Semantic hierarchical multi-path video backhaul method and system for vehicle insurance evidence collection and claim settlement

Abstract

The application provides a semantic hierarchical multi-path video backhaul method and system for vehicle insurance evidence collection and claim settlement. After collision, report or work order triggering, the terminal extracts frame sequence within the time window around the triggering time. The final semantic importance of each frame is generated by combining end-side rapid evaluation and improved convolutional neural network, the frames are classified, and the ROI-based hierarchical coding is implemented on the evidence focus area. The system perceives the link quality such as Bluetooth or satellite in parallel, determines the distribution share and backup degree, performs fountain code type forward error correction on the fragments, and delivers them according to the priority in multi-link concurrency. Each coding block carries content fingerprint, chain integrity label and terminal digital signature. The server side completes deduplication and verification, performs directional retransmission according to the priority for the missing data, and forms the link level arrival record. The scheme realizes the priority arrival and verifiable of key evidence in weak network and heterogeneous network environment, considers the cost and energy consumption control, and is suitable for forms such as driving recorders, mobile phones and vehicle gateway.

Description

Technical Field

[0001] This invention relates to the field of video processing technology, specifically to a semantically graded multi-path video backhaul method and system for vehicle insurance evidence collection and claims processing. Background Technology

[0002] With the online transformation of auto insurance services, accident scene footage is widely used for liability determination, rapid damage assessment, and dispute review. Current practices largely rely on dashcams or mobile apps to upload the entire video after an accident, or to stream it in real-time via a single path. However, complex road environments and varying coverage conditions mean that uplink links are frequently affected by congestion, obstruction, and handover, leading to delays or loss of critical segments. In the event of packet loss, traditional connection-based transmission is prone to header blocking, requiring retransmission of entire segments, resulting in long processing times and high bandwidth consumption. Furthermore, user terminals often possess cellular, Bluetooth, and even satellite communication capabilities, but existing systems have limited utilization of multiple communication modes, lacking sufficient multi-path parallelism and adaptive strategies, and failing to ensure the "first come, first served" availability of evidence in weak network conditions.

[0003] At the encoding and scheduling level, common solutions process the entire video using a fixed bitrate or a simple adaptive bitrate, lacking fine-grained priority encoding and bandwidth allocation for evidence areas strongly related to claims, such as license plates, collision contact points, and pedestrian upper bodies. While some solutions introduce keyframe or thumbnail pre-transmission, they fail to establish a stable time window around the trigger moment and do not link end-side sensors (accelerometers, airbags, brakes, etc.) with video scheduling, making it difficult to guarantee the identifiability and arrival order of key time periods and evidence areas. Especially in multi-link concurrent scenarios, the lack of joint assessment and share allocation of link bandwidth, latency, fluctuations, and loss ratios results in redundant delivery often remaining at a fixed ratio, making steady-state convergence difficult and failing to balance cost and energy consumption.

[0004] In edge perception and representation, existing lightweight CNNs mostly employ general activation functions and static feature extraction, lacking direct gating mechanisms to determine "which frame is more important or which region is more critical." Even when attention or loss weighting is introduced, it often remains at the training target layer or channel level, failing to effectively inject "first semantic importance" and sensor events into the inference stage. This results in insufficient feature response of evidence regions during weak network compression, affecting the selection of subsequent encoding and transmission strategies.

[0005] There are also gaps in evidence compliance and verifiability. Many systems only record upload time and account information, lacking block-level data integrity protection and source verification; mechanisms such as content hashing, context association, trusted timestamps, and client-side signatures are not in a closed loop, making it difficult to prove that images have not been tampered with and that the transmission order is authentic and reliable in disputed scenarios. In addition, breakpoint resumption is mostly done at the file level, failing to accurately recover fragments; multi-path deduplication and confirmation feedback are coarse-grained, leading to duplicate billing and invalid retransmissions. Regarding privacy governance, client-side minimization of non-involved faces and license plates is insufficient, and data minimization and compliant record keeping are incomplete.

[0006] The integration with the claims processing system is also problematic. Existing processes largely rely on users uploading large files at once or manually supplementing documents, lacking a mechanism for "precisely pulling back a specific time window / evidence area based on the work order." When network interruptions or frequent link switching occur, the system struggles to differentiate retransmissions based on arrival confirmation, easily leading to long-tail waiting times. In summary, existing technologies have shortcomings in semantic classification, focus encoding, adaptive parallelism across multiple communication modes, block-level verifiability, and differentiated retransmission, making it difficult to simultaneously meet the comprehensive requirements of "priority arrival of key evidence, controllable cost and energy consumption, and credible evidence chain" under complex network conditions. Therefore, it is necessary to propose a full-link solution that completes semantic evaluation and gating on the endpoint, performs link awareness and intelligent distribution on the transmission side, and implements deduplication, reassembly, and evidence collection closed loop on the server side. Summary of the Invention

[0007] To address the aforementioned problems in existing technologies, this invention proposes a semantically graded multi-path video backhaul scheme for vehicle insurance evidence collection and claims processing. Upon collision, report, or work order triggering, the terminal extracts the frame sequence within a time window around the trigger time. Combining on-device rapid evaluation with an improved convolutional neural network to generate the final semantic importance of each frame, the frames are graded, and ROI-based hierarchical encoding is implemented for the evidence focus area. The system simultaneously senses the quality of links such as cellular, Wi-Fi direct, V2X, Bluetooth, or satellite, determines the distribution share and backup rate, performs fountain code-based forward error correction on the segments, and delivers them concurrently across multiple links according to priority. Each encoded block carries a content fingerprint, a chain-like integrity tag, a trusted timestamp, and a terminal digital signature. The server performs deduplication and verification, reassembles the data when coverage reaches a threshold, prioritizes and retransmits missing data, and forms a link-level arrival record. Local caching occurs when a link is interrupted, and differentiated breakpoint resumption is performed based on the confirmation result after recovery. This solution enables priority access and verification of key evidence in weak and heterogeneous network environments, while taking into account cost and energy consumption control. It is applicable to various forms such as dashcams, mobile devices, and vehicle gateways.

[0008] This application provides a semantically graded multi-path video backhaul method for vehicle insurance evidence collection and claims processing, including the following steps: S1: Triggered Acquisition: When a collision is triggered, a user reports a case, or a claim work order is issued, the terminal extracts continuous image frames within a preset time window centered on the trigger time. S2: Semantic classification: The first semantic importance of each frame is calculated on the edge side and input together with the frame data into the improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance. S3: Focus Encoding: Perform hierarchical encoding of the evidence focus region of each frame with ROI constraints to obtain source fragments with focus mask markings; S4: Path awareness: Parallel detection of link status in at least two different communication modes; S5: Allocation and Loss Mitigation: Based on the available throughput, round-trip time, arrival interval fluctuation and loss ratio of each link, determine the distribution share and backup degree for each link; perform out-of-order recoverable forward error correction coding on the source fragments to generate coding blocks, and deliver the coding blocks to multiple links for parallel transmission according to frame priority. S6: Verifiable backhaul: Generate a content fingerprint for each coded block and write the preceding fingerprint in the order of transmission to form a chain integrity tag. Attach a trusted timestamp and generate a digital signature with the terminal key. The server side performs deduplication, verification, and reassembly based on the content identifier, and prioritizes and sends missing data back to the source for supplementary transmission. S7: Resume interrupted download: When a path is interrupted, unacknowledged data blocks are cached locally, and resumed based on the arrival of the acknowledgment result after any path becomes available.

[0009] Preferably, the first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the triggering time, and the license plate image recognition accuracy.

[0010] Preferably, the end-side convolutional neural network includes a lightweight backbone, sensor feature branches, and a multi-task output head; an amplitude gating unit based on the first semantic importance is set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branches, thereby outputting the final semantic importance, evidence mask, key points, and sharpness evaluation results.

[0011] Preferably, the path-aware link status parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio; and optionally, signal quality indicators, path availability indicators, unit data cost, and unit energy consumption labels are collected; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup rate; the communication mode includes one or more of cellular communication, Wi-Fi direct connection, vehicle-to-everything (V2X), Bluetooth relay, or satellite fallback.

[0012] Preferably, the distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

[0013] This invention also provides a semantically graded multi-path video backhaul system for vehicle insurance evidence collection and claims processing, comprising: Trigger acquisition module: When a collision is triggered, a user reports a case, or a claim work order is issued, the terminal extracts continuous image frames within a preset time window centered on the trigger time; Semantic classification module: The first semantic importance of each frame is calculated on the edge side and input together with the frame data into an improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance. Focus encoding module: Performs hierarchical encoding of ROI constraints on the evidence focus area of ​​each frame to obtain source fragments with focus mask markings; Path awareness module: detects the link status of at least two different communication modes in parallel; Allocation and Loss Prevention Module: Based on the available throughput, round-trip delay, arrival interval fluctuation and loss ratio of each path, determine the distribution share and backup degree for each path; perform out-of-order recoverable forward error correction coding on the source fragments to generate coding blocks, and deliver the coding blocks to multiple paths for parallel transmission according to frame priority; Verifiable backhaul module: Generates content fingerprint for each coded block and writes it into the preceding fingerprint in the order of transmission to form a chain integrity tag, attaches a trusted timestamp and generates a digital signature with the terminal key; The server side performs deduplication, verification and reassembly based on the content identifier, and backhauls missing data to the source according to priority. Resume interrupted download module: When the path is interrupted, unacknowledged data blocks are cached locally, and resumed download based on the arrival of the acknowledgment result after any path is restored to availability.

[0014] Preferably, the first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the triggering time, and the license plate image recognition accuracy.

[0015] Preferably, the end-side convolutional neural network includes a lightweight backbone, sensor feature branches, and a multi-task output head; an amplitude gating unit based on the first semantic importance is set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branches, thereby outputting the final semantic importance, evidence mask, key points, and sharpness evaluation results.

[0016] Preferably, the path-aware link status parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio; and optionally, signal quality indicators, path availability flags, unit data cost, and unit energy consumption tags are collected; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup rate.

[0017] Preferably, the distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

[0018] This invention provides a semantically hierarchical multi-path video backhaul method and system for vehicle insurance evidence collection and claims processing, which achieves the following beneficial technical effects: 1. This solution first obtains the initial semantic importance through lightweight evaluation on the edge side, and then implements gated activation and sensor-driven channel modulation in mid-to-high-order feature layers using an improved convolutional neural network, significantly improving the expressive power of key evidence such as license plates, collision contact areas, and the upper body of pedestrians. During the encoding stage, a layered strategy with ROI constraints is applied to the key evidence: the base layer ensures "clear visibility and priority," while the enhancement layer is responsible for improving overall clarity; in weak network or congestion situations, only the base layer and the key areas are preserved to fully reconstruct the crucial facts. Combined with "focus mask marking," the scheduler sets priority queues and dual communication mode redundancy to avoid evidence gaps caused by single-link interruptions. Compared to overall balanced compression and indiscriminate push, this solution achieves "evidence first, clarity later," improving recognizability at the same bitrate and maintaining credible images under limited bitrates, significantly reducing secondary transmission and communication costs.

[0019] 2. The system of this invention can simultaneously sense the available throughput, round-trip latency, arrival interval fluctuations, and loss ratios of links such as cellular, Wi-Fi direct, V2X, Bluetooth, or satellite, and can combine these with signal quality, cost, and energy consumption tags to form a comprehensive profile. The scheduling end describes the load and redundancy intensity of each link using "distribution share and backup degree," and adopts a windowed strict priority approach: first the basic layer, then the enhancement layer; within the same priority, weighted rotation is performed according to share, and link degradation is immediately reassigned, with differentiated resumption after recovery. Source fragmentation undergoes fountain code-type forward error correction, and reassembly can be performed once the reception coverage meets the standard, avoiding full-segment retransmission; the end side maintains breakpoint resumption and deduplication confirmation to reduce duplicate billing. The policy engine dynamically reduces the enhancement layer ratio of high-cost links based on monthly traffic, power consumption, and hotspot availability, prioritizing the use of low-cost, stable-quality channels to achieve the comprehensive optimization of arrival rate, latency, cost, and energy consumption.

[0020] 3. Each coded block of this invention carries a content fingerprint and references the preceding fingerprint to form a chain-like integrity tag. It also attaches a trusted timestamp and is digitally signed by the terminal key. The server performs deduplication, signing and time verification, chain continuity checks, and quality statistics on arriving data. Reassembly is triggered after the coverage meets a threshold, and missing or unqualified areas are prioritized for retransmission back to the source. The system automatically generates auditable records at the link and window levels, including arrival time, deduplication hits, reassembly progress, and reassignment trajectory, which can be exported as reports for claims and dispute review. Regarding privacy governance, the terminal side minimizes the processing of non-involved faces and license plates and uses local encrypted caching. Transmission and storage are protected by keys throughout. The solution also provides a precise reverse-pull interface with the work order system, realizing a closed-loop retransmission "by event ID and time window," making the evidence chain traceable, verifiable, and operable. Attached Figure Description

[0021] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0022] Figure 1 This is a flowchart of the steps of a semantic hierarchical multi-path video backhaul method for vehicle insurance evidence collection and claims processing according to the present invention. Figure 2 This is a schematic diagram of a semantic hierarchical multipath video backhaul system for vehicle insurance evidence collection and claims processing according to the present invention. Detailed Implementation

[0023] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0024] Example 1: In view of the aforementioned problems mentioned in the prior art, and in order to solve the above technical problems, as shown in the appendix. Figure 1 As shown: This application provides a semantically graded multi-path video backhaul method for vehicle insurance evidence collection and claims processing, including the following steps: S1: Trigger Acquisition: Upon collision triggering, user reporting, or issuance of a claim work order, the terminal extracts continuous image frames within a preset time window centered on the trigger moment. In one embodiment, the dashcam (pre-installed / aftermarket) features a front / rear camera at 1080p@30fps, H.265 encoding; an IMU (three-axis accelerometer, gyroscope); GNSS timing; CAN / OBD reading of airbag / brake signals; and a built-in "accident button." A normally open "ring-shaped video buffer" covers 90–120 seconds (first 60s + last 60s), GOP1s, with keyframes aligned to the whole second. Trigger source and priority collision: Acceleration reaches a preset impact threshold and the duration exceeds the de-shake threshold; or the airbag trigger position changes. Manual reporting: The user presses the built-in accident button. Work order issuance: A claim work order arrives at the device from the cloud, including the approximate accident time. Trigger priority: Airbag triggering > IMU impact > Manual reporting > Work order issuance. When airbag triggering is present, t0 is determined by the airbag triggering edge time. Otherwise, t0 is determined by the IMU peak occurrence time within the current window; if there are multiple peaks, the largest peak is used. When only a manual report is made, the button press time is used as t0. When a work order is triggered, the work order given time is used as the initial t0; if the difference from the most recent significant IMU peak is within ±60 seconds, then the peak time is used to correct t0. Δt value and clipping: The default Δt=60 seconds, truncated by 30 seconds before and after t0; if a severe collision is detected (e.g., strong IMU impact or airbag triggering), it automatically expands to Δt=120 seconds (60 seconds before and after); for minor scrapes, it can converge to Δt=40 seconds (20 seconds before + 20 seconds after). Buffer freeze and subsequent acquisition: The segment before t0 of the circular buffer is "frozen" at the moment of triggering; recording continues until the "subsequent segment" is fully acquired. If power is cut off before the subsequent data collection is complete, a supercapacitor will provide 20-30 seconds of power to complete the critical disk placement; the uncompleted portion will be marked as "subsequent data collection duration" in the event metadata. Time synchronization and annotation frames are timestamped using a monotonic clock; the offset from GNSS / NTP is also recorded and written into the event metadata. Multiple cameras are encapsulated into "multi-channel synchronization segments" using a unified time base, and camera intrinsic / extrinsic parameter snapshots are recorded for subsequent spatiotemporal alignment. Anti-false triggering and road bump filtering: short acceleration pulses without significant vehicle speed changes are downweighted or do not trigger; trigger cooldown time: after a trigger is completed, it will not be repeatedly triggered by similar small pulses within 10 seconds; if a false trigger is manually canceled by the user within 30 seconds, the event is marked as "invalid" and the space is reclaimed.

[0025] In one embodiment, the mobile app + OBD Bluetooth (low modification cost) operation mode features the app constantly running in the foreground / background; using the main camera at 1080p@30fps, reducing to 720p if necessary; using BLE connection to OBD to read vehicle speed, braking, and airbag status; and utilizing the phone's built-in IMU and GNSS. A circular buffer is used for the foreground for 90 seconds; the background power-saving mode buffers for 45 seconds (reducing frame rate to 24fps and bitrate). The trigger source is an OBD report of airbag triggering or safety-related fault code activation; the phone's IMU detects a significant impact accompanied by a sudden drop in speed; the "Report an Incident" button is activated in the app; and a work order is issued from the cloud. t0 and Δtt0 are determined by prioritizing "OBD airbag > IMU peak value > report button > work order time"; the default Δt = 60 seconds; in cases of personal injury or high-speed driving, the work order issuance time is extended to 90 seconds; when the battery level is <15%, the time automatically converges to 40 seconds, prompting the user to quickly move to the side and retake static scene photos. To address insufficient foreground information, if the app has just been activated, resulting in insufficient foreground information, an event will still be generated, and the "foreground availability duration" will be noted in the metadata. The UI will guide the user to capture a panoramic view of the scene and collision details as supplementary evidence. Storage and security: Event fragments are stored locally with encrypted encryption (based on the system security zone key); automatic uploading is allowed after network recovery. Automatic extension of the background is allowed when Wi-Fi is available (e.g., continuing collection for another 20 seconds) to improve the completeness of the evidence.

[0026] In one embodiment, the vehicle-mounted system / ADAS gateway (multi-camera multi-sensor fusion) hardware and interface vehicle-mounted system SoC connects to front / rear / left / right GMSL2 cameras for multi-frame synchronization; connects to CAN / FlexRay, GNSS, and 5G modules; and can read ADAS events such as AEB and FCW. Triggering sources for ADAS intervention include: AEB execution, high-confidence FCW alarm, BSD side and rear warning triggering; vehicle side collision / airbag signal; manual reporting (central control screen button); and work order issuance. The t0 value strategy uses the braking execution command time as t0 when AEB is triggered; when there is no AEB, it uses the highest confidence ADAS collision warning time or IMU peak time as t0; in manual reporting / work order scenarios, t0 is retrospectively adjusted or corrected according to the aforementioned principles. Δt and multi-channel synchronous export: The default Δt = 120 seconds (60 seconds each for front and rear), with full-channel synchronous trimming; simultaneously, CAN segments (vehicle speed, steering angle, braking status, turn signals) aligned with video time and key ADAS event logs are exported. If the vehicle continues to move and the subsequent data collection is not yet complete, an extension of 20-30 seconds is allowed to complete the subsequent data collection, provided it does not affect driving. Consistency and verification use GNSSPPS or PTP to ensure the time consistency of multi-camera and bus data; a unique event ID (event_uuid) is generated as the association key for subsequent encoding, transmission and server reassembly; frame loss statistics and exposure anomalies of each camera are written into the event metadata to facilitate backend evaluation of evidence reliability.

[0027] S2: Semantic classification: The first semantic importance of each frame is calculated on the edge side and input together with the frame data into the improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance. In one implementation example, the calculation of the first semantic importance (fast edge scoring as a priori) is input to the synchronous video: frame-by-frame input (e.g., 1080p / 30fps), while a low-resolution copy is retained for fast detection. Sensors: IMU (accelerometer, angular velocity), airbag trigger position, vehicle speed / braking, etc., are aligned with the video frames via a monotonic clock. Three types of core signals and generation methods: Human / vehicle keypoint confidence and stability are assessed by performing lightweight target detection and keypoint estimation on low-resolution frames to obtain confidence values ​​for keypoints such as the upper body of pedestrians, front and rear corners of vehicles, lights, and license plate candidates. Combining optical flow or keypoint tracking from adjacent frames, the continuity and jitter amplitude of keypoint trajectories are evaluated; targets near the collision contact area are given higher weights. Adding and subtracting scores are performed by combining sharpness and occlusion judgment (e.g., Laplacian variance, visibility ratio) to obtain a "usable and stable" keypoint score. Inertial event intensity related to the trigger moment: Within a small window before and after the trigger, the peak value of acceleration / angular velocity and its duration are identified, and the relative distance to the trigger moment is recorded. If a sudden drop in vehicle speed, abnormal braking, or airbag activation is observed simultaneously, the relevance level is increased; a level is decreased for regular bumps or deceleration. Frames closer to the activation time and with higher intensity are given higher scores. License plate readability and stability are assessed using a lightweight license plate detector to locate candidate regions, followed by affine correction, sharpening, and contrast stretching before edge-side OCR. Readability levels are assigned based on character confidence, consistency verification, and glare / blurring / occlusion assessment; a small bonus is awarded when adjacent frames have consistent and stable text. Scoring and robustness processing combine the above three items into a "first semantic importance" with configurable weights. By default, key point items have the highest weight, followed by inertial relevance, and then license plate readability; if airbag activation or high-speed impact is detected, the influence of the inertial item is temporarily increased. When an item is missing, it is not recorded as zero; a minimum score is used, and the weight of that item is distributed among other items to avoid drastic changes in priority. A slight temporal smoothing is applied to the scores of individual frames to reduce jitter; finally, the prior importance of each frame is obtained, which is used for gating and reference in subsequent deep models.

[0028] An improved edge-side convolutional neural network (structure and function) incorporates "amplitude-gated activation driven by first semantic importance" into a lightweight backbone and introduces a "sensor feature branch" to generate channel scaling and bias, dynamically modulating mid-to-high-order features. The multi-task output head simultaneously provides evidence masks, keypoints / contact areas, sharpness assessments, and "final semantic importance." The lightweight backbone structure employs a mobile-friendly network (such as a small variant of MobileNetV3 or an equivalent structure), divided into four stages, extracting texture, shape, and semantic features from the bottom up. This facilitates INT8 quantization and ARM / NPU deployment, ensuring real-time performance on the edge. The sensor feature branch concatenates inertial event intensity, airbag trigger position, vehicle speed / braking, etc., aligned with the current frame into a short vector, which is then processed through two or three small fully connected layers and normalized to obtain the "sensor embedding." This embedding is subsequently used to generate channel scaling and bias, modulating the backbone features in an event-aware manner. Semantic Importance Gated Activation (an improvement in this scheme, while maintaining the same key technologies): Using the "first semantic importance" as the gain signal, amplitude gating units are placed in the key layers of the backbone: in the middle layer (where object structure is significant), local structures are amplified / suppressed; in the high layer (where semantic aggregation is significant), channels related to the evidence focus are enhanced, while background and noise channels are suppressed; a global-level "total control gain" is performed before entering the output head. The placement order follows "normalization / attention first, then gating, then merging with residuals," which neither disrupts statistics nor reduces additional latency. Multi-task Output Head: Evidence Region Segmentation Head: Outputs masks for license plates, pedestrian upper bodies, collision contact areas, brake lights, etc., with resolution obtained by upsampling at a medium scale in the backbone. Key Point / Contact Area Head: Supplements vehicle / human key points and contact candidate boxes to assist in ROI accuracy. Clarity / Readability Evaluation Head: Provides perceptual quality levels for license plates and contact areas, allowing the encoder to set a lower limit. Semantic Importance Scoring Head: Integrates global pooling features, sensor embeddings, and the "first semantic importance" to output the "final semantic importance." The end-to-end latency is controlled to the tens of milliseconds; the model size is several megabytes, and after quantization, it can run in real time on common automotive / mobile SoCs. When temperature or power is limited, it can degrade to enabling gating activation only at higher levels to maintain keyframe discrimination performance.

[0029] Priority allocation and output integration (from "prior" to "final") simultaneously obtains the "first semantic importance" (prior) and the "final semantic importance" (deep model output) for each frame. Based on the "final semantic importance," and combined with business thresholds, at least two levels are defined: high priority and low priority (expandable to three levels). The evidence mask is slightly expanded and output along with the frame metadata, then handed over to the encoder for hierarchical encoding with ROI constraints: the evidence region of high-priority frames has a stricter quality lower limit, while non-evidence regions can be more strongly compressed in weak network conditions. Segments that "must be sent first" (such as the evidence region of the base layer) are marked to subsequent scheduling modules to support multi-link dual-mode redundancy and weak network first-reach strategy.

[0030] In one embodiment, the first semantic importance of each frame in the calculated frame sequence, and the first semantic importance of the k-th frame... for:

[0031] in, , , These represent the weight parameters, Indicates the first Average confidence level of the vehicle body in the frame; Indicates the first The frame represents the aligned peak correlation of the IMU; Indicates the first Confidence of license plate OCR.

[0032] The activation function used in the improved convolutional neural network model at the endpoints :

[0033] in, The previous network layer outputs a feature vector. Represents the hyperbolic tangent function. This is the preset scaling factor.

[0034] The link weight is calculated based on the bandwidth, round-trip time, and packet loss rate of each link. and redundancy ratio ,include:

[0035]

[0036] in, For the first Available bandwidth of each link For the first Round-trip delay of the link; For the first Packet loss rate of each link; This is the upper limit of the redundancy ratio. , These are non-negative weighting coefficients.

[0037] In some embodiments, the edge-side improved convolutional neural network (CNN) structure, the functions of each branch, and the specific placement and triggering logic of the "semantic importance gating activation function" are described.

[0038] 1) Operating Environment and Input. Edge Platform: ARMA76×2 or equivalent automotive SoC; supports INT8 quantization and NEON instructions; optional NPU. Input: 1080p per frame; generates two resolutions on the edge: 640×360 for detection / segmentation; 224×224 for global scoring. Auxiliary Signals: IMU (accelerometer / angular velocity), airbag trigger position, vehicle speed / braking. Aligned with video using a monotonic clock.

[0039] 2) Overall Model Structure (Lightweight Multi-Head). Backbone: A modified MobileNetV3-Small network, divided into four stages (Stage 1~Stage 4), with an overall downsampling of 1 / 32. Each stage consists of several InvertedResidualBlocks, including pointwise convolutions, depthwise separable convolutions, pointwise linear convolutions, and batch normalization / attention modules. Sensor Branch: Acceleration peaks, airbag triggering, vehicle speed, braking, etc., aligned with the current frame are concatenated into short vectors, which are then processed through two fully connected layers and activation layers to obtain "sensor embeddings" for subsequent channel scaling and biasing. Multi-Head Output Evidence Region Segmentation Head: A lightweight decoder (upsampling + depthwise separable convolutions + skip connections) outputs low-resolution masks, with categories including license plates, pedestrian upper bodies, collision contact areas, brake lights, etc. Keypoint / Contact Area Detection Head: Outputs keypoints and candidate contact boxes. Clarity / Readability Regression Head: Outputs readability scores for license plates and contact areas, used as the encoder's quality lower limit. Semantic importance scoring head: It integrates global pooling features, sensor embeddings, and "first semantic importance" to output the final semantic importance score.

[0040] S3: Focus Encoding: Perform hierarchical encoding of the ROI constraint on the evidence focus area of ​​each frame to obtain source segments with focus mask markings; in some embodiments, 1) Focus Mask Preparation (following the output of S2) Input: Evidence focus mask for each frame (categories include license plate, pedestrian upper body, collision contact area, brake light, etc.), final semantic importance and priority. Cleaning and Merging: Perform a light smoothing and 3-8 pixel morphological expansion on the mask to avoid jagged edges; when multiple categories overlap, merge according to priority, with evidence priority from high to low as follows: license plate ≈ contact area > pedestrian upper body > others. Low-Resolution Adaptation: Align the mask according to the macroblock / CTU grid used by the encoder to ensure that the ROI boundary falls on the block boundary controllable by the encoder. Temporal Stabilization: Use the same object ID across frames for the same target, limit the maximum percentage change of mask area in each frame, and reduce bitrate jitter caused by "flickering".

[0041] 2) Layered Encoding and ROI Constraints (End-Side Encoder Configuration): Encoding Layers: Base Layer (BL): Prioritizes "evidence first, clarity later." Recommended resolution: 480p–720p, 12–15fps, ensuring minimum clarity and readability in key evidence areas. Enhancement Layer (EL): Aims for overall image quality improvement. Recommended resolution: 1080p, 24–30fps, can be transmitted when network conditions are good or low-cost links are available. Code control and GOPGOP are approximately 1 second, triggering a forced keyframe near the event; periodic intra-frame refresh is enabled in weak network conditions to ensure local recovery. Bitrate mode adopts "constrained variable bitrate": BL has a lower limit protection, ROI has an inherent quality baseline; EL expands elastically with network margin. ROI constraint implementation: For chips with hardware supporting ROI / QP maps: set a lower quantization step size for blocks falling within the mask (higher quality), and moderately increase quantization in non-ROI areas (more code-efficient). For chips that do not support ROI: adopt a dual-path strategy (see "Compatibility and Degradation"), using patch streams to cover key areas. Multi-camera consistency: Multiple cameras use a common time base and a unified GOP rhythm; the ROI rules of each channel are consistent to avoid timing misalignment during reassembly.

[0042] 3) Generate source fragments with focus mask markings. Fragment granularity is based on a "scheduling window" (e.g., 1 second), outputting fragments one by one for each frame within that window. Each frame contains at least three types of fragments: Metadata fragments: containing window ID, frame timestamp, frame priority, mask category and polygon / raster digest, coding level (BL / EL), GOP location, privacy markers, etc. BL video fragments: carrying the bitstream segment of the base layer frame; segments related to the evidence focus are marked "must be sent first". EL video fragments: carrying the bitstream segment of the enhancement layer frame; appended transmission on good networks or low-cost links. Fragment size: 1–4KB recommended to avoid MTU fragmentation; each fragment has a monotonically increasing sequence number and intra-frame offset for easy out-of-order reception and reassembly. Mask markings are recorded in the metadata fragments: a list of mask categories, boundary points or raster digests, the corresponding coding block range, and whether privacy processing is involved. For BL / EL video segments, add lightweight tags: whether it covers an ROI, which type of ROI it belongs to, and whether it is "must be sent first". The interface S4 with the subsequent stage can directly read the "must be sent first" and "belongs to ROI" tags and preferentially map them to links with at least two different communication modes; S5 uses the content identifier and time information of the segment to establish a hash chain and forensic signature.

[0043] S4: Path Awareness: Parallel detection of link states in at least two different communication modes; calculation of the path state parameters (bandwidth, round-trip time, and packet loss rate) for each link. Link weight and redundancy ratio The fragments are then encoded using fountain codes with forward error correction to obtain multiple coded blocks, which are then mapped to multiple links for transmission according to priority; Number of transmissions per link for: .

[0044] S5: Allocation and Loss Mitigation: Based on the available throughput, round-trip time, arrival interval fluctuation, and loss ratio of each link, determine the allocation share and redundancy for each link; perform out-of-order recoverable forward error correction coding on source fragments to generate coded blocks, and deliver the coded blocks to multiple links for parallel transmission according to frame priority; 1) Input and Time Granularity, Scheduling Window: A window of 1 second (adjustable according to the model) processes all source fragments generated in that second. Hierarchy and Priority: Fragments within each window are hierarchically divided into a base layer and an enhancement layer, and classified into at least two levels of high / medium / low importance according to frame level. Link Status: Probe results from the previous stage, including the available throughput, round-trip time, arrival interval fluctuation, loss ratio, and optional signal quality, cost, and energy consumption labels for each link.

[0045] 2) Determine the "distribution share" and "backup ratio" for each link. Distribution share (who sends more): First, perform threshold filtering: only retain links that meet the minimum requirements of the base layer (throughput not lower than the lower limit, latency not higher than the upper limit, and loss ratio not too high). Rank links that pass the threshold according to overall quality; the better the quality, the larger the share. If there are cost or power constraints, prioritize allocating enhancement layer shares to low-cost or low-power links. Form the "sendable quota" for each link in this window, which can be understood as the number of tokens sent (not a fixed number of packets, it can change with the window). Backup ratio (how many redundancies to send): The backup ratio of the base layer is higher than that of the enhancement layer; base layer fragments related to evidence focus require "at least one copy of each of two different communication modes." Links with more fluctuating network quality and higher loss ratios have more redundancy; high-quality links have less redundancy to control overhead. When cost or power is tight, compress enhancement layer redundancy first, retaining redundancy for the base layer and evidence focus.

[0046] 3) Unordered recoverable forward error correction coding (fountain code type) and a "pending block" list, window-level coding: Basic layer source fragments within the window are encoded together to generate sufficient coded blocks; enhancement layers are encoded separately to generate another set of coded blocks. Decoding trigger: The receiver can reassemble the blocks once it has accumulated enough different coded blocks; they do not need to arrive in strict order. List-based management: Two "pending lists" are generated: a basic layer list (with evidence focus fragments marked as "must be prioritized" and "dual-mode transmission") and an enhancement layer list.

[0047] 4) Prioritize and deliver coded blocks to multiple links (mapping and queues), with strict priority: always schedule the base layer first; within the base layer, prioritize evidence focus blocks with "must-priority" markings. Dual-mode guarantee: For "must-priority" blocks, select two different communication modes and assign one copy to each; if a mode is temporarily unavailable, another mode will take over until the unavailable mode is restored. Weighted round-robin dispatch: Within the same priority, perform weighted round-robin dispatch based on the "sendable quota" of each link: when it is a link's turn, take several blocks from the head of the list and add them to the link's send queue until its quota is exhausted. Queue implementation: each link has an independent send queue (base layer queue + enhancement layer queue), and each link's "link send thread" sends according to congestion control and rhythm.

[0048] 5) Parallel Transmission and Loss Prevention Strategies (End-Side): Rate and Pace: Initially, a "slow start" approach is adopted to gradually increase the transmission rate, avoiding sudden congestion; the pace is more conservative for high-fluctuation links. Out-of-Order Tolerance: Fountain codes allow out-of-order arrival, and the end-side does not need to queue for sequence; only the time and hierarchical metadata of each block needs to be ensured. Instant Reassignment: If a link is detected to have increased jitter, a sudden increase in loss, or a direct interruption within this window, the unsent quota is immediately reassigned to other links; reassignment prioritizes links with "different communication modes" from the original link to improve the arrival probability. Deduplication Pre-processing: The end-side calculates a content identifier for each coded block to prevent the same block from being repeatedly reassigned on different links, exceeding the expected redundancy. Cost / Energy Threshold: When the cellular package approaches the threshold or the battery is insufficient, the allocation quota for the enhancement layer is automatically downgraded, while the quota for the basic layer is maintained; when Wi-Fi / V2X is present, priority is given to these low-cost links.

[0049] 6) Server-side confirmation, reassembly, and client-side adaptive arrival confirmation: After deduplication based on content identifiers, the server returns a "received list" in bitmap or digest format. The client removes confirmed blocks from the pending list accordingly. Reassembly determination: When the arrival coverage of the base layer reaches a preset threshold, reassembly and rendering of that window are triggered immediately; the enhancement layer is reassembled sequentially when resources allow. Targeted replenishment: If critical blocks are still missing during reassembly, the server initiates a back-to-origin request to the client in the order of "base layer evidence focus, then base layer ordinary, and finally enhancement layer," prioritizing communication modes different from those of failed links. Cross-window convergence: The server periodically returns the effective arrival count, average arrival latency, duplication rate, and failure reasons for each link; the client automatically fine-tunes the "sendable quota" and "backup degree" of each link in the next window, so that the allocation converges to a stable state within several windows.

[0050] 7) Typical numerical example: Within the same window, there are 100 base layer source fragments and several enhancement layer fragments. The detection results show that Wi-Fi has the best quality, followed by 5G, while V2X fluctuates the most. In this round of allocation: Wi-Fi receives the largest transmission quota, followed by 5G, with V2X primarily handling supplementation and redundancy; base layer blocks relevant to the evidence focus are first transmitted once each on Wi-Fi and 5G to ensure dual-mode transmission; the remaining base layer blocks are filled using a weighted rotation of "Wi-Fi→5G→V2X" to fill the quota; if Wi-Fi fluctuates midway, the remaining quota is immediately reassigned to 5G and V2X. Once the reassembly threshold is met, the server first outputs usable base layer video; for missing enhancement layers or local details, they are retransmitted back to the source according to priority.

[0051] 8) Boundary and Fault Tolerance for Short-Term Strong Fluctuations: To avoid frequent reassignments caused by back-and-forth jitter, a minimum keep-alive duration and cooldown time are set; link quality recovery requires meeting standards for multiple consecutive sampling cycles. Long-Term Single-Link Availability: If only the cell remains available, the system automatically downgrades the bitrate and only dispatches base layer and evidence focus; after restoring other links, the enhancement layer backlog is cleared. Power Outage / Reboot: The pending list and confirmed bitmaps are saved to disk, and the system resumes from the last progress after a reboot; seamless integration with the S7 breakpoint resume strategy.

[0052] S6: Verifiable backhaul: Generate a content fingerprint for each coded block and write the preceding fingerprint in the order of transmission to form a chain integrity tag. Attach a trusted timestamp and generate a digital signature with the terminal key. The server side performs deduplication, verification, and reassembly based on the content identifier, and prioritizes and sends missing data back to the source for supplementary transmission. S7: Resume Transmission: When a path is interrupted, unacknowledged data blocks are cached locally. Once any path becomes available again, differentiated resumption is performed based on the arrival confirmation result. In some embodiments, S6 is verifiable backhaul (end-side → server): 1) End-side: The source of the "verifiable" encapsulated block of the encoded block: from the fountain code / forward error correction output of the previous stage. Each block has an event ID, window ID (e.g., one window per second), level (base layer / enhanced layer), frame number, whether it is ROI, priority, and other context. Content fingerprint: Calculate the content fingerprint (fixed-length digest) for the block payload. This fingerprint serves as the primary key for subsequent deduplication and verification. Chained integrity tag: Write "content fingerprint of the previous block" in the header of the current block; write "chain start point" in the starting block of each window. This forms a chain of trusted time stamps in the order of transmission: first request time tickets from the Time Stamp Authority (TSA) to be written in the block header; if the site is offline, write "local security clock + GNSS / NTP offset" and set "to be re-signed", and fill in the TSA tickets block by block after the network is restored. Digital Signature: Within the Terminal Security Unit (TEE / SE / TPM), the device's private key is used to sign the "content fingerprint + previous fingerprint + time information + key metadata (event / window / level / frame / ROI / priority)"; the block header also carries the device certificate chain or device ID for verification. Local Log: Records the first transmission time, link type used, signature digest, and size of each block; used for resuming interrupted transmissions and troubleshooting disputes. 2) End-side: Multi-link dispatch before transmission and deduplication: Blocks at the base layer with ROI are sent at least once in each of two different communication modes; enhanced layers are sent selectively based on network and cost conditions. Transmission Deduplication: The end-side controls the "backup degree" based on content_id to avoid meaningless duplication of the same block on multiple links.

[0053] 3) Server: Deduplication, verification, and reassembly. Unified entry point and deduplication: All link traffic enters the same receiving layer; deduplication is performed using content_id, duplicate arrivals are counted only once, but the "arrival link type and time" are retained as evidence. Signature verification: Signatures are verified using the device certificate chain; the certificate is bound to the insurance / terminal information when the device is connected. Blocks that fail to be signed are marked as "suspicious," retained but not involved in reassembly. Time verification: TSA tickets are verified first; if only local time is available, "clock deviation description" is recorded in the event report, and updated after the signatures are replenished. Chain integrity check: Check the continuity of "previous fingerprint reference" by window_id; if broken, targeted replenishment is triggered; if it cannot be replenished, the window is marked "chain incomplete" and the evidence level is reduced. Coverage judgment and reassembly: When the receiving coverage of a certain window's base layer reaches a preset threshold, decoding, reassembly, and rendering of that window are performed immediately, prioritizing the restoration of high-priority time periods; the enhancement layer is reassembly performed when conditions permit. Reverse fetch (targeted retransmission): If a critical block is found to be missing or the ROI clarity does not meet the policy threshold, the server initiates a back-to-origin request to the end side, including: target event / window / frame range, priority order (BL-ROI first, then BL non-ROI, then EL), suggested link type (prioritizing different from the previous failure mode), retry limit and validity period. 4) Evidence collection and reporting, recording dimensions: block level (content_id, arrival link / time, signature and time ticket verification status, participation in reorganization), window level (coverage, reorganization time, missing list), link level (successful arrivals, latency percentile, loss estimate, number of reassignments). Output: Automatically generates an "Evidence Chain Integrity Report", including chain verification results, time verification results, deduplication hits, reorganization progress and anomaly descriptions; provides both PDF and JSON formats, and writes back to the claims work order.

[0054] S7 Resume Transmission (Offline Cache → Resume Handshake → Differential Resumption), 1) Local Cache and Index, Storage Unit: Event-level directory + window-level subdirectory; each encoded block is stored as a file or KV item, metadata is the same as above; static encryption is performed using the device key. Index and WAL: Write to a temporary file first, then atomically rename; and write indexes such as content_id to the sequential log (WAL) to ensure recovery after power failure. Priority Queue: The cache manager maintains three queues in the order of "BL-ROI > BL-Normal > EL"; each entry contains the number of sending attempts, the last attempt time, and the expiration time. 2) Offline and Interruption Detection, Interruption Detection: The link is determined to be interrupted by heartbeat loss, continuous timeout of probe packets, or system network status callback; the link is set to "backoff", and block dispatch to it is suspended. Offline Mode: If all links are unavailable, enter the offline cache: collection and encoding continue, all new blocks are entered into the database and marked "unconfirmed"; if the trusted time is unreachable, write "pending signature". 3) Resume handshake and difference synchronization. Resumption trigger: When any link transitions from "backoff" to "ready / active," a handshake is immediately initiated. Handshake content: The endpoint sends a "window list summary" (event ID, list of window IDs, summary or bitmap of content_ids held by each window); the server responds with a "received list," indicating the range of received and missing data. Difference set: The endpoint arranges the "required transmission list" accordingly, strictly following the priority order: BL-ROI → BL-normal → EL; and prioritizes sending communication modes different from the previous failure.

[0055] 4) Resumption scheduling and throttling, rate control: Slow start is used in the initial recovery phase, gradually increasing to the normal rate to avoid conflicts with probes; the transmission rhythm of high-fluctuation links is more conservative. Multi-link parallel processing: If more than one link is being recovered, BL-ROI uses dual communication mode redundancy; remaining blocks are distributed to each link according to real-time weights. Arrival confirmation loop closure: The server confirms receipt based on content_id; the end side marks the block as "confirmed" and removes it from the cache accordingly; unconfirmed blocks proceed to the next round.

[0056] 5) Capacity and Retirement Strategy: Quota: BL and ROI occupy the majority of space; EL only occupies the remaining quota. Retirement order: First delete confirmed blocks; then delete the oldest EL; only BL non-ROI are retired when there is still insufficient space; BL-ROI are only allowed to be deleted if the retention period has expired, and an "irreversible deletion" log is recorded. Power saving mode: When the power is below the threshold, automatic voltage reduction encoding (reduced resolution / frame rate) is implemented, only BL and ROI are cached, and EL is paused.

[0057] Preferably, the first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the triggering time, and the license plate image recognition accuracy.

[0058] Preferably, the end-side convolutional neural network includes a lightweight backbone, sensor feature branches, and a multi-task output head; an amplitude gating unit based on the first semantic importance is set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branches, thereby outputting the final semantic importance, evidence mask, key points, and sharpness evaluation results.

[0059] Preferably, the path-aware link status parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio; and optionally, signal quality indicators, path availability indicators, unit data cost, and unit energy consumption labels are collected; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup rate; the communication mode includes one or more of cellular communication, Wi-Fi direct connection, vehicle-to-everything (V2X), Bluetooth relay, or satellite fallback.

[0060] Preferably, the distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

[0061] This invention also provides a semantically hierarchical multi-path video backhaul system for vehicle insurance evidence collection and claims processing, such as... Figure 2 As shown, the vehicle-side hardware configuration is as follows: 1) Sensing and acquisition domain: external / interior cameras: front, rear, and optional interior; interface is MIPICSI-2 or GMSL2 (coaxial) + deserializer. Microphone array (optional single microphone): PDM / I 2 S, used for acoustic evidence such as collision sounds and glass breakage. Inertial Measurement Unit (IMU): Triaxial Accelerometer / Gyroscope. Vehicle Bus Interface: OBD-II / CAN / LIN / FlexRay, connected via an isolated transceiver; can read airbag screws, I... 2 C / SPI. Positioning and timing GNSS: UARTNMEA + PPS (second pulse). Trigger, braking, vehicle speed, steering angle, etc. Trigger / status inputs: accident hardware button, collision sensor / airbag trigger line (GPIO).

[0062] 2) Computing and Trusted Domain: In-vehicle computing platform SoC: Multi-core CPU + NPU / GPU / DSP, built-in ISP. System memory: LPDDR4 / LPDDR5. Local storage: eMMC / UFS or NVMeSSD (event ring cache, breakpoint resume cache). Trusted and secure unit: SE / TPM or TEE (private key storage, block-level signing, certificate verification). Temperature control and health monitoring: Temperature sensor, fan / hot plate control (optional).

[0063] 3) Communication and Network Domain (Multi-communication Modes): Cellular 4G / 5G Module: PCIe or USB3; Main and Secondary Antennas (MIMO); eSIM / SIM Card Slot. Wi-Fi / Bluetooth Combination Module: SDIO / PCIe + UARTHCI; Supports AP / STA Dual Operation. Vehicle-to-Everything (V2X) Module: DSRC or C-V2X (PC5); PCIe / Ethernet; Independent GNSS Timing Input. Satellite Communication Terminal (Optional): UART / SPI; External Satellite Antenna. Vehicle Ethernet / USB-C: Interconnection with Roadside Units (RSUs), Engineering Maintenance Units, or External Hotspots. 4) Power Supply and HMI Domain: Vehicle Power Supply and Backup: 12V Conversion, Overvoltage and Overcurrent Protection, Ignition Signal Detection, Supercapacitor / Small Capacity Battery (Power-Off Write Disk). HMI: Emergency Button, Indicator Lights, Buzzer, Optional Small Screen.

[0064] Vehicle-side connectivity: Power supply link: Vehicle 12V → Power Management Board → SoC, Camera, Communication Module, GNSS, Storage, SE; During power failure, a supercapacitor maintains critical disk writing and safe shutdown. Video link: Front / rear camera → (GMSL2 coaxial) → Deserializer → SoC's CSI-2; SoC's ISP handles noise reduction / distortion correction → Semantic hierarchical CNN / ROI module → Layered encoder. Audio link: Microphone → PDM / I 2 S → SoC Audio Subsystem → Acoustic Event Labeling. Time Base and Synchronization: GNSSPPS → SoC GPIO generates hard interrupts; NMEA time stamp → clock synchronization service; multi-camera and bus data are stamped according to a unified time base. Vehicle Status: CAN / FlexRay / OBD → isolated transceiver → SoC bus controller; airbag / trigger line → GPIO (high-priority interrupt). Trusted Link: SoC ↔ SE / TPM (I 2C / SPI: Generates / stores keys and signs "encoded block digest + previous sequence digest + time information + event metadata"; certificate chains are bound to devices upon registration. Storage and caching: SoC↔ eMMC / UFS / NVMe (ring video buffer, source fragments, encoded blocks, acknowledgment bitmaps, WAL logs). Multimode network: SoC↔ 5G (PCIe / USB3), Wi-Fi / BT (SDIO / PCIe+UART), V2X (PCIe / Ethernet), satellite (UART); each module has its own external antenna. Link management service cycle detects bandwidth / latency / jitter / loss and cost / energy consumption tags → multipath scheduler allocates "distribution share and redundancy" accordingly. "Source fragments with focus markers" output by the encoder → FEC / fountain code block output → enters "5G queue / Wi-Fi queue / V2X queue / satellite queue" according to priority and link share. Data Flow Overview (End Side): Acquisition Alignment → (Semantic Hierarchical CNN Output Mask and Importance) → ROI Hierarchical Encoding → Fragmentation Marking (BL / EL, Whether ROI, Priority) → FEC Block Generation → Verifiable Encapsulation (Content Fingerprint, Chain Reference, Time Ticket, Terminal Signature) → Multi-Link Parallel Transmission → Arrival Confirmation / Deduplication Receipt → Resume Transmission and Differential Retransmission

[0065] Cloud / Central Side Collaborative Hardware (Paired with Vehicle End), Entry Point and Load Balancing: Multi-regional Access Nodes (Anycast / GLB), TLS Termination. Transmission Access and Deduplication Layer: Deduplication by Content Fingerprint and Recording Arrival Link and Time. FEC Decoding and Reassembly Cluster: CPU / GPU Hybrid, Basic Layer Reassembly Triggered When Window Coverage Meets Standard. Time and Signature Verification: TSA (Timestamp Authority) and HSM (Hardware Security Module) Support Time Credentials and Certificate Chain Verification. Storage and Indexing: Object Storage (Videos and Clips), Time-Series Database / Relational Database (Block-Level Index, Link Log, Confirmation Bitmap). Scheduling and Origin Service: Sending Directed Retransmission Instructions, Cost Policies, and Thresholds to the End Side; Generating an "Evidence Chain Integrity Report" and Writing Back to the Claims Work Order. Observability and Alarms: Link Health, Arrival Rate, Reassembly Latency, and Abnormal Signature Monitoring. Cloud-device connection relationship: Multiple links on the device side go up in parallel → Cloud access deduplication → Verification / reassembly → Source return command and confirmation receipt go down via the same or different links; All interactions are associated with event ID / window ID.

[0066] Alternative implementation (connection differences): Mobile App + OBD, with camera and calculations performed on the phone; OBD-II Bluetooth provides vehicle speed / braking / airbag status to the phone; multiple communication modes primarily use the phone's 4G / 5G / Wi-Fi; the rest of the process is the same. Integrated vehicle infotainment / ADAS gateway, with the camera directly connected to the vehicle infotainment system via GMSL2 / FPD-Link; ADAS events (AEB / FCW) are triggered locally via the internal bus; V2X directly connects to the RSU via the vehicle's Ethernet; other connections are the same as before.

[0067] Centralized SoC Star / Bus Hybrid Topology: The SoC acts as the central hub, connecting downwards to data acquisition and vehicle-side buses, sideways to trusted systems and storage, and upwards to multi-mode communication. Unified Hard Timebase: GNSSPPS + monotonic clock ensures alignment of frames, sensors, signatures, and transmission logs. Multi-Link Decoupling: Each communication module has an independent network interface and transmission queue; link status is provided by a unified "link management service." Verifiable Closed Loop: Block-level fingerprints, chained references, trusted timestamps, and terminal signatures are generated on the edge, with cloud verification and report output, meeting forensic compliance requirements.

[0068] The system includes: a trigger acquisition module: when a collision is triggered, a user reports a case, or a claim work order is issued, the terminal extracts continuous image frames within a preset time window centered on the trigger time; Semantic classification module: The first semantic importance of each frame is calculated on the edge side and input together with the frame data into an improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance. Focus encoding module: Performs hierarchical encoding of ROI constraints on the evidence focus area of ​​each frame to obtain source fragments with focus mask markings; Path awareness module: detects the link status of at least two different communication modes in parallel; Allocation and Loss Prevention Module: Based on the available throughput, round-trip delay, arrival interval fluctuation and loss ratio of each path, determine the distribution share and backup degree for each path; perform out-of-order recoverable forward error correction coding on the source fragments to generate coding blocks, and deliver the coding blocks to multiple paths for parallel transmission according to frame priority; Verifiable backhaul module: Generates content fingerprint for each coded block and writes it into the preceding fingerprint in the order of transmission to form a chain integrity tag, attaches a trusted timestamp and generates a digital signature with the terminal key; The server side performs deduplication, verification and reassembly based on the content identifier, and backhauls missing data to the source according to priority. Resume interrupted download module: When the path is interrupted, unacknowledged data blocks are cached locally, and resumed download based on the arrival of the acknowledgment result after any path is restored to availability.

[0069] Preferably, the first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the triggering time, and the license plate image recognition accuracy.

[0070] Preferably, the end-side convolutional neural network includes a lightweight backbone, sensor feature branches, and a multi-task output head; an amplitude gating unit based on the first semantic importance is set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branches, thereby outputting the final semantic importance, evidence mask, key points, and sharpness evaluation results.

[0071] Preferably, the path-aware link status parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio; and optionally, signal quality indicators, path availability flags, unit data cost, and unit energy consumption tags are collected; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup rate.

[0072] Preferably, the distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

[0073] This invention provides a semantically hierarchical multi-path video backhaul method and system for vehicle insurance evidence collection and claims processing, which achieves the following beneficial technical effects: 1. This solution first obtains the initial semantic importance through lightweight evaluation on the edge side, and then implements gated activation and sensor-driven channel modulation in mid-to-high-order feature layers using an improved convolutional neural network, significantly improving the expressive power of key evidence such as license plates, collision contact areas, and the upper body of pedestrians. During the encoding stage, a layered strategy with ROI constraints is applied to the key evidence: the base layer ensures "clear visibility and priority," while the enhancement layer is responsible for improving overall clarity; in weak network or congestion situations, only the base layer and the key areas are preserved to fully reconstruct the crucial facts. Combined with "focus mask marking," the scheduler sets priority queues and dual communication mode redundancy to avoid evidence gaps caused by single-link interruptions. Compared to overall balanced compression and indiscriminate push, this solution achieves "evidence first, clarity later," improving recognizability at the same bitrate and maintaining credible images under limited bitrates, significantly reducing secondary transmission and communication costs.

[0074] 2. The system of this invention can simultaneously sense the available throughput, round-trip latency, arrival interval fluctuations, and loss ratios of links such as cellular, Wi-Fi direct, V2X, Bluetooth, or satellite, and can combine these with signal quality, cost, and energy consumption tags to form a comprehensive profile. The scheduling end describes the load and redundancy intensity of each link using "distribution share and backup degree," and adopts a windowed strict priority approach: first the basic layer, then the enhancement layer; within the same priority, weighted rotation is performed according to share, and link degradation is immediately reassigned, with differentiated resumption after recovery. Source fragmentation undergoes fountain code-type forward error correction, and reassembly can be performed once the reception coverage meets the standard, avoiding full-segment retransmission; the end side maintains breakpoint resumption and deduplication confirmation to reduce duplicate billing. The policy engine dynamically reduces the enhancement layer ratio of high-cost links based on monthly traffic, power consumption, and hotspot availability, prioritizing the use of low-cost, stable-quality channels to achieve the comprehensive optimization of arrival rate, latency, cost, and energy consumption.

[0075] 3. Each coded block of this invention carries a content fingerprint and references the preceding fingerprint to form a chain-like integrity tag. It also attaches a trusted timestamp and is digitally signed by the terminal key. The server performs deduplication, signing and time verification, chain continuity checks, and quality statistics on arriving data. Reassembly is triggered after the coverage meets a threshold, and missing or unqualified areas are prioritized for retransmission back to the source. The system automatically generates auditable records at the link and window levels, including arrival time, deduplication hits, reassembly progress, and reassignment trajectory, which can be exported as reports for claims and dispute review. Regarding privacy governance, the terminal side minimizes the processing of non-involved faces and license plates and uses local encrypted caching. Transmission and storage are protected by keys throughout. The solution also provides a precise reverse-pull interface with the work order system, realizing a closed-loop retransmission "by event ID and time window," making the evidence chain traceable, verifiable, and operable.

[0076] The above provides a detailed description of a semantic hierarchical multi-path video backhaul method and system for vehicle insurance evidence collection and claims. Specific examples have been used to illustrate the principles and implementation methods of this invention. The descriptions of the above embodiments are only for the purpose of helping to understand the core ideas of this invention. At the same time, for those skilled in the art, there will be changes in the specific implementation methods and application scope based on the ideas and methods of this invention. Therefore, the content of this specification should not be construed as a limitation of this invention.

Claims

1. A semantically hierarchical multi-path video backhaul method for vehicle insurance evidence collection and claims processing, characterized in that, Including the following steps: S1: Triggered Acquisition: When a collision is triggered, a user reports a case, or a claim work order is issued, the terminal extracts continuous image frames within a preset time window centered on the trigger time. S2: Semantic Hierarchy: The edge-side calculates the first semantic importance of each frame and inputs it together with the frame data into the improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance; the edge-side convolutional neural network includes a lightweight backbone, sensor feature branches and a multi-task output head; amplitude gating units based on the first semantic importance are set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branch, thereby outputting the final semantic importance, evidence mask, key points and sharpness evaluation results; S3: Focus Encoding: Perform hierarchical encoding of the evidence focus area of ​​each frame with ROI constraints to obtain source fragments with focus mask markings; S4: Path awareness: Parallel detection of link status in at least two different communication modes; S5: Allocation and Loss Mitigation: Based on the available throughput, round-trip time, arrival interval fluctuation and loss ratio of each link, determine the distribution share and backup degree for each link; perform out-of-order recoverable forward error correction coding on the source fragments to generate coding blocks, and deliver the coding blocks to multiple links for parallel transmission according to frame priority. S6: Verifiable backhaul: Generate a content fingerprint for each coded block and write the preceding fingerprint in the order of transmission to form a chain of integrity tags, attach a trusted timestamp and generate a digital signature with the terminal key; The server performs deduplication, verification and reassembly based on content identifiers, and prioritizes missing data for targeted transmission back to the source. S7: Resume interrupted download: When a path is interrupted, unacknowledged data blocks are cached locally, and resumed based on the arrival of the acknowledgment result after any path becomes available.

2. The semantically graded multi-path video backhaul method for vehicle insurance evidence collection and claims processing as described in claim 1, characterized in that, The first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the trigger time, and the license plate image recognition accuracy.

3. The semantically graded multi-path video backhaul method for vehicle insurance evidence collection and claims processing as described in claim 1, characterized in that, The path-aware link state parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio. It can optionally collect signal quality indicators, path availability indicators, unit data cost and unit energy consumption labels; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup degree; the communication mode includes one or more of cellular communication, Wi-Fi direct connection, vehicle-to-everything (V2X), Bluetooth relay or satellite fallback.

4. The semantically graded multi-path video backhaul method for vehicle insurance evidence collection and claims processing as described in claim 1, characterized in that, The distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

5. A semantically graded multi-path video backhaul system for vehicle insurance evidence collection and claims processing, characterized in that: include: Trigger acquisition module: When a collision is triggered, a user reports a case, or a claim work order is issued, the terminal extracts continuous image frames within a preset time window centered on the trigger time; Semantic grading module: The edge side calculates the first semantic importance of each frame and inputs it together with the frame data into an improved edge-side convolutional neural network to obtain the final semantic importance of each frame; the frames are divided into at least two priority levels according to the importance; the edge-side convolutional neural network includes a lightweight backbone, sensor feature branches, and a multi-task output head; amplitude gating units based on the first semantic importance are set in the mid-to-high-order feature layers of the lightweight backbone to amplify the response of important frames in the evidence focus area and suppress non-critical background; at the same time, the features are dynamically modulated by the channel scaling and bias generated by the sensor branch, thereby outputting the final semantic importance, evidence mask, key points, and sharpness evaluation results; Focus encoding module: Performs hierarchical encoding of ROI constraints on the evidence focus area of ​​each frame to obtain source fragments with focus mask markings; Path awareness module: detects the link status of at least two different communication modes in parallel; Allocation and Loss Prevention Module: Based on the available throughput, round-trip delay, arrival interval fluctuation and loss ratio of each path, determine the distribution share and backup degree for each path; perform out-of-order recoverable forward error correction coding on the source fragments to generate coding blocks, and deliver the coding blocks to multiple paths for parallel transmission according to frame priority; Verifiable backhaul module: Generates content fingerprint for each coded block and writes the preceding fingerprint in the order of transmission to form a chain of integrity tags, attaches a trusted timestamp and generates a digital signature with the terminal key; The server performs deduplication, verification and reassembly based on content identifiers, and prioritizes missing data for targeted transmission back to the source. Resume interrupted download module: When the path is interrupted, unacknowledged data blocks are cached locally, and resumed download based on the arrival of the acknowledgment result after any path is restored to availability.

6. The semantically graded multi-path video backhaul system for vehicle insurance evidence collection and claims processing as described in claim 5, characterized in that, The first semantic importance consists of at least two of the following: the pixel grayscale offset of the pedestrian or vehicle key point, the relative time distance to the trigger time, and the license plate image recognition accuracy.

7. The semantically graded multi-path video backhaul system for vehicle insurance evidence collection and claims processing as described in claim 5, characterized in that, The path-aware link state parameters include at least: available throughput, round-trip time, arrival interval fluctuation, and loss ratio. It can also optionally collect signal quality indicators, path availability flags, unit data cost and unit energy consumption tags; the parameters are acquired at a preset sampling period and smoothed before being used to determine the distribution share and backup degree.

8. The semantic hierarchical multi-path video backhaul system for vehicle insurance evidence collection and claims processing as described in claim 5, characterized in that, The distribution of the encoded blocks adopts a windowed strict priority strategy: within the same window, the base layer queue is scheduled before the enhancement layer queue; under the same priority, weighted rotation is performed according to the distribution share of each path; when the path quality is lower than the threshold, its uncompleted share is immediately reassigned to other paths; the base layer data corresponding to the evidence focus is delivered at least once in two communication modes to improve the arrival rate.

Images