clock source for each sensing device. The clock source broadcasts precise timestamps to all sensing devices at fixed intervals. When collecting data, each sensing device binds and records the currently received global timestamp with the collected data. After completing data collection, each sensing device uniformly transmits the timestamp-integrated sensing data to the edge AI computing unit. The edge AI computing unit performs time-dimension alignment on all sensing data using the timestamps as an index. Only after timestamp alignment can the multi-source sensing data enter the subsequent fusion processing flow.

[0053] The time synchronization mechanism adopts a hardware-triggered synchronization method, that is, the clock source simultaneously sends hardware trigger signals to each sensing device, triggering each device to perform data acquisition at the same time, thereby controlling the synchronization error between multi-source sensing data to within 1 millisecond.

[0054] The basis for setting the synchronization error tolerance at 1 millisecond is as follows: When a vehicle travels at a typical urban road speed of 60 kilometers per hour, the vehicle displacement within a 1-millisecond time window is approximately 1.67 centimeters. This displacement, relative to the pixel displacement of the target in the camera's field of view, is less than the actual physical size corresponding to a single pixel. Therefore, a synchronization accuracy of 1 millisecond is sufficient to ensure that the spatial correspondence between different sensor data at the same moment does not undergo perceptible misalignment during subsequent fusion processing.

[0055] The method for determining the actual synchronization error is as follows: A synchronization monitoring timer is maintained within the edge AI computing unit. Any randomly selected sensing device is used as a reference device, and its data timestamp is recorded. The timestamps of all other sensing devices are also recorded synchronously. The differences between the timestamps of the other devices and the timestamp of the reference device are calculated one by one, and the absolute value is the maximum value among these differences. The measured synchronization error is compared with a 1-millisecond threshold. If the measured value is less than or equal to the threshold, the current synchronization accuracy is deemed to meet the requirements. If the measured value is greater than the threshold, the edge AI computing unit sends a synchronization accuracy adjustment command to the clock source. The clock source responds to the synchronization accuracy adjustment command by shortening the period of the hardware trigger signal or increasing its own clock frequency until the measured synchronization error falls back to within the 1-millisecond threshold.

[0056] Step S205 enables all the various sensing devices distributed throughout the vehicle to operate under a unified time reference, precisely aligning the data collected independently in time. Accurate temporal registration of multi-source data is a prerequisite for subsequent cross-modal fusion inference by the AI ​​model. If there is a temporal discrepancy, the association between visual targets and radar targets will be misaligned, directly affecting the accuracy of scene understanding.

[0057] In summary, after step S2, a comprehensive three-dimensional perception network covering both the external environment and the interior space of the vehicle is completed. This perception network collects raw data from multiple dimensions, including visual images, radar echoes, auxiliary lighting, and vital signs, and ensures the consistency of data across these dimensions over time through a time synchronization mechanism, providing a data source for subsequent scene understanding and feature extraction by the edge AI computing unit.

[0058] The next step, S3, involves deploying a lightweight deep learning model locally on the edge computing unit and using this model to perform scene understanding and feature extraction on multi-source perception data. This includes the following sub-steps.

[0059] Step S301: Configure the neural network accelerator hardware inside the automotive-grade edge AI computing unit. The neural network accelerator adopts a dedicated computing architecture design, which includes multiple parallel computing units, a dedicated storage buffer, and a data transfer engine. The parallel computing units are responsible for performing matrix multiplication and addition operations for each layer of the neural network. The dedicated storage buffer is used to cache intermediate layer feature map data to reduce the number of accesses to off-chip memory. The data transfer engine is responsible for transferring data between the parallel computing units and the dedicated storage buffer.

[0060] The neural network accelerator performs instruction-level optimizations for convolutional neural networks, recurrent neural networks, and Transformer architectures, enabling various mainstream deep learning model structures to achieve high computational efficiency on the neural network accelerator.

[0061] The neural network accelerator supports mixed computation using two low-precision data formats: INT8 and FP16. INT8 is used for forward propagation computation during the inference phase, while FP16 is used for specific computational layers requiring higher numerical precision. By employing mixed computation with these low-precision data formats, inference accuracy remains essentially unchanged, while computational throughput is improved and power consumption is reduced.

[0062] Step S302: Perform lightweight compression processing on the AI ​​model of the scene understanding engine. Before deployment, the AI ​​model of the scene understanding engine needs to go through three compression processes: model quantization, pruning, and knowledge distillation.

[0063] The model quantization process converts the weight parameters and activation values ​​of each network layer in the original model from 32-bit floating-point representation to 8-bit integer representation. After the conversion, the AI ​​model of the scene understanding engine performs integer inference operations on the neural network accelerator.

[0064] In the pruning process, for each convolutional kernel and each fully connected neuron in the scene understanding engine's AI model, the contribution of that kernel or neuron to the model's final output accuracy is calculated. The contribution is calculated by successively removing individual convolutional kernels or neurons from the standard validation dataset, recording the decrease in accuracy of the scene understanding engine's AI model on the validation set before and after removal, and this decrease value is the contribution score of the removed object.

[0065] The contribution scores of all convolutional kernels and neurons are sorted in ascending order, and the bottom 20% of objects are identified as redundant. The weight parameters of the redundant objects are set to zero and removed from the AI ​​model computation graph of the scene understanding engine.

[0066] The 20% redundancy threshold was determined based on the following criteria. A grid search experiment was conducted on a typical in-vehicle scenario dataset to test the model accuracy preservation and compression ratio under multiple candidate thresholds. With the constraint that the compressed accuracy should not be less than 98% of the original model's accuracy, the threshold corresponding to the maximum compression ratio that meets this accuracy constraint was selected. Experiments determined this threshold to be 20%.

[0067] The knowledge distillation process uses a pre-trained large teacher model as guidance, employing the soft labels output by the teacher model on a standard dataset as additional supervision signals to train a lightweight student model. The loss function of the knowledge distillation process consists of two parts.

[0068] The first part is the cross-entropy loss between the student model's predictions and the true labels, used to ensure the student model's basic classification ability. The second part is the KL divergence loss between the student model's predictions and the soft labels output by the teacher model, used to bring the student model's output probability distribution closer to the teacher model's output probability distribution. The two parts of the loss are weighted and summed according to preset weights to obtain the total loss. The student model is iteratively trained with the goal of minimizing the total loss, so that the lightweight student model inherits the class discrimination ability and generalization performance of the teacher model.

[0069] During the distillation training process, the training set images are input into the teacher model and the student model respectively. The soft labels output by the teacher model and the prediction results output by the student model are used together to calculate the loss function. Gradient backpropagation only updates the parameters of the student model, while the parameters of the teacher model remain frozen.

[0070] After the above three compression processes, the number of AI model parameters of the scene understanding engine is reduced to less than one-tenth of the original model, and the amount of computation is reduced to less than one-fifth of the original model, meeting the requirements for real-time operation on the automotive-grade edge AI computing unit embedded hardware platform.

[0071] Step S303: The multi-source sensing data is preprocessed and input into the AI ​​model of the scene understanding engine for inference. After receiving multi-source data from various sensing devices, the edge AI computing unit first executes the data preprocessing process. The preprocessing process includes four operations: image distortion correction, color space conversion, size normalization, and data standardization.

[0072] Image distortion correction involves inversely transforming the original image using the camera's intrinsic parameter matrix and distortion coefficients to eliminate image stretching and bending caused by lens distortion. Color space conversion transforms the original color space output by the camera into the color space used by the AI ​​model of the scene understanding engine during training.

[0073] Size normalization scales images of different resolutions and aspect ratios to a fixed size required by the AI ​​model input layer of the scene understanding engine. Data standardization centers and scales the pixel values ​​of each channel of the normalized image to make its distribution conform to the data distribution during the training of the AI ​​model of the scene understanding engine.

[0074] The preprocessed perception data is converted into a standard tensor format and fed into the neural network accelerator as input data for the AI ​​model of the scene understanding engine to perform inference calculations.

[0075] Step S304: Obtain the scene understanding results and feature vector representations output by the AI ​​model of the scene understanding engine. After the AI ​​model of the scene understanding engine completes inference, it outputs two types of information. The first type is the structured scene understanding results, including object detection boxes and their corresponding category labels, semantic segmentation masks, and instance tracking identifiers. The object detection boxes are rectangular boxes that mark the spatial location of each pedestrian, vehicle, traffic sign, and obstacle in the image. The rectangles are defined by four parameters: the horizontal and vertical coordinates of the upper left corner, the width of the box, and the height of the box. The category labels indicate the category to which each detected object belongs.

[0076] The semantic segmentation mask is a pixel-level category matrix with the same resolution as the input image, where each element represents the semantic category of that pixel. Instance tracking identifiers are unique numbers assigned to each detected target, used to associate the same target across multiple frames. The second category consists of high-dimensional semantic feature vectors extracted by a deep network. These vectors are represented as floating-point arrays and can be used for subsequent retrieval matching and similarity calculation.

[0077] In step S305, the inference results are processed by multiple functional sub-modules within the scene understanding engine. The scene understanding engine includes an object detection sub-module, a semantic segmentation sub-module, a scene classification sub-module, and an action recognition sub-module.

[0078] The object detection submodule parses the object detection boxes and class labels, and outputs the bounding box coordinates, class confidence, and attribute label for each object in the image.

[0079] The semantic segmentation submodule parses the semantic segmentation mask and outputs spatial distribution information of different semantic categories such as road areas, passable areas, buildings, and vegetation.

[0080] The scene classification submodule performs global pooling on the high-dimensional semantic feature vector of the entire image and inputs it into the fully connected classification layer, outputting high-level semantic labels for weather condition type, road type, and geographical location type.

[0081] The behavior recognition submodule performs temporal action analysis on pedestrian or vehicle passenger targets detected in multiple consecutive frames of images. It extracts the keypoint position change sequences of each target in consecutive frames and inputs these sequences into a temporal convolutional network for classification. The temporal convolutional network outputs a probability vector with a dimension equal to the number of action categories. Each element in the probability vector has a value between 0 and 1, and the sum of its elements equals 1. The category corresponding to the highest probability value in the probability vector is determined as the action category of that target; this highest probability value is the confidence score.

[0082] The above four sub-modules are scheduled and executed in a pipelined parallel manner on the neural network accelerator. That is, the inference tasks of different sub-modules are executed overlappingly on different parallel computing units of the neural network accelerator in sequence according to the data flow.

[0083] In summary, after step S3, the in-vehicle dual-sided LED advertising display system completes the localized deployment of the scene understanding engine on the edge computing unit. The scene understanding engine utilizes neural network accelerator hardware and lightweight AI models to convert the multi-source perception data collected in step S2 into semantic-level understanding results of the external environment and the state of passengers inside the vehicle in real time. The structured environmental information and passenger feature information output by the scene understanding engine will serve as input to the advertising content strategy generation module in step S4.

[0084] Regarding step S4 in the method of this embodiment, personalized advertising content strategies are generated for the exterior display surface and the interior display surface based on AI analysis results, specifically including the following sub-steps.

[0085] Step S401: The vehicle exterior advertising content strategy generation unit receives the vehicle exterior scene analysis results output by the scene understanding engine.

[0086] The scene understanding engine outputs vehicle exterior scene analysis results comprising four dimensions: pedestrian detection results and attribute features, environmental weather conditions, road congestion level, and surrounding building types. Pedestrian detection results include the number of detected pedestrians, their spatial location, walking direction, and dwell state. Pedestrian attribute features include gender and age group estimation, clothing style analysis, and gaze direction detection, used to construct advertising audience profiles.

[0087] Environmental weather conditions include weather types such as sunny, cloudy, rainy, snowy, and hazy, as well as light intensity levels. Road congestion levels are categorized into three traffic statuses: smooth flow, slow traffic, and congested. Surrounding building types include commercial areas, residential areas, office areas, hospitals, schools, and other similar locations.

[0088] In step S402, the vehicle exterior advertising content strategy generation unit assesses the potential audience size based on the number and distribution density of pedestrians.

[0089] The vehicle-mounted advertising content generation unit counts the total number of pedestrians detected in the current image frame and calculates the pedestrian distribution density on the image plane. The image is divided into multiple equally sized grid regions, and the number of pedestrian target center points falling into each grid region is counted. Grids with more than a threshold number of pedestrians falling into them are considered dense grids, and the proportion of dense grids to the total number of grids is the distribution density.

[0090] If the total number of pedestrians exceeds the first preset threshold and the distribution density exceeds the first preset density threshold, the current environment is determined to have a large potential audience, and brand display advertising content with a wide reach is prioritized. If the total number of pedestrians does not exceed the first preset threshold or the distribution density does not exceed the first preset density threshold, the current audience is determined to be limited, and the screening criteria for brand display advertising are skipped; content matching is directly performed based on pedestrian attribute characteristics.

[0091] Step S403: The vehicle exterior advertising content strategy generation unit performs audience profile matching based on pedestrian attribute characteristics. The unit summarizes and statistically analyzes the attribute characteristics of each pedestrian target, including gender ratio, age group proportion, and clothing style distribution. The statistical results are then matched with the preset target audience profiles of each advertising material in the advertising content library based on similarity.

[0092] Each ad creative in the ad content library is pre-labeled with a target audience profile tag, which includes attribute values ​​in three dimensions: gender orientation, age range, and interest category.

[0093] The gender matching score is calculated as follows: The percentage of females in the current pedestrian group is calculated, and the difference between this percentage and the gender preference value in the target audience profile of the advertising creative is calculated. Subtracting this difference from 1 gives the gender matching score. The age group matching score is calculated as follows: The distribution vector of the current pedestrian group across each age group is calculated, and the cosine similarity is calculated between this vector and the distribution vector of each age group in the target audience profile of the advertising creative. The resulting cosine similarity value is the age group matching score.

[0094] The interest category matching degree is calculated by taking the dot product of the current pedestrian group's clothing style distribution and the interest category distribution in the target audience profile of the advertising material, and then normalizing it to obtain the interest category matching degree.

[0095] The overall matching score is obtained by summing the three values: gender matching, age group matching, and interest category matching. All ad creatives in the ad content library are traversed, and the overall matching score between each ad creative and the current pedestrian group is calculated sequentially. The ad creative with the highest overall matching score is selected as the recommended ad content for the current scenario.

[0096] In step S404, the vehicle exterior advertising content strategy generation unit adjusts the color parameters of the advertising content based on environmental weather conditions and lighting conditions. The unit dynamically adjusts the color saturation and brightness of the selected advertising content based on environmental weather conditions and light intensity levels. When the environmental weather is sunny and the light intensity level is strong light, the saturation value of the original image is increased by 15%, and the brightness value is increased by 10% based on the original image brightness value.

[0097] When the ambient weather conditions are rainy, snowy, or foggy with low light intensity, increase the saturation value by 25% and the brightness value by 20% based on the original image. When the ambient weather conditions are cloudy with normal light intensity, keep the color saturation and brightness of the original image unchanged.

[0098] In step S405, the vehicle-mounted advertising content strategy generation unit adjusts the playback duration and update frequency of the advertising content according to the road congestion level. When the road congestion level is "congested," the playback duration of a single advertising piece is set to 15 seconds, and the update frequency is set to once every 60 seconds.

[0099] When traffic congestion is slow, the playback duration of a single advertisement is set to 10 seconds, and the update frequency is set to once every 40 seconds. When traffic congestion is smooth, the playback duration of a single advertisement is set to 5 seconds, and the update frequency is set to once every 20 seconds. The total number of advertisements played per unit time remains the same under all three configurations.

[0100] In step S406, the in-vehicle advertising content strategy generation unit receives the in-vehicle scene analysis results output by the scene understanding engine. The in-vehicle scene analysis results output by the scene understanding engine include information in four dimensions: passenger count, age distribution of each passenger, emotional state assessment, and gaze behavior detection. The passenger count includes the number of front-seat occupants and rear-seat passengers, as well as their spatial distribution on the seats.

[0101] The age distribution includes estimated proportions for five age groups: children, adolescents, young adults, middle-aged adults, and the elderly. Emotional state assessment includes classification results for four emotion categories: happy, relaxed, tired, and anxious, along with corresponding confidence scores. Gaze behavior detection includes determining whether each passenger's gaze is directed towards the in-vehicle display screen, the duration of gaze, and the location of the gaze area.

[0102] In step S407, the in-vehicle advertising content strategy generation unit determines the target audience for advertising based on the number of passengers and seat distribution. The in-vehicle advertising content strategy generation unit counts the total number of passengers in the vehicle. If the total number of passengers is equal to 1, it is determined that the current scenario is a single-person ride, and that passenger is taken as the sole target audience for advertising. The passenger's age group, emotional state, and gaze behavior characteristics are extracted, and the process proceeds to step S408 to execute the precise advertising recommendation process based on single user characteristics.

[0103] If the total number of passengers is greater than 1, it is determined that the current situation is a multi-person travel scenario. The age group, emotional state and gaze behavior characteristics of each passenger are summarized, the common indicators of each characteristic are calculated, and the process of selecting advertisements based on the common needs of the group is carried out in step S408.

[0104] In step S408, the in-vehicle advertising content strategy generation unit adjusts the language style and visual design elements of the advertising content based on the age distribution of passengers. In a single-passenger scenario, the in-vehicle advertising content strategy generation unit selects the corresponding advertising language style and visual design template based on the current passenger's age range. When the passenger's age range is children or teenagers, an advertising design template with a lively language style, bright colors, and animation elements is selected.

[0105] When targeting middle-aged or elderly passengers, choose an advertising design template with larger font sizes, clear information hierarchy, and high contrast. When targeting young passengers, choose an advertising design template with a fashionable language style and simple visual elements.

[0106] In multi-passenger scenarios, the in-vehicle advertising content strategy generation unit statistically analyzes the age distribution of each passenger, identifying the age group with the highest percentage of passengers as the mainstream age group, and using the advertising language style and visual design template corresponding to this mainstream age group as the basis. Simultaneously, it checks for the existence of a secondary mainstream age group with a percentage exceeding 30%. If such a group exists, design elements corresponding to this secondary mainstream age group are integrated into the base template.

[0107] In step S409, the in-vehicle advertising content strategy generation unit adjusts the type selection of advertising content based on the passenger's emotional state assessment results. The unit calculates the average confidence score for each passenger's emotional state classification and confidence score, and determines the emotional category with the highest average confidence score as the dominant emotional state in the vehicle. When the dominant emotional state is fatigue, light and entertaining advertising content is prioritized.

[0108] When the dominant emotional state is anxiety, prioritize pushing soothing and relaxing advertisements. When the dominant emotional state is natural or happy, push informational or promotional advertisements. If the mean confidence score for each emotional category is below 0.6, the reliability of the current emotional state recognition result is deemed insufficient, the current advertisement type remains unchanged, and no type switching operation is triggered.

[0109] In step S410, the in-vehicle advertising content strategy generation unit adjusts the push priority of advertising content based on the passenger gaze behavior detection results. The in-vehicle advertising content strategy generation unit calculates the comprehensive attention score of each passenger to the in-vehicle display screen based on the direction of each passenger's gaze and the duration of their gaze.

[0110] When a passenger's gaze is directed towards the in-vehicle display screen, a base score of 10 is awarded, with an additional 1 point for every 2 seconds or more of sustained gaze. When a passenger's gaze is not directed towards the in-vehicle display screen, a base score of 0 is awarded. The overall attention score for the in-vehicle display screen is obtained by summing the combined attention scores of all passengers.

[0111] If the overall attention score is above 15, it is determined that the current passenger is paying high attention to the in-vehicle display screen, and the current advertising content is kept at the same priority. If the overall attention score is below or equal to 15, it is determined that the current passenger is not paying enough attention to the in-vehicle display screen, the priority of the currently playing advertising content is reduced, and alternative advertising content is switched in advance.

[0112] In step S411, the exterior advertising content strategy generation unit and the interior advertising content strategy generation unit run independently and in parallel, while sharing the analysis results of the scene understanding engine.

[0113] The exterior advertising content strategy generation unit and the interior advertising content strategy generation unit each run their own strategy generation logic and content selection algorithm. Their decision-making processes are independent and do not interfere with each other. During operation, both generation units read analysis results from the scene understanding engine; the exterior advertising content strategy generation unit reads the exterior scene analysis results, and the interior advertising content strategy generation unit reads the interior scene analysis results.

[0114] After each of the two generation units completes the generation of advertising content strategies, they send the generated advertising content data packets to the same queue to be distributed, waiting for the collaborative control unit to extract the data packets from the queue and perform synchronous distribution processing.

[0115] For step S5 in the method of this embodiment, a dedicated collaborative control unit is designed to synchronously distribute the AI ​​decision results to the dual-sided LED display screen, and is implemented according to the following steps.

[0116] Step S501: Configure the hardware architecture of the collaborative control unit. The collaborative control unit is implemented using a high-performance microprocessor or system-on-a-chip solution. The microprocessor integrates dual display control interfaces, multi-channel high-speed data bus interfaces, and local storage control interfaces.

[0117] The dual display control interfaces connect to the external display control driver and the internal display control driver respectively, supporting concurrent driving of both display outputs. A multi-channel high-speed data bus interface connects to the AI ​​decision module, receiving advertising content data packets output by the AI ​​decision module. A local storage control interface connects to external storage devices, caching advertising content data and display parameter configuration data to be sent.

[0118] Step S502: Establish a display driver interface adapter layer for the dual-sided display screen inside the collaborative control unit. The display driver interface adapter layer encapsulates the driving control details for dual-sided LED displays with different resolutions, refresh rates, and pixel arrangements, and provides a unified software abstraction interface for upper-layer applications.

[0119] The display driver interface adapter layer supports dynamic configuration of display parameters, including online adjustment of display resolution, refresh rate, color depth, brightness contrast, and color temperature gamma curve, and is adapted and configured according to the specifications of the actual deployed LED display module.

[0120] Step S503: Receive and parse the advertising content data packet output by the AI ​​decision module. The collaborative control unit receives the advertising content data packet output by the AI ​​decision module through the high-speed data bus interface. The data packet contains advertising content data for the exterior display surface, advertising content data for the interior display surface, and timestamp information.

[0121] The collaborative control unit parses and verifies the received data packets. The verification method involves checking whether the cyclic redundancy check (CRC) code carried in the data packet matches the locally recalculated checksum. If they match, the data packet is deemed complete, and the advertising content data for the exterior display and the interior display are extracted. If they do not match, a retransmission request is sent to the AI ​​decision-making module.

[0122] The retransmission request carries the original number of the current data packet. After receiving the retransmission request, the AI ​​decision module retrieves the data packet with the corresponding number from its local send cache and retransmits it. If the verification still fails after three consecutive retransmission requests, the collaborative control unit reports a communication link anomaly to the upper-layer system and suspends the update operation of the current advertisement content.

[0123] Step S504: According to the synchronization timing control strategy, a distribution time window is allocated to the two content data streams. The collaborative control unit determines the synchronization start time of the advertising content on both displays based on the timestamp information in the data packets and the current system time. The synchronization timing control strategy is defined as follows: The expected value of the synchronization start time is obtained by adding a fixed transmission link delay budget value to the time when the edge AI computing unit completes the advertising content decision and outputs the data packet.

[0124] The specific components of the transmission link delay budget are as follows: 5 milliseconds for high-speed data bus transmission delay, 3 milliseconds for display driver circuit parsing delay, and 2 milliseconds for LED pixel refresh delay, totaling 10 milliseconds. The values ​​of each delay component are based on the measured values ​​of the frame transmission time of the selected automotive Ethernet physical layer chip, the instruction parsing cycle of the display driver chip, and the column data latching and row scanning cycle of the LED driver circuit under typical operating conditions.

[0125] The collaborative control unit aligns the release time windows of the external and internal displays to the synchronization start time. The time deviation between the two display updates is controlled within the refresh interval of two adjacent frames. For an LED display with a refresh rate of 60Hz, the single frame refresh period is 16.7 milliseconds, and the time deviation does not exceed 16.7 milliseconds.

[0126] Step S505: Convert the parsed content data into a display cache data format that the display driver module can recognize.

[0127] The collaborative control unit performs format conversion operations on the two extracted advertising content data streams. The format conversion process includes image scaling, color space conversion, pixel format conversion, and display buffer filling.

[0128] The image scaling operation scales the original image size to the actual resolution size of the target LED display screen using a bilinear interpolation algorithm. The scaled image corresponds one-to-one with the physical pixel array of the display screen. The color space conversion operation converts the color space of the advertising content data to the color space required by the LED display screen driver chip.

[0129] The pixel format conversion operation converts the color depth and channel arrangement of each pixel to the format required by the display module. The display buffer fill operation fills the converted pixel data into the corresponding address positions in the display buffer according to the row and column scan order of the target display, forming complete frame buffer data.

[0130] Step S506: Send display update commands and content data to the exterior display control driver and the interior display control driver respectively through the display driver interface.

[0131] When the synchronization start time arrives, the collaborative control unit simultaneously sends display update commands and converted frame buffer data to the external display control driver and the internal display control driver through two display control interfaces. After receiving the corresponding frame buffer data, the external display control driver and the internal display control driver transmit the display data to each pixel of the LED matrix through row scanning drive and column data drive circuits, controlling the emission color and brightness of each pixel to achieve the display rendering of the complete advertising image.

[0132] Step S507: Mechanical structure configuration of the double-sided LED display module. The double-sided LED display module uses a double-sided copper-clad flexible circuit board as the electrical connection substrate. LED lamp matrix and driver chip are soldered to the front and back sides of the flexible circuit board, respectively. The display frame is reinforced with a metal profile, which also serves as a heat dissipation path. The front and back sides of the flexible circuit board are electrically interconnected through metallized vias, allowing the LED lamp matrix on both sides to share the same set of drive control signals and power supply network.

[0133] In step S508, the collaborative control unit monitors the operating status of the dual-sided LED display and performs fault diagnosis. The collaborative control unit periodically reads the status register of the display driver chip through the display driver interface. The status register stores the current display operating mode, temperature monitoring data, and power consumption monitoring data. The temperature monitoring data, in degrees Celsius, is collected in real-time by the temperature sensor integrated within the driver chip and written to the status register. The power consumption monitoring data, in milliwatts, is collected in real-time by the current detection circuit integrated within the driver chip and written to the status register.

[0134] The collaborative control unit compares the temperature monitoring data it reads with a temperature threshold of 85 degrees Celsius. If the temperature exceeds 85 degrees Celsius, the brightness output of the display screen is reduced to 60% of the current brightness value. The collaborative control unit also compares the power consumption monitoring data it reads with a power consumption threshold of 30 watts. If the power consumption exceeds 30 watts, the maximum brightness of subsequent advertising content is limited to 80% of the current brightness value.

[0135] When the collaborative control unit detects that the display driver chip returns an error flag, it determines that a display abnormality has occurred. The error flag is defined as bit 3 in the status register being set to 1. If the collaborative control unit does not receive a response after reading the status register twice consecutively, it determines that a communication timeout has occurred. In either of the above situations, the collaborative control unit reports the fault type to the upper-level system and triggers the fault handling procedure.

[0136] The fault handling process includes sending a software reset command to the display driver chip. If the error flag is still read three times in a row after the software reset or the communication timeout occurs, the power output of the corresponding display surface is turned off and an unrecoverable fault is reported to the upper-level system.

[0137] In summary, the in-vehicle dual-sided LED advertising display system completes the construction of a full distribution channel from AI decision results to the physical images on both displays. The collaborative control unit, acting as the intermediate hub connecting the AI ​​decision module and the display driver module, handles the reception and parsing of advertising content data packets, synchronous timing allocation, format conversion processing, and dual-channel synchronous distribution. This distribution channel ensures precise time synchronization between the two displays, enabling the exterior and interior displays to present their respective adapted advertising content simultaneously, providing the display output foundation for the feedback optimization mechanism in step S6.

[0138] Finally, in step S6, a feedback optimization mechanism based on display effect data is established. The feedback optimization module collects display effect data such as the time pedestrians spend outside the vehicle and the interactive behavior of passengers inside the vehicle. The display effect data is used as a feedback signal to input into the online learning module of the AI ​​model to achieve dynamic optimization and adjustment of the AI ​​model parameters. This includes the following sub-steps.

[0139] Step S601: Deploy external display effect data acquisition equipment by installing external passenger flow statistics cameras on the side or roof of the vehicle. The field of view of the external passenger flow statistics cameras covers the pedestrian activity area around the vehicle. The external passenger flow statistics cameras use high-resolution wide-angle lenses and are used with dedicated passenger flow statistics algorithm software to quantitatively count the number of pedestrians, their flow direction, and their dwell time.

[0140] The passenger flow statistics algorithm is based on video stream analysis and employs deep learning-based personnel detection and tracking technology. The algorithm assigns a unique tracking identifier to each pedestrian appearing in the video frame and calculates the dwell time and movement trajectory of each pedestrian within the monitored area. A dwell time exceeding 3 seconds is considered a valid stop, and the timestamp and duration of this event are recorded.

[0141] The vehicle's external pedestrian flow monitoring camera simultaneously tracks pedestrians' facial orientation and gaze behavior. When a pedestrian's gaze is detected as directed towards the external LED display screen, it is recorded as a valid advertising exposure event.

[0142] Step S602: Deploy in-vehicle display effect data acquisition equipment. In-vehicle interactive sensors include three types: touch screen interactive sensors, voice interactive sensors, and gesture recognition sensors.

[0143] Touchscreen interaction sensors are integrated into the surface of the vehicle's LED display screen to detect passenger touch events, including touch location, touch duration, and swipe direction. When a passenger clicks on a hyperlink in advertising content or performs other interactive actions, it is recorded as a valid interaction event.

[0144] Voice interaction sensors collect voice commands and feedback from passengers inside the vehicle, and use voice recognition and semantic understanding technologies to analyze passengers' evaluation and attitude towards advertising content.

[0145] Gesture recognition sensors detect passengers' hand movements using infrared depth cameras or millimeter-wave radar. When a passenger makes a gesture such as pointing at advertising content or zooming in and out of the advertising image, it is recorded as a gesture interaction event.

[0146] In step S603, the data processing and analysis subsystem of the feedback optimization module cleans, aggregates, and analyzes the collected display effect data. The data cleaning process filters out sensor noise and outlier interference. The data aggregation process statistically summarizes the raw data according to time windows or event batches, generating aggregated statistical values ​​for various performance indicators. The data analysis process uses correlation analysis and attribution analysis to evaluate the relationship between various advertising content and display effects, identify the common characteristics of high-performing advertising content, and identify areas for improvement for low-performing advertising content.

[0147] After the above processing, training sample data is generated for online learning by the AI ​​model. Each training sample contains an input feature vector and a corresponding effect label. The input feature vector is composed of semantic features from various dimensions of the current scene understanding results, including pedestrian detection results and attribute features, environmental weather condition encoding, road congestion level encoding, number of passengers in the vehicle, and age distribution. The effect label is a scalar value, obtained by weighted summation of the normalized value of the duration of dwelling outside the vehicle and the normalized value of the number of interaction events inside the vehicle, used to quantitatively evaluate the display effect of the current advertising content in the corresponding scene.

[0148] In step S604, the online learning module of the AI ​​model fine-tunes and updates the model parameters based on the latest performance data provided by the feedback optimization module. The online learning module of the AI ​​model adopts an incremental learning strategy. The online learning module first adds the latest performance data samples to the model training dataset, and then performs several rounds of mini-batch gradient descent update iterations to adjust the mapping relationship parameters between scene understanding and ad content selection, so that the model learns the latest environment-ad performance correlation pattern.

[0149] The specific process of mini-batch gradient descent update iteration is as follows: Each time, a fixed number of samples are randomly selected from the updated training dataset, and the gradient of the loss function with respect to each model parameter is calculated. The loss function uses mean squared error; the input to the loss function is the predicted value of the model under the current scene features, and the output is the square of the difference between the predicted value and the actual label. After averaging the gradients of the sampled samples, each model parameter is adjusted by one step along the gradient descent direction. When the decrease in the loss function value is less than a preset threshold after several consecutive iterations, the model is considered to have converged, and the current batch update iteration stops.

[0150] Step S605: The online learning module adopts a conservative update strategy to avoid overfitting of the model parameters. The conservative update strategy is as follows: during each parameter update, the update step size of the model parameters is limited to within a preset upper limit. The suggested update amount in the current gradient direction is calculated. If the absolute value of the suggested update amount exceeds the preset upper limit, the actual update amount is compressed to the preset upper limit.

[0151] Simultaneously, an elastic weight consolidation mechanism is introduced to record the historical importance of each parameter in the model, applying additional update penalties to parameters with higher historical importance. Historical importance is measured by the cumulative value of the squared gradients of each parameter during historical training; a larger cumulative squared gradient value indicates a higher degree of participation and importance of the parameter in historical tasks. The update penalty is implemented by multiplying the historical importance by a penalty coefficient and then adding it to the loss function, thus subjecting important parameters to additional constraints during updates. Through this conservative update strategy, the model does not experience drastic parameter drift on a small number of new samples, while retaining the effective knowledge learned from historical data.

[0152] In summary, after step S6, for example, the system acquires pedestrian dwell time and exposure data for exterior advertisements through external passenger flow statistics cameras, and acquires passenger click, voice, and gesture interaction data for in-vehicle advertisements through in-vehicle interaction sensors. After cleaning and aggregation by the data processing and analysis subsystem, training samples are generated. Then, the online learning module of the AI ​​model continuously updates the model parameters using a mini-batch gradient descent method. This feedback optimization mechanism enables the system to continuously improve itself based on real advertising display effect data, forming a continuous iterative cycle of perception, decision-making, display, feedback, and optimization.

[0153] To facilitate understanding, the following detailed explanation of the complete workflow of the vehicle-mounted dual-sided LED advertising display method with edge AI computing power, using a specific application example, is provided below.

[0154] Imagine a commercial display vehicle equipped with the system provided by this invention driving on a road in an urban commercial area. The vehicle is equipped with dual-sided LED display modules on both sides and the roof, which can simultaneously display advertising content to pedestrians outside the vehicle and passengers inside the vehicle.

[0155] The vehicle's external environment perception camera and millimeter-wave radar continuously collect information about the vehicle's surrounding environment. After the collected data is analyzed and reasoned by the AI ​​model of the scene understanding engine, it is detected that there are multiple pedestrians waiting for the pedestrian crossing signal at the zebra crossing ahead. The analysis results of the forward-looking camera image show that there are about eight pedestrians in the area. The pedestrian attribute analysis shows that four of them are young professional women, three are middle-aged men, and one is an elderly woman. At the same time, the scene classification results output by the scene understanding engine detect that the weather conditions of this section of road are sunny, the light intensity is moderate, and the road congestion level is smooth.

[0156] The vehicle exterior advertising content generation unit comprehensively analyzes the vehicle exterior scene results output by the scene understanding engine. Based on the number and density of pedestrians, it determines that the audience is large in the current environment, with young professional women as the main group. Based on the pedestrian attribute matching results, it decides to push a cosmetics brand's promotional advertisement to the vehicle exterior display surface. In step S404, the color parameters of the advertisement are dynamically adjusted according to the current sunny weather and moderate light intensity, using high-saturation colors and dynamic display modes adapted to the current environment to enhance visual appeal.

[0157] Meanwhile, the in-vehicle environment perception infrared camera collects image data of passengers inside the vehicle. After analysis by the behavior recognition submodule of the scene understanding engine, it is detected that there is a passenger sitting in the front passenger seat. Passenger age group analysis shows that the passenger is a young male, emotional state assessment shows that the passenger is in a relaxed state, and gaze behavior detection shows that his gaze is continuously directed towards the in-vehicle display screen.

[0158] The in-vehicle advertising content generation unit makes a comprehensive judgment based on the in-vehicle scene analysis results output by the scene understanding engine. Based on the passenger's age group and emotional state, it determines that the current passenger is paying close attention to the in-vehicle display screen and decides to push a promotional advertisement of a new energy vehicle brand to the in-vehicle display screen. The advertisement highlights the vehicle's performance and technological configuration to match the preferences of the young male target audience.

[0159] The collaborative control unit synchronously distributes the two advertising content decisions generated by the exterior advertising content strategy generation unit and the interior advertising content strategy generation unit to the exterior display surface control driver and the interior display surface control driver, so that the exterior display screen and the interior display screen start displaying their respective advertising content at the same time.

[0160] The vehicle-exterior pedestrian flow statistics camera continuously monitors the lingering and looking behavior of pedestrians in the zebra crossing area. The statistics show that during the advertising display period, six pedestrians lingered for more than the 3-second effective lingering threshold set in step S601. Among them, the facial orientation of three pedestrians showed that they were paying attention to the advertisement on the vehicle-exterior display screen, forming three effective advertising exposure events.

[0161] The data processing and analysis subsystem of the feedback optimization module records the aforementioned display effect data as training samples and inputs them into the online learning module of the AI ​​model. Based on the incremental learning and conservative update strategies disclosed in steps S604 to S605, the online learning module uses the effect label obtained by weighted summation of the normalized value of the vehicle's dwell time and the normalized value of the exposure event count as the optimization objective. It then performs small-batch gradient descent update iterations to adjust the mapping relationship parameters between scene understanding and ad content selection. The model thus learns that pushing beauty ads to young women in scenarios like waiting at zebra crossings in commercial areas can obtain higher pedestrian attention, and will prioritize recommending similar ad content when encountering similar scenarios in the future.

[0162] The in-car touch sensor detected that a passenger clicked on an advertisement on the in-car display screen, viewing a vehicle configuration details page within the advertisement. The feedback optimization module also recorded this interaction data as a training sample and input it into the AI ​​model's online learning module. The online learning module incorporated the normalized count of in-car interaction events into the effect label calculation, using passenger clicks, voice commands, and gesture interactions as positive feedback signals to update model parameters. The model thus learned the preference characteristics of young male passengers for vehicle performance display advertisements, and will prioritize recommending advertisements matching their preferences for this passenger in subsequent trips.

[0163] On the other hand, the in-vehicle dual-sided LED advertising display system with edge AI computing power disclosed in this invention includes: The automotive-grade edge AI computing unit has a built-in neural network accelerator, general-purpose computing processor and high-speed communication interface, which is used to deploy scene understanding engine and perform AI inference on multi-source perception data. The perception module includes onboard cameras, millimeter-wave radar, and infrared sensors deployed on the exterior and interior of the vehicle body to collect data on the external environment and passenger data. The time synchronization module is used to synchronize the data collected by the vehicle camera, millimeter-wave radar and infrared sensor through a unified clock source; The scene understanding engine is deployed locally on the automotive-grade edge AI computing unit. It includes a lightweight and compressed AI model for the scene understanding engine, which is used to perform inference on multi-source perception data after time synchronization and output scene understanding results. The vehicle exterior advertising content strategy generation unit is used to receive the vehicle exterior scene analysis results output by the scene understanding engine, and generate advertising content strategies for the vehicle exterior display surface based on pedestrian detection results and attribute characteristics, environmental weather conditions, and road congestion levels. The in-vehicle advertising content strategy generation unit receives the in-vehicle scene analysis results output by the scene understanding engine and generates advertising content strategies for the in-vehicle display surface based on the statistics of the number of passengers in the vehicle, age distribution, emotional state assessment, and gaze behavior detection. The collaborative control unit has a built-in dual display control interface and a multi-channel high-speed data bus interface. It is used to receive advertising content data packets output by the AI ​​decision module, align the delivery time windows of the exterior display and the interior display to the same synchronous start time, and send display update instructions and frame buffer data to the exterior display control driver and the interior display control driver synchronously through the dual display control interface respectively. A dual-sided LED display module includes an exterior LED display and an interior LED display. The exterior LED display is controlled by the exterior display surface to display advertising content, and the interior LED display is controlled by the interior display surface to display advertising content. The feedback optimization module, including an external passenger flow statistics camera and an in-vehicle interactive sensor, is used to collect display effect data and generate training samples through the data processing and analysis subsystem. These samples are then used by the online learning module of the scene understanding engine's AI model to perform incremental learning and update the model parameters.

[0164] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention. Therefore, the embodiments should be regarded as exemplary and non-limiting in all respects.

[0165] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. A method for displaying vehicle-mounted dual-sided LED advertisements with edge AI computing power, characterized in that, include: Establish a real-time communication link between an automotive-grade edge AI computing unit and a dual-sided LED display screen; Deploy vehicle-mounted cameras, millimeter-wave radar, and infrared sensors to construct a three-dimensional perception network covering the external environment and the interior space of the vehicle, and synchronize the data collected by the vehicle-mounted cameras, millimeter-wave radar, and infrared sensors through a unified clock source; A scene understanding engine is deployed locally in the automotive-grade edge AI computing unit. The scene understanding engine includes a lightweight compressed AI model of the scene understanding engine. Multi-source perception data that has been synchronized with time is input into the AI ​​model of the scene understanding engine to perform inference and obtain scene understanding results. Based on the scenario understanding results, differentiated advertising content strategies are generated for the exterior and interior display surfaces of the vehicle, respectively. The collaborative control unit synchronously distributes the advertising content data packets corresponding to the differentiated advertising content strategy to the exterior display control driver and the interior display control driver. The exterior display control driver and the interior display control driver then drive the exterior LED display and the interior LED display to display the advertising content, respectively.

2. The method according to claim 1, characterized in that, The scene understanding engine's AI model performs lightweight compression processing, including: Model quantization is performed on the weight parameters and activation values ​​of each network layer in the AI ​​model of the scene understanding engine, converting the 32-bit floating-point representation into an 8-bit integer representation; For each convolutional kernel and each fully connected layer neuron in the AI ​​model of the scene understanding engine, a contribution score is calculated. The contribution score is calculated as follows: a single convolutional kernel or a single neuron is removed successively from the standard validation dataset, and the accuracy decrease of the AI ​​model of the scene understanding engine on the validation set before and after the removal is recorded. The accuracy decrease value is used as the contribution score of the removed object. The contribution scores of all convolutional kernels and all neurons are sorted in ascending order. The last 20% of objects are identified as redundant objects. The weight parameters of the redundant objects are set to zero and removed from the computation graph of the AI ​​model of the scene understanding engine. Using a pre-trained large teacher model as a guide, the soft labels output by the pre-trained large teacher model on a standard dataset are used as additional supervision signals to train a lightweight student model.

3. The method according to claim 1, characterized in that, The scene understanding engine includes a behavior recognition submodule. The behavior recognition submodule performs temporal action analysis on pedestrian targets or in-vehicle passenger targets detected in multiple consecutive frames of images, including: Extract the sequence of joint position changes for each target in consecutive frames; The sequence of changes in the joint positions is input into a temporal convolutional network for classification. The temporal convolutional network outputs a probability vector with a dimension equal to the number of action categories. The value of each element in the probability vector is between 0 and 1, and the sum of the elements is equal to 1. The category corresponding to the maximum probability value in the probability vector is determined as the action category of the target. The maximum probability value in the probability vector is the confidence score.

4. The method according to claim 1, characterized in that, The strategy for generating advertising content for the vehicle exterior display surface based on the scene understanding results includes: The attribute characteristics of each pedestrian target are summarized and statistically analyzed, and the statistical results are matched with the target audience profiles preset in each advertising material in the advertising content library. The percentage of females in the current pedestrian group is statistically analyzed. The difference between the percentage of females and the gender tendency value in the target audience profile of the advertising material is calculated. The gender matching degree is obtained by subtracting the difference from 1. The distribution vector of the current pedestrian group in each age group is statistically analyzed, and the cosine similarity is calculated with the distribution vector of each age group in the target audience profile of the advertising material. The resulting cosine similarity value is the age group matching degree. The dot product of the current pedestrian group's clothing style distribution and the interest category distribution in the target audience profile of the advertising material is calculated, and the interest category matching degree is obtained after normalization. The overall matching score is obtained by summing the gender matching score, the age group matching score, and the interest category matching score. Iterate through all ad creatives in the ad content library and select the ad creative with the highest overall match score as the recommended ad content.

5. The method according to claim 1, characterized in that, The strategy for generating advertising content for the in-vehicle display surface based on the scene understanding results includes: The overall attention score of each passenger to the in-vehicle display screen is calculated based on the direction of their gaze and the duration of their gaze. A base score of 10 points is given when a passenger's gaze is directed toward the in-vehicle display screen, with an additional 1 point for every 2 seconds or more of sustained gaze; a base score of 0 points is given when a passenger's gaze is not directed toward the in-vehicle display screen. The overall attention score is obtained by summing the combined attention scores of all passengers in the vehicle. When the overall attention score is higher than 15, it is determined that the current passengers in the vehicle are paying high attention to the in-vehicle display screen, and the current advertising content is pushed with the same priority. When the overall attention score is lower than or equal to 15, it is determined that the current passengers in the vehicle are not paying enough attention to the in-vehicle display screen, the push priority of the currently playing advertising content is reduced, and the alternative advertising content is switched in advance.

6. The method according to claim 1, characterized in that, The strategy for generating advertising content for the vehicle exterior display surface based on the scenario understanding results also includes: The color parameters of the advertisement content are adjusted according to environmental weather conditions and lighting conditions; When the ambient weather conditions are sunny and the light intensity level is strong, the saturation value of the original image is increased by 15% and the brightness value of the original image is increased by 10%; when the ambient weather conditions are rainy, snowy, or hazy and the light intensity level is low, the saturation value of the original image is increased by 25% and the brightness value of the original image is increased by 20%; when the ambient weather conditions are cloudy and the light intensity level is normal, the color saturation and brightness of the original image remain unchanged. The playback duration and update frequency of the advertisement content are configured according to the level of road congestion; When the road congestion level is "congested," the playback duration of a single advertisement is set to 15 seconds, and the update frequency is set to once every 60 seconds; when the road congestion level is "slow," the playback duration of a single advertisement is set to 10 seconds, and the update frequency is set to once every 40 seconds; when the road congestion level is "smooth," the playback duration of a single advertisement is set to 5 seconds, and the update frequency is set to once every 20 seconds.

7. The method according to claim 1, characterized in that, Before inputting the time-synchronized multi-source perception data into the AI ​​model of the scene understanding engine to perform inference, the process also includes a preprocessing procedure for the multi-source perception data. The preprocessing procedure includes: Image distortion correction uses the camera intrinsic parameter matrix and distortion coefficients to perform an inverse transformation on the original image, eliminating image stretching and bending caused by lens distortion. Color space conversion converts the original color space output by the camera into the color space used by the AI ​​model of the scene understanding engine during training. Size normalization scales images of different resolutions and aspect ratios to a fixed size required by the AI ​​model input layer of the scene understanding engine. Data standardization involves centering and scaling the pixel values ​​of each channel in the normalized image to ensure that the distribution of pixel values ​​in each channel conforms to the data distribution during the training of the AI ​​model of the scene understanding engine.

8. The method according to claim 1, characterized in that, The step of synchronously distributing advertising content data packets to the exterior display control driver and the interior display control driver via the collaborative control unit includes: The collaborative control unit receives advertising content data packets output by the AI ​​decision module through a high-speed data bus interface. The advertising content data packets include advertising content data for the exterior display surface, advertising content data for the interior display surface, and timestamp information. The collaborative control unit determines the synchronization start time of the advertising content on the dual-sided LED display screen based on the timestamp information in the advertising content data packet and the current system time. The synchronization start time is based on the time when the automotive-grade edge AI computing unit completes the advertising content decision and outputs the advertising content data packet, plus the transmission link delay budget value. The transmission link delay budget is composed of: a high-speed data bus transmission delay of 5 milliseconds, a display driver circuit parsing delay of 3 milliseconds, and an LED pixel refresh delay of 2 milliseconds. The sum of the high-speed data bus transmission delay, the display driver circuit parsing delay, and the LED pixel refresh delay is 10 milliseconds. The collaborative control unit aligns the release time windows of the exterior display and the interior display to the synchronization start time, and controls the time deviation of the two display updates to be within the display refresh interval of two adjacent frames. When the synchronization start time arrives, the collaborative control unit simultaneously sends display update commands and frame buffer data to the external display control driver and the internal display control driver through the dual display control interface.

9. The method according to claim 1, characterized in that, This also includes establishing a feedback optimization mechanism based on display performance data: Deploy external passenger flow statistics cameras to quantitatively count the number of pedestrians, their flow direction, and their dwell time. Deploy in-vehicle interaction sensors, including touch screen interaction sensors, voice interaction sensors, and gesture recognition sensors; The online learning module of the AI ​​model of the scene understanding engine adopts an incremental learning strategy, adding the latest performance data samples to the model training dataset, performing small-batch gradient descent updates and iterations, and adjusting the mapping relationship parameters between scene understanding and ad content selection. Each time, a fixed number of samples are randomly drawn from the updated training dataset, and the gradient of the loss function with respect to each model parameter is calculated. The loss function adopts the mean squared error form. The input of the loss function is the prediction effect value of the AI ​​model of the scene understanding engine under the current scene features. The output of the loss function is the square of the difference between the prediction effect value and the actual effect label. After averaging the gradients of the extracted samples, adjust the parameters of each model by one step size along the gradient descent direction; When the decrease in the loss function value is less than a preset threshold after multiple consecutive iterations, it is determined that the AI ​​model of the scene understanding engine has converged, and the update iteration of the current batch is stopped. The online learning module of the AI ​​model of the scene understanding engine adopts a conservative update strategy, which limits the update step size of the model parameters of the AI ​​model of the scene understanding engine to within a preset upper limit. Calculate the suggested update amount in the current gradient direction. If the absolute value of the suggested update amount exceeds a preset upper limit, compress the actual update amount to the preset upper limit. At the same time, an elastic weight consolidation mechanism is introduced to record the historical importance of each parameter in the AI ​​model of the scene understanding engine. The historical importance is measured by the cumulative value of the squared gradient of each parameter during the historical training process. The historical importance is multiplied by a penalty coefficient and then added to the loss function, so that parameters with higher historical importance are subject to additional constraints when updating.

10. A vehicle-mounted dual-sided LED advertising display system with edge AI computing power, characterized in that, include: The automotive-grade edge AI computing unit has a built-in neural network accelerator, general-purpose computing processor and high-speed communication interface, which is used to deploy scene understanding engine and perform AI inference on multi-source perception data. The perception module includes onboard cameras, millimeter-wave radar, and infrared sensors deployed on the exterior and interior of the vehicle body to collect data on the external environment and passenger data. The time synchronization module is used to synchronize the data collected by the vehicle camera, the millimeter-wave radar and the infrared sensor using a unified clock source. The scene understanding engine, deployed locally on the automotive-grade edge AI computing unit, includes a lightweight and compressed AI model of the scene understanding engine, which is used to perform inference on multi-source perception data after time synchronization and output scene understanding results. The vehicle exterior advertising content strategy generation unit is used to receive the vehicle exterior scene analysis results output by the scene understanding engine, and generate the advertising content strategy for the vehicle exterior display surface based on pedestrian detection results and attribute characteristics, environmental weather conditions, and road congestion. The in-vehicle advertising content strategy generation unit is used to receive the in-vehicle scene analysis results output by the scene understanding engine, and generate advertising content strategies for the in-vehicle display surface based on the statistics of the number of passengers in the vehicle, age distribution, emotional state assessment and gaze behavior detection. The collaborative control unit has a built-in dual display control interface and a multi-channel high-speed data bus interface. It is used to receive advertising content data packets output by the AI ​​decision module, align the delivery time windows of the exterior display and the interior display to the same synchronous start time, and send display update instructions and frame buffer data to the exterior display control driver and the interior display control driver synchronously through the dual display control interface respectively. A dual-sided LED display module includes an exterior LED display and an interior LED display. The exterior LED display is controlled by the exterior display surface to display advertising content, and the interior LED display is controlled by the interior display surface to display advertising content. The feedback optimization module includes an external passenger flow statistics camera and an in-vehicle interactive sensor, which are used to collect display effect data and generate training samples through the data processing and analysis subsystem. These samples are then used by the online learning module of the AI ​​model of the scene understanding engine to perform incremental learning to update the model parameters.

Images