Train video data adaptive transmission method and system based on edge calculation

Abstract

The invention relates to the technical field of rail transit communication and data processing, discloses a train video data adaptive transmission method and system based on edge computing, and aims to solve the problem of video return failure caused by bandwidth fluctuation in a complex train environment. The scheme is characterized by comprising the following steps: acquiring a train running state, GIS terrain and network parameters in real time; the vehicle-mounted computing equipment is used for carrying out AI target detection to identify video abnormity; predicting a bandwidth trend in combination with GIS attributes; and dynamically generating and executing a transmission strategy, wherein the transmission strategy comprises compression transmission in a normal state, ROI (Region of Interest) enhanced transmission in an abnormal state and key frame priority transmission in a weak network in combination with local storage medium supplementary transmission. According to the application, the bandwidth utilization rate can be optimized, and the continuity and certainty of key security data return in an extreme geographical environment are ensured.

Description

Technical Field

[0001] This invention belongs to the field of rail transit communication and data processing, and specifically relates to an adaptive transmission method and system for train video data based on edge computing. Background Technology

[0002] With the rapid evolution of global rail transit technology, intelligent high-speed rail has become a core component of the modern transportation system. To ensure the absolute safety of high-speed trains in complex operating environments and improve operational efficiency, deploying high-density, full-coverage high-definition surveillance cameras in key areas such as the pantograph, driver's cab, passenger compartment, and exterior of the carriages has become standard practice in intelligent rail transit systems. This massive amount of video data not only carries the task of real-time monitoring of train operation status but also provides crucial data support for subsequent accident tracing, fault early warning, and emergency response. Against this backdrop, how to construct an efficient, stable, and environmentally aware video data feedback system has become one of the core issues in ensuring railway transportation safety.

[0003] In the existing railway communication architecture, the 5G-R (5G for Railway) wireless communication network serves as the primary channel for carrying video services, aiming to provide high-bandwidth, low-latency transmission guarantees. Traditional technical solutions typically employ a full-volume video backhaul mechanism based on a fixed bit rate or preset bandwidth allocation. In this mode, the high-definition video stream acquired by the onboard unit is treated as a homogenized data load and uploaded in real time according to uniform encoding parameters and transmission priorities. Under relatively ideal network conditions and low-speed train operation, this solution can maintain good image quality continuity and real-time performance, ensuring the ground control center's visual control of the train's status.

[0004] However, with the continuous increase in train speeds (such as reaching 300 km / h and above) and the increasing complexity of operating conditions, the aforementioned traditional fixed transmission mechanism has revealed profound inherent limitations at the principle level. Specifically, the Doppler frequency shift effect caused by high-speed movement and frequent base station handovers lead to drastic fluctuations in the physical layer parameters of the wireless channel, resulting in nonlinear jitter in the uplink bandwidth and random packet loss. More seriously, when trains travel through complex geographical environments such as tunnels, deep trenches, or high embankments, the rapid deterioration of the electromagnetic environment often leads to a precipitous attenuation of the network signal. In such an extremely unstable network environment, the full-rate transmission mechanism based on a fixed bit rate, lacking sensitivity to changes in channel capacity, is highly susceptible to buffer overflow at the transmitting end, which manifests at the ground end as severe video tearing, stuttering, or even transmission interruption.

[0005] Furthermore, existing technologies often overlook the semantic value differences in video content at the security level when processing video data. Current solutions for dealing with network congestion often employ a strategy of blindly and indiscriminately reducing global resolution or increasing compression ratio. This "one-size-fits-all" approach presents a significant technical paradox: during the "normal" periods when trains are running smoothly for the vast majority of the time, the system consumes valuable bandwidth resources for the full transmission of redundant data; however, in critical moments such as foreign object intrusion, fires in carriages, or physical altercations, due to limited network bandwidth and the system's inability to recognize the importance of the footage, core security data is often submerged in low-quality bitstreams or key frames are lost due to packet loss. This lack of semantic awareness directly results in ground control centers receiving only blurry or intermittent residual footage when high-definition details are most needed for decision-making, thus creating significant security risks.

[0006] At its core, the existing train video transmission system faces a deep-seated contradiction between limited resources and sudden, unpredictable demands. On the one hand, the spectrum resources of 5G-R private networks are relatively scarce and highly volatile in specific geographical areas; on the other hand, for train security, the value distribution of video data exhibits extremely high sparsity and unevenness. Due to the lack of effective edge sensing and intelligent scheduling methods, existing technical solutions cannot achieve dynamic decoupling and reconstruction of video content value and network channel status at the train end. Furthermore, when facing extreme conditions such as weak network conditions in tunnels, existing technical approaches typically only offer an inefficient compromise between "ensuring image quality" and "maintaining transmission continuity," making it difficult to achieve deterministic protection of critical security information in complex and dynamic environments.

[0007] Therefore, how to combine the real-time processing capabilities of edge computing to achieve multi-dimensional collaborative perception of the operating environment, network status, and video semantics at the front end of the train, and on this basis to build an adaptive transmission system that can automatically adjust the transmission strategy according to the urgency of business and take into account both bandwidth utilization and data security, has become a key challenge and a technical problem that needs to be solved by those skilled in the art. Summary of the Invention

[0008] The technical problem to be solved by this invention is to address the technical defects of 5G-R wireless communication networks in high-speed trains operating at high speeds and in complex geographical environments (such as tunnels, deep trenches, and mountainous areas), which are caused by frequent base station switching, Doppler frequency shift effect, and signal attenuation leading to extremely unstable bandwidth, resulting in video transmission stuttering, loss of critical safety footage, and inability to prioritize real-time transmission of emergencies. The invention provides a train video data adaptive transmission method and system based on edge computing. To achieve the above-mentioned objectives, the present invention adopts the following technical solution: an adaptive transmission method for train video data based on edge computing, the method being applied to a collaborative transmission architecture consisting of an onboard unit and a ground control center, the method comprising the following steps: Step 1: Real-time acquisition of train operation status data and real-time network performance parameters of the currently accessed wireless communication network; wherein, the operation status data includes the train's current real-time speed, real-time geographical location information, and terrain attributes within a predetermined distance ahead of the train based on the geographical location information; the real-time network performance parameters include the uplink available bandwidth of the current channel, packet loss rate, end-to-end network latency, and Doppler frequency shift caused by the high-speed movement of the train; Step 2: Acquiring real-time raw video streams captured by onboard cameras, and performing frame-by-frame analysis of the real-time raw video streams using a target detection model pre-deployed in the onboard computing device, outputting structured video anomaly detection results including anomaly event type labels, anomaly target bounding box position coordinates, and identification confidence scores; Step 3: Constructing a bandwidth prediction model based on the real-time operation status data and the real-time network performance parameters; the bandwidth The prediction model uses a weighted moving average algorithm combined with historical bandwidth performance and introduces an environmental feedforward correction factor determined by the terrain attributes to dynamically predict the current available network bandwidth threshold within a future transmission cycle. Step four: Based on the current available network bandwidth threshold and the semantic information of the video anomaly detection results, a target video adaptive transmission strategy is dynamically generated and executed from multiple preset transmission strategies. Step five: When the video anomaly detection results show an abnormal event and the real-time network performance parameters are lower than the second preset threshold, an extreme environment protection strategy is executed, i.e., key frame images of the abnormal event are captured and alarm text information is generated, which is encapsulated in the control plane channel protocol data unit of the wireless communication network and sent preferentially. At the same time, the high-definition video clip corresponding to the abnormal event is stored in the local storage medium. Step six: The real-time network performance parameters are continuously monitored. When the network performance is detected to recover to the first preset threshold, and the first preset threshold is greater than the second preset threshold, and this state is maintained for a preset stable window duration, the breakpoint resume mechanism is automatically triggered, and the high-definition video clip stored in the local storage medium is retransmitted to the ground control center using the remaining bandwidth.This invention also provides an adaptive transmission system for train video data based on edge computing. The system includes: a status acquisition module for acquiring train operation status data, GIS map data, and physical layer information of the wireless communication network. The status acquisition module has a data alignment unit to ensure consistency of operation status data and network parameters in timestamps across different dimensions; an edge detection module integrated into the onboard computing device, which performs real-time semantic parsing of the video stream and generates a structured anomaly detection report by running a target detection model; a threshold evaluation module for running a bandwidth prediction algorithm, which dynamically calculates the predicted effective transmission bandwidth for the current and future predetermined time periods based on the multi-dimensional data provided by the status acquisition module and a preset environmental feature library; a strategy scheduling module, which switches between a first transmission strategy, a second transmission strategy, or a third transmission strategy based on the bandwidth prediction value and the anomaly detection report through a built-in decision logic matrix and outputs corresponding encoding control parameters; and an execution transmission module, which includes a multi-protocol video encoding unit and a communication scheduling unit, for resampling, ROI extraction encoding, keyframe extraction, or text encapsulation of the video stream according to the instructions of the strategy scheduling module, and pushing the generated data packets to the ground control center via a wireless link. Furthermore, as a preferred technical solution of the present invention, in step one, the process of obtaining the Doppler frequency shift offset includes: the vehicle-mounted integrated sensing unit according to the formula. Calculate the predicted value, where The center frequency of the carrier. For the real-time speed of the train, The speed of electromagnetic wave propagation. The angle between the train's direction of travel and the electromagnetic wave propagation direction is used. The predicted value is fused with the actual received signal frequency deviation fed back by the wireless communication module using Kalman filtering to accurately assess the physical layer stability of the current channel, and the real-time network performance parameters are corrected based on the physical layer stability. Further, as a preferred embodiment of the present invention, the on-board computing device is preferably an on-board edge computing node, which can be equipped with a high-performance neural network processing unit (NPU). The target detection model undergoes pruning and symmetric quantization based on 8 bits or 16 bits before operation, employs a feature extraction network based on depthwise separable convolution, and achieves multi-scale feature fusion through a feature pyramid structure (FPN) to reduce inference latency on the NPU. Further, as a preferred technical solution of the present invention, the process of constructing the bandwidth prediction model specifically includes: smoothing the real-time sampled bandwidth data using a weighted moving average algorithm, wherein the weight of the near-real-time sampled value is set to 0.7 and the weight of the historical average is set to 0.3; introducing an environmental feedforward correction mechanism: based on the geographical location information, when it is determined that the train is about to enter the tunnel area, deep trench section, or tall building distribution area marked by the terrain attribute within a preset distance, the bandwidth prediction model adjusts the current available network bandwidth threshold in advance according to the preset signal attenuation curve. Specifically, the process of dynamically generating the video adaptive transmission strategy preferably includes: when the video anomaly detection result shows that there is no anomaly in the current picture, and the current available network bandwidth threshold is greater than a first preset threshold (e.g., 20Mbps), the first transmission strategy is executed. Under this strategy, edge computing nodes reduce the frame rate of the original video stream to a first preset frame rate (e.g., 5fps) using a time sampler, and downsample the video frames to achieve a first preset resolution. Then, they employ the H.265 standard with a high fixed quantization parameter (QP) for high compression ratio encoding to maintain low bitrate inspection of the train's basic status by the ground center. When the video anomaly detection result indicates an abnormal event, and the current network performance is between a first and a second preset threshold, a second transmission strategy is executed. The system initiates a Region of Interest (ROI) extraction algorithm, and the edge computing nodes divide the video frames into core ROI regions and non-core background regions based on the bounding box coordinates. A high proportion of the bitrate budget is allocated to the core ROI regions, and lossless or high-fidelity encoding is used; simultaneously, spatial domain downsampling or fuzzy compression is performed on the non-core background regions. Finally, the two data sets are merged and transmitted via a 5G-R dedicated high-priority network slice.Furthermore, as the core innovative mechanism of this invention, the specific execution logic of the extreme environment protection strategy is as follows: the edge computing node immediately stops sending continuous video streams, retrieves the original I-frame images before and after the anomaly occurrence from the video stream buffer, and compresses them to a preset size limit; the alarm text information includes timestamp, train number, carriage location, anomaly type, and geographic coordinates; the keyframe images and alarm text information are sent through the physical layer control channel or narrowband emergency communication link; simultaneously, after associating the full-resolution original video segments with a unique event index identifier through the storage controller, they are stored in real time in a high-speed non-volatile memory (NVMe SSD) as the local storage medium. As a refined implementation of this invention, the ROI extraction algorithm also introduces a visual saliency weighting mechanism: when an abnormal target is detected, the visual association area within a preset range around the target is determined by calculating the image contrast, color gradient, and edge density, and these areas are included in the encoding scope of the core ROI area. For example, when an abnormal pantograph sparking is detected, the range is automatically expanded to 50 centimeters around the target to capture more complete physical failure features. Furthermore, at the system architecture level, the edge computing node adopts a dual-buffer queue architecture. The first buffer queue temporarily stores real-time data to be sent, while the second buffer queue, relying on NVMe SSDs, specifically stores video segments to be retransmitted. The policy scheduling module also introduces a dynamic channel quality assessment (CQI) metric as a feedback loop. When the packet loss rate exceeds a preset threshold multiple times consecutively, the encoding bit rate is automatically reduced, and the redundancy of forward error correction (FEC) coding is increased. In addition, the execution sending module supports a reliable transmission protocol based on UDP and a custom ARQ mechanism to provide selective retransmission guarantees. The electronic device involved in this invention includes at least one processor and a memory communicatively connected to the processor. The memory stores computer instructions that can be executed by the processor. When the processor calls and executes the instructions, the system can automatically complete the entire process from state awareness, edge AI detection, bandwidth prediction to policy dynamic scheduling. As a beneficial effect of this invention, this technical solution, by introducing edge computing nodes and intelligent content awareness mechanisms, completely changes the limitations of "content blindness" and "bandwidth rigidity" in traditional train video backhaul. First, regarding bandwidth resource utilization, this invention saves over 70% of ineffective redundant bandwidth through dimensionality reduction and frame extraction transmission under normal anomaly conditions, significantly alleviating the communication pressure on the 5G-R private network in multi-vehicle concurrent scenarios. Second, in terms of security assurance performance, it achieves deep coupling between video content semantics and transmission priority. In the event of sudden anomalies such as foreign object intrusion or fire, the system can ensure the clarity of key details through ROI enhancement technology, or forcibly break through communication bottlenecks in weak network environments through a combination strategy of "text + keyframes," ensuring that the ground control center can obtain alarm information immediately, thus achieving deterministic assurance of critical security data backhaul.Finally, this invention possesses strong environmental adaptability. Through the synergy of GIS location awareness and a breakpoint resume mechanism, it solves the continuity problem of video surveillance services in extreme geographical environments such as tunnels. The core logic of this invention lies in exchanging expensive and unstable spatial wireless bandwidth resources for computing resources at the vehicle edge. By achieving a structured understanding of video content, it enables differentiated services (QoS) based on data value. This constitutes the core technical feature that distinguishes this invention from existing fixed bitrate backhaul technologies, possessing significant inventiveness and non-obviousness. Attached Figure Description

[0009] Figure 1 A flowchart illustrating an adaptive transmission method for train video data based on edge computing, provided in an embodiment of the present invention;

[0010] Figure 2 A structural block diagram of a train video data adaptive transmission system based on edge computing is provided in an embodiment of the present invention.

[0011] Figure 3 This is a schematic diagram of the internal logic processing of the edge detection module in an embodiment of the present invention;

[0012] Figure 4 This is a schematic diagram illustrating the principle of ROI region extraction and video encoding processing in an embodiment of the present invention;

[0013] Figure 5 This is a flowchart illustrating the operation of the third transmission strategy and the breakpoint resumption mechanism in this embodiment of the invention.

[0014] Figure 6 This is a schematic diagram of the structure of an electronic device for performing the method according to an embodiment of the present invention.

[0015] The attached diagram is labeled as follows: 1. Status acquisition module; 2. Edge detection module; 3. Threshold evaluation module; 4. Policy scheduling module; 5. Execution and transmission module; 6. Vehicle-mounted integrated sensing unit; 7. Edge computing node; 8. High-performance neural network processing unit (NPU); 9. High-speed non-volatile memory (NVMe SSD); 10. Processor; 11. Memory; 12. Policy scheduling engine. Detailed Implementation

[0016] The technical solution provided by the present invention will be described in detail below with reference to the accompanying drawings and specific embodiments.

[0017] According to one embodiment of the present invention, a method and system for adaptive transmission of train video data based on edge computing is disclosed. The core logic lies in constructing a closed-loop control system driven by intelligent edge perception at the train end. Through real-time fusion analysis of the train's operating environment, network physical layer performance, and video content semantics, on-demand allocation of communication resources is achieved. In this embodiment, the system is deployed on a high-speed train operating at 350 km / h, and its external communication environment encompasses open fields, deep trenches, long tunnels, and urban high-rise building clusters.

[0018] In its implementation, this invention first achieves synchronous acquisition of multi-source heterogeneous data through an onboard integrated sensing unit 6 deployed on the train. The onboard integrated sensing unit 6 is interconnected with the train's core control computer via an onboard Ethernet or CAN bus. The real-time operational status data it acquires includes the train's current real-time speed and real-time geographic location coordinates from the BeiDou Navigation Satellite System (BDS) or Global Positioning System (GPS) obtained via a multi-mode GNSS receiver. To enable early perception of complex environments, the unit integrates a high-precision Geographic Information System (GIS) database. This database records the terrain attributes of the entire train route with decimeter-level accuracy, including but not limited to precise mileage markers at tunnel entrances, the length distribution inside tunnels, slope heights in deep trench sections, and the overlapping sectors of 5G-R base stations along the route. During train operation, the status acquisition module 1, based on the current geographic location coordinates, pre-retrieves terrain features within a 5-kilometer radius ahead from the GIS database and parameterizes these features as terrain attenuation prediction factors.

[0019] Furthermore, the vehicle-mounted integrated sensing unit 6 extracts real-time network performance parameters of the wireless communication network through the underlying driver interface of the wireless communication module. These parameters include, but are not limited to, the reference signal received power (RSRP), signal-to-interference-plus-noise ratio (SINR), uplink available bandwidth, packet loss rate, and end-to-end network latency of the current channel. Specifically, considering the impact of Doppler frequency shift caused by high-speed movement on inter-symbol interference in 5G-R Orthogonal Frequency Division Multiplexing (OFDM), the system uses the formula... Calculate the predicted value of the Doppler frequency shift. Among them, The center frequency of the carrier (e.g., 4.9 GHz). For the train's real-time speed, At the speed of light, This is the angle between the train's direction of travel and the direction pointed to by the ground base station. This predicted value is fused with the actual frequency deviation fed back by the wireless module using Kalman filtering to accurately assess the physical layer stability of the current channel, providing a basis for correction in subsequent bandwidth predictions.

[0020] In the video perception dimension, this invention acquires real-time raw video streams through onboard high-definition cameras (such as pantograph monitoring units, driver's cab panoramic cameras, and passenger corridor cameras) distributed at key locations on the train. These video streams typically use 4K or 1080P resolution, with a frame rate maintained between 30fps and 60fps. The raw video streams are aggregated to edge computing node 7 via an onboard switch. This edge computing node 7 is equipped with a high-performance neural network processing unit (NPU) 8, capable of providing computing power of no less than 20 TOPS. Inside the edge computing node 7, a deeply optimized lightweight object detection model is pre-loaded and runs.

[0021] Specifically, the lightweight target detection model employs an improved ShuffleNetV2 as the feature extraction backbone network, reduces the number of computational parameters by extensively using depthwise separable convolutions, and utilizes a Feature Pyramid Network (FPN) at the model output to enhance the ability to capture anomalous targets of different sizes. Before deployment, the model undergoes 8-bit or 16-bit symmetric quantization, reducing its inference latency on the embedded NPU to less than 20ms. The edge detection module 2 scans the input video stream frame by frame, detecting objects including but not limited to abnormal sparks in the pantograph area, foreign objects caught in the net, driver fatigue, and sudden violent conflicts inside the passenger compartment. The model outputs a highly structured JSON data packet containing the classification label of the anomalous event, the normalized bounding box coordinates (xmin, ymin, xmax, ymax) of the target in the current frame, and a detection confidence score.

[0022] Subsequently, the system enters the threshold evaluation and bandwidth prediction stage. The threshold evaluation module 3 utilizes the multi-dimensional input provided by the state acquisition module 1 to construct a dynamic bandwidth prediction model. This model employs a weighted moving average (WMA) algorithm, assigning higher weights (e.g., 0.7) to near-real-time sampled values ​​and lower weights (e.g., 0.3) to historical values. More importantly, the model introduces an environmental feedforward correction mechanism: when GIS information indicates that the train is about to enter the tunnel within 500 meters, the model will pre-adjust the current available network bandwidth threshold based on a preset tunnel attenuation profile, even if the RSRP fed back from the physical layer is still high. This location-based prediction effectively avoids transmission link interruptions caused by signal abrupt changes, giving the policy scheduling engine 12 sufficient time to perform encoder parameter switching.

[0023] The strategy scheduling module 4, as the core decision-making unit of the system, receives bandwidth prediction results and video anomaly detection reports. This module integrates a multi-criteria decision matrix. Specifically, when the anomaly confidence level returned by the edge detection module 2 is lower than a preset threshold (e.g., 0.4), and the bandwidth prediction value is greater than 20Mbps, the strategy scheduling module 4 instructs the sending module 5 to activate the first transmission strategy. Under this strategy, the video encoder is configured in "saving mode" and executes a first preset frame rate. Specifically, the original 30fps stream is reduced to 5fps using a time sampler, and the resolution is downsampled from 3840x2160 (4K) to 1280x720 (720P). Simultaneously, the quantization parameter (QP) of the H.265 encoder is set to a large fixed value (e.g., 35-40) to maintain an extremely low output bit rate. At this time, the ground center only receives discontinuous, low-bitrate footage for routine monitoring, and the bandwidth occupancy rate is reduced by more than 95% compared to the original stream.

[0024] When the edge detection module 2 identifies a high-value anomalous event (such as an unidentified object attached to the pantograph area) and the detection confidence exceeds 0.85, while the predicted bandwidth is between 1Mbps and 20Mbps, the system switches to the second transmission strategy. At this stage, the edge computing node 7 initiates the Region of Interest (ROI) extraction logic. Based on the bounding box coordinates returned by the target detection, the system dynamically divides the video frame into core ROI regions and non-core background regions. For core ROI regions, the encoder allocates the majority of the bitrate budget, employs near-lossless quantization parameters (QP values ​​of 15-20), and forces fine-grained encoding of intra-frame prediction blocks; while for background regions, maximized spatial domain downsampling or high-intensity blur filters are used. This content-aware encoding method ensures that the features of anomalous targets (such as cracks on the pantograph) can be clearly identified at the ground end, meeting the requirement for evidence fixation. The processed bitstream is encapsulated in a 5G-R network slice with extremely high scheduling priority or a bearer with QoS level (QCI=1) for transmission.

[0025] In extreme environments, such as when a train enters the central section of a long tunnel or crosses a communication blind spot, if the real-time network performance parameters are below 1Mbps and there are still pending abnormal events, the system triggers the third transmission strategy. The sending module 5 immediately cuts off the transmission of the continuous video stream and instead performs keyframe extraction. The system retrieves the original images (I-frames) of three frames before and after the time of the abnormality from the video stream buffer and compresses them into high-quality JPG images with a size of less than 50KB. Simultaneously, the alarm information is encapsulated into a very short structured text, including geographic coordinates, timestamp, abnormality category, and warning level. This "text + keyframe" combination packet is transmitted through the physical layer control channel or a narrowband emergency communication link (such as a GSM-R backup link), ensuring that the ground control center can still achieve second-level early warning perception even with extremely narrow bandwidth.

[0026] Meanwhile, the high-speed non-volatile memory (NVMe SSD) 9 inside the edge computing node 7 acts as a "black box." The storage controller tags the full-resolution, high-frame-rate raw video data from 30 seconds before to 30 seconds after the anomaly and stores it locally. When the train leaves the weak network area, the threshold assessment module 3 detects that the network performance has recovered to above 20Mbps and remained stable for more than 10 seconds, and the breakpoint resume mechanism is activated. The policy scheduling module 4 sorts the events according to their severity, extracts the segments to be retransmitted from the memory, and uses the remaining idle bandwidth outside the current real-time transmission task to transmit the high-definition segments back through a reliable transmission protocol that supports ARQ (Automatic Repeat Request). After receiving the data, the ground server uses timestamp alignment technology to stitch together the real-time low-definition stream, the emergency image stream, and the subsequently retransmitted high-definition stream to reconstruct a complete view of the entire event process.

[0027] As a specific application example of this invention, the system demonstrated extremely high robustness within an operating segment spanning 1000 kilometers. The following table shows a performance comparison between the technical solution of this invention (example) and a traditional fixed bit rate transmission scheme (comparative example) under different terrain conditions:

[0028] Experimental scenario description Terrain Environment Type Average access bandwidth (Mbps) Transmission strategy Video packet loss / stuttering rate Critical event detection rate Bandwidth savings Example 1 open plains 45.2 First transmission strategy 0.02% 99.1% 88.5% Comparative Example 1 open plains 45.2 4K Full Backhaul 4.80% 92.4% 0% Example 2 Long tunnel 0.8 Third transmission strategy 0.00% (Image) 98.5% 99.8% Comparative Example 2 Long tunnel 0.8 1080P return 98.20% (Interrupted) 5.2% 0% Example 3 Deep trenches in mountainous areas 8.5 Second transmission strategy 1.15% 97.8% 72.3% Comparative Example 3 Deep trenches in mountainous areas 8.5 720P Full Download 22.40% 81.6% 0%

[0029] The quantitative data above shows that in open areas, this invention achieves nearly 90% bandwidth savings by adaptively reducing the redundancy of unnecessary images, thus leaving ample margin for communication with other trains in the same section. In tunnel scenarios with extremely poor network conditions, traditional solutions suffer from link collapse due to maintaining rigid transmission, while this invention ensures 100% accessibility of critical security data through a "dimensionality reduction" image backhaul strategy, greatly improving the system's survivability.

[0030] Furthermore, as a refinement of this embodiment, the edge detection module 2 employs an enhanced dataset tailored to the specific railway environment during the model training phase. Specifically, the dataset includes samples reflecting complex operating conditions such as low-light noise in tunnels, pantograph spark glare, and snow cover during winter. In the training loss function design, a focal loss is introduced to address the imbalance between outlier and background samples. Actual testing shows that in tunnel environments with illumination below 5 Lux, the model achieves a recall rate approximately 18 percentage points higher for identifying intruding objects compared to the general model.

[0031] Furthermore, the system described in this invention features deep customization at the network protocol stack layer. When performing real-time transmission, the sending module 5 employs an enhanced UDP-based transport protocol that incorporates Dynamic Forward Error Correction (FEC) coding. The policy scheduling module 4 monitors the channel's CQI (Channel Quality Indicator) feedback in real time. When channel quality deteriorates, the system automatically increases the proportion of redundant error correction packets, trading computational power for transmission determinism. However, when performing local high-definition video retransmission, the system switches to a TCP-based congestion control protocol, utilizing its stringent acknowledgment mechanism to ensure binary-level data integrity.

[0032] In the electronic device implementation scheme of this invention, the edge computing node 7 adopts an industrial-grade embedded architecture, and its internal bus adopts the PCIe 4.0 standard to support high-speed data exchange between the NPU and the video capture card. The high-speed non-volatile memory (NVMe SSD) 9 adopts SLC or enhanced MLC NAND flash memory and has power-loss protection logic to ensure that the recorded video metadata can be completely flashed into the flash memory in the event of an emergency power outage on the train.

[0033] In summary, this invention breaks away from the old model of "blindly collecting and indiscriminately transmitting" video in train surveillance by constructing an edge intelligent node with deep content awareness and network environment prediction capabilities at the vehicle-mounted end. It seeks a dynamic optimal balance among the three dimensions of "visible," "clearly visible," and "transmittable" based on real-time environmental conditions and security business requirements. Especially in railway application scenarios where 5G-R private network bandwidth resources are limited and high-speed movement leads to severe signal attenuation, this invention not only solves the reliability problem of video transmission but also builds an intelligent defense line with millisecond-level response capabilities for train operation safety through the priority transmission of structured information. The parameter configurations and strategy logic demonstrated in this embodiment are all optimal choices based on actual engineering test data, and have significant guiding significance for the intelligent upgrading of rail transit video surveillance systems.

[0034] Furthermore, the strategy scheduling module 4 incorporates the concept of a "historical network heatmap" into the decision-making process. The system can automatically learn the average throughput characteristics of the coverage areas of each base station along the line and store them as auxiliary input parameters in the GIS database. When a train enters a base station coverage sector with a high historical packet loss rate, the system will enter an early warning state earlier than relying solely on physical layer sampling, compressing the bit rate in advance to avoid image mosaic caused by instantaneous packet loss, thus further optimizing the monitoring experience for ground personnel. This self-learning transmission mechanism enables the system to automatically adapt to line aging and environmental changes during long-term operation, reducing manual maintenance costs.

[0035] As a further improvement to the system architecture of this invention, a dual watchdog monitoring mechanism is adopted between the edge computing node 7 and the on-board integrated sensing unit 6. Once an abnormal lock-up of the NPU computing unit or a crash in the video processing process is detected, the system will automatically reset the relevant modules within 500ms and forcibly switch to the most basic keyframe transmission mode during the reset. This integrated hardware and software redundancy design ensures that the system can maintain uninterrupted monitoring services in the long-term, high-load train operation environment.

[0036] In the edge computing-based adaptive transmission system for train video data provided by this invention, the various modules interact through tightly defined internal API interfaces. For example, the threshold evaluation module 3 transmits a standardized "network credibility index (0.0-1.0)" to the policy scheduling module 4, while the edge detection module 2 transmits an "event urgency matrix." This decoupled design enables the system to have good scalability. In the future, with the evolution of 5G-R or the deployment of 6G technology, only the underlying communication driver of the sending module 5 needs to be updated, without large-scale reconstruction of the upper-layer AI analysis logic and scheduling strategy.

[0037] To verify the maintainability of the present invention in actual deployment, the system also possesses adaptive traffic shaping capabilities when executing the retransmission mechanism. Specifically, the strategy scheduling module 4 dynamically adjusts the maximum upper limit traffic for video retransmission based on the bandwidth usage of other critical services on the train (such as signal system data and dispatch voice), ensuring that the retransmission task does not encroach on the transmission resources of critical train operation data. This global resource scheduling strategy demonstrates the deep customization advantages of the present invention on a high-security mobile platform like a train.

[0038] Finally, the system described in this invention supports remote configuration and firmware upgrades. The ground control center can dynamically adjust various threshold parameters of the lightweight target detection model (such as the sensitivity of fire alarm detection) based on seasonal changes or specific security level requirements via an encrypted downlink, and even remotely update the model version of edge nodes. This "edge-cloud collaborative" governance architecture enables each train node distributed across a wide area of ​​railway lines to always maintain the latest intelligent analysis capabilities, further enhancing the intelligence level of the entire railway operation and maintenance system.

[0039] In summary, by integrating advanced edge computing power, deep learning models, content-aware coding technology, and GIS-based bandwidth prediction algorithms, this invention provides a complete, efficient, and highly feasible closed-loop solution for train video transmission. It not only solves the bottleneck problems in high-speed mobile communication at the technical level but also provides a reliable architecture capable of adapting to extreme environments and ensuring the absolute delivery of critical information at the engineering practice level. This is not merely a simple superposition of communication technology and artificial intelligence in the railway industry, but a deep integration and innovation based on the essential needs of the business, possessing extremely high technological advancement and industry universality. The method and system proposed in this invention are fully compatible with the existing railway communication standard system and can provide solid infrastructure support for the integrated audiovisual operation and maintenance of next-generation intelligent high-speed railways.

[0040] In a specific implementation, to address the multipath effects and shadow fading that trains may encounter during high-speed travel, the strategy scheduling module 4 of this invention further integrates differential prediction logic for Channel Quality Assessment (CQI). Specifically, the system not only focuses on the absolute value of CQI but also on its slope over a very short period. When the slope is negative and the absolute value exceeds a preset slope threshold, even if the current RSRP still meets the requirements for high-definition transmission, the system will predict an "imminent severe fading" and intervene in the video encoding process in advance, shortening the I-frame interval (GOP length) by half to enhance the bitstream's error resilience and recovery speed. This microscopic insight into channel fluctuation trends, combined with the aforementioned macroscopic GIS geographic perception, constitutes the unique dual-scale prediction architecture of this invention, significantly reducing the probability of instantaneous interruptions in video backhaul.

[0041] Meanwhile, this invention also introduces a visual saliency weighting mechanism for the ROI extraction algorithm. Based on the detection of abnormal targets (such as foreign objects), the system determines auxiliary information within a 50-100 pixel range around the target by calculating image contrast, color gradient, and edge density. This auxiliary information often includes key visual features such as the boundary between the foreign object and the background, and the stress deformation at the contact point. By incorporating these regions into the ROI high-fidelity encoding scope, this invention can provide a richer visual background for ground-based expert diagnosis than a single target detection box, which has extremely high practical value in precision diagnostic scenarios such as pantograph damage analysis.

[0042] In terms of storage management in this embodiment, edge computing node 7 adopts a circular buffer management strategy. During normal system operation, all video streams are temporarily stored in RAM (Random Access Memory) in a rolling manner (storage time is approximately 2 minutes). Only when the policy scheduling engine 12 determines that a level 3 or higher (high priority) alarm event has occurred, the storage controller quickly dumps the temporary data in RAM to NVMe SSD persistent storage. This two-tier architecture of "RAM buffer + SSD persistence" avoids the problem of reduced lifespan caused by frequent SSD writes and ensures that critical historical footage (the preceding 30 seconds) before the anomaly occurs can be accurately captured. After the retransmission is completed, the system automatically releases the corresponding SSD block resources to maintain the continuous availability of storage space.

[0043] On the ground control center side, the corresponding decoding and adaptation unit is not only responsible for restoring video images but also for "metadata visualization." It overlays the structured detection boxes, confidence levels, timestamps, and other information returned by the edge detection module 2 onto the OSD (On-Screen Display) layer of the video image in real time and simultaneously marks them on an electronic map along with geographic coordinates. This multi-dimensional situational awareness interface significantly reduces the response time of on-duty dispatchers to emergencies. Furthermore, the ground-based decoder incorporates a packet loss compensation algorithm, which utilizes residual redundant information carried in the transmission protocol to smooth out minor transmission errors, ensuring the continuity of visual presentation.

[0044] The electronic equipment hardware involved in this invention employs stringent electromagnetic compatibility (EMC) hardening treatment during the PCB design phase to withstand the severe electromagnetic interference generated by the high-voltage pantograph of trains. The chassis of edge computing node 7 adopts an all-aluminum alloy finned heat dissipation design to ensure that the NPU can still operate at full frequency without overheating and frequency throttling in the high-temperature environment inside tunnels in summer (ambient temperature can reach above 50°C). This comprehensive engineering consideration, from algorithms, protocols, software architecture to hardware processes, jointly supports the long-term reliable operation of the technical solution of this invention in complex industrial environments.

[0045] In summary, this invention, through intelligent decision-making at the onboard edge, completely transforms the traditional "bandwidth-oriented" backhaul logic into a "value-oriented" intelligent transmission logic. In the context of railway transportation, a sector with extremely high safety requirements, this content-sensitive adaptive mechanism not only saves expensive spectrum resources but, more importantly, provides a technically mandated guarantee for the train's "right to visual survival" under extreme conditions. This approach of transforming edge computing nodes into part of the train's "brain" reflects the evolving trend of intelligent and autonomous development in modern rail transit equipment. All technical details of this embodiment can be fine-tuned according to specific line conditions and vehicle models, without departing from the scope of protection defined by the claims of this invention.

Claims

1. An edge computing-based adaptive transmission method for train video data, characterized in that, The method is applied to a collaborative transmission architecture consisting of an on-board unit and a ground control center, and the method includes the following steps: Step 1: Real-time acquisition of train operation status data and real-time network performance parameters of the currently accessed wireless communication network; wherein, the operation status data includes the train's current real-time speed, real-time geographical location information, and terrain attributes within a predetermined distance ahead of the train based on the geographical location information; the real-time network performance parameters include the current channel's uplink available bandwidth, packet loss rate, end-to-end network latency, and Doppler frequency shift caused by the train's high-speed movement; Step 2: Acquire the real-time raw video stream captured by the vehicle-mounted camera, and perform frame-by-frame analysis of the real-time raw video stream using a target detection model pre-deployed in the vehicle-mounted computing device, and output a structured video anomaly detection result containing anomaly event type labels, anomaly target bounding box position coordinates, and recognition confidence scores; Step 3: Based on the real-time operating status data and the real-time network performance parameters, construct a bandwidth prediction model; the bandwidth prediction model uses a weighted moving average algorithm combined with historical bandwidth performance, and introduces an environmental feedforward correction factor determined by the terrain attributes to dynamically predict the current available network bandwidth threshold in the next transmission cycle; Step 4: Based on the current available network bandwidth threshold and the semantic information of the video anomaly detection results, dynamically generate and execute the target video adaptive transmission strategy from multiple preset transmission strategies; Step 5: When the video anomaly detection result shows that there is an abnormal event and the real-time network performance parameter is lower than the second preset threshold, the extreme environment protection strategy is executed, that is, the key frame image of the abnormal event is captured and alarm text information is generated, which is encapsulated in the control plane channel protocol data unit of the wireless communication network and sent first, while the high-definition video clip corresponding to the abnormal event is stored in the local storage medium. Step Six: Continuously monitor the real-time network performance parameters. When the network performance recovers to the first preset threshold, and the first preset threshold is greater than the second preset threshold, and this state is maintained for a preset stable window duration, the breakpoint resume mechanism is automatically triggered to retransmit the high-definition video clips stored in the local storage medium to the ground control center using the remaining bandwidth.

2. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, In step one, the process of obtaining the Doppler frequency shift offset includes: according to the formula Calculate the predicted value, where The center frequency of the carrier. For the real-time speed of the train, The speed of electromagnetic wave propagation. The angle between the train's direction of travel and the electromagnetic wave propagation direction is used. The predicted value is fused with the actual received signal frequency deviation fed back by the wireless communication module using Kalman filtering to accurately assess the physical layer stability of the current channel, and the real-time network performance parameters are corrected based on the physical layer stability.

3. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, The vehicle-mounted computing device is an edge computing node, and the target detection model runs in the high-performance neural network processing unit (NPU) of the edge computing node. The target detection model is a lightweight model, which adopts a feature extraction network based on depthwise separable convolution and combines it with a feature pyramid structure (FPN) to achieve multi-scale feature fusion. Before running, it undergoes pruning and symmetric quantization based on 8 bits or 16 bits. The video anomaly detection results are output as structured data packets in JSON format, and the bounding box position coordinates are represented by normalized quadruples.

4. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, In step three, the process of constructing the bandwidth prediction model specifically includes: smoothing the real-time sampled bandwidth data using a weighted moving average algorithm, wherein the weight of the near-real-time sampled value is set to 0.7 and the weight of the historical average is set to 0.3; introducing an environmental feedforward correction mechanism: based on the geographical location information, when it is determined that the train is about to enter the tunnel area, deep trench section, or tall building distribution area marked by the terrain attribute within a preset distance, the bandwidth prediction model adjusts the current network available bandwidth threshold in advance according to the preset signal attenuation curve, so as to realize the early perception of the degradation of communication link quality.

5. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, In step four, the process of dynamically generating the video adaptive transmission strategy includes: when the video anomaly detection result shows that there is no anomaly in the current picture, and the current network available bandwidth threshold is greater than the first preset threshold, the first transmission strategy is executed; under the first transmission strategy, the on-board computing device reduces the frame rate of the original video stream to the first preset frame sampling rate through the time sampler, and performs downsampling processing on the video frames to achieve the first preset resolution, and then uses the H.265 standard and sets a high fixed quantization parameter QP to perform high compression ratio encoding in order to maintain the low bit rate inspection of the basic status of the train by the ground center.

6. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, In step four, the process of dynamically generating the video adaptive transmission strategy further includes: when the video anomaly detection result shows an abnormal event and the real-time network performance parameters are between a first preset threshold and a second preset threshold, a second transmission strategy is executed; under the second transmission strategy, a Region of Interest (ROI) extraction algorithm is started, and the vehicle-mounted computing device divides the video frame into a core ROI region and a non-core background region according to the bounding box position coordinates; a high proportion of bitrate budget is allocated to the core ROI region and lossless encoding or high-fidelity encoding with a low compression ratio is used, while spatial domain downsampling or fuzzy compression processing is performed on the non-core background region; the processed core ROI region data and background data are merged into a bitstream and transmitted through a dedicated high-priority network slice of the wireless communication network.

7. The adaptive transmission method for train video data based on edge computing according to claim 6, characterized in that, The ROI extraction algorithm also introduces a visual saliency weighting mechanism: when an abnormal target is detected, the visual association region within a preset range around the target is determined by calculating the image contrast, color gradient and edge density, and the visual association region is included in the encoding scope of the core ROI region to capture the physical failure edge features of the abnormal target.

8. The adaptive transmission method for train video data based on edge computing according to claim 1, characterized in that, In step five, the specific execution logic of the extreme environment protection strategy is as follows: the on-board computing device immediately stops the continuous video stream transmission task, retrieves the original I-frame images before and after the time of the anomaly from the video stream buffer, and compresses them to within a preset size limit; the alarm text information includes timestamp, train number, carriage location, anomaly type, and geographic coordinates; at the same time, the storage controller associates the full-resolution original video segments containing the preset duration before and after the anomaly with a unique event index identifier and writes them into the local storage medium.

9. An adaptive transmission system for train video data based on edge computing, characterized in that, The system includes: The status acquisition module is used to acquire train operation status data, geographic information system (GIS) map data, and physical layer information of wireless communication network. The status acquisition module is equipped with a data alignment unit to ensure the consistency of operation status data and network parameters in different dimensions in terms of timestamps. The edge detection module, integrated into the vehicle's computing device, runs a target detection model to perform real-time semantic parsing of the video stream and generate a structured anomaly detection report. The threshold evaluation module is used to run the bandwidth prediction algorithm and dynamically calculate the effective transmission bandwidth prediction value for the current and future predetermined time periods based on the multidimensional data provided by the status acquisition module and the preset environmental feature library. The strategy scheduling module, based on the bandwidth prediction value and the anomaly detection report, switches between executing the first transmission strategy, the second transmission strategy, or the third transmission strategy through a built-in decision logic matrix, and outputs the corresponding encoding control parameters. The execution transmission module includes a multi-protocol video encoding unit and a communication scheduling unit, which are used to resample the video stream, extract and encode ROI, extract keyframes or encapsulate text according to the instructions of the strategy scheduling module, and push the generated data packets to the ground control center through a wireless link.

10. The adaptive transmission system for train video data based on edge computing according to claim 9, characterized in that, The onboard computing device employs a dual-buffer queue architecture. The first buffer queue temporarily stores real-time data to be transmitted, while the second buffer queue, relying on the local storage medium, serves as persistent storage and is specifically used to store video segments to be retransmitted under extreme environment protection strategies. The policy scheduling module also introduces a dynamic channel quality evaluation (CQI) index as a feedback loop. When the packet loss rate reported by the execution sending module exceeds a preset threshold multiple times consecutively, the policy scheduling module automatically instructs the execution sending module to reduce the encoding bit rate and increase the redundancy of forward error correction (FEC) coding. The system also includes a decoding adaptation unit located at the ground control center. This unit automatically selects matching decoding parameters based on the received policy identifier and automatically stitches the subsequently retransmitted high-definition video segments with the real-time stream on the timeline to reconstruct a complete view of the event.

Images