A task dynamic allocation method for edge computing

Abstract

The application provides a task dynamic allocation method for edge computing, comprising the following steps: determining an obstacle avoidance task priority according to a computing power allocation demand, preferentially allocating edge computing power to a front obstacle avoidance area of a vehicle body, and reallocating the computing power according to a video stream instantaneous surge if a front-view sensor bandwidth occupancy is lower than a preset threshold; calculating remaining computing power resources according to optimized storage allocation, reallocating the remaining computing power resources to the obstacle avoidance task, judging whether the remaining computing power meets a preset real-time decision threshold, optimizing the obstacle avoidance task processing efficiency if the remaining computing power meets the preset real-time decision threshold, and obtaining obstacle avoidance processing computing power; processing video acquisition data according to the obstacle avoidance processing computing power, obtaining a final obstacle avoidance decision through collaborative optimization of real-time obstacle avoidance decision data and computing power scheduling analysis results, and determining a continuously optimized task allocation state according to the final obstacle avoidance decision.

Description

Technical Field

[0001] This invention relates to the field of information technology, and in particular to a method for dynamic task allocation for edge computing. Background Technology

[0002] Edge computing, as a core technology supporting environmental perception in autonomous vehicles, can significantly improve real-time decision-making capabilities by dynamically allocating computing tasks to edge nodes, which is particularly crucial in emergency obstacle avoidance scenarios. Autonomous vehicles rely on high-frequency video acquisition and real-time data processing to cope with complex road conditions and ensure safety and reliability. However, existing methods rely too heavily on static resource allocation strategies, making it difficult to adapt to sudden data surges caused by unexpected events. This leads to uneven distribution of computing power among edge nodes, affecting the response speed of critical tasks. Furthermore, during network outages, local caching strategies often prioritize storing data from a specific area, ignoring the perception needs of other areas, resulting in information loss. In emergency obstacle avoidance scenarios, the core challenge lies in balancing the surge in computing power demand caused by high-frequency video acquisition with the dynamic adjustment of perception data priorities. Sudden surges in video streams can quickly saturate the computing resources of edge nodes. Especially when a sudden obstacle triggers high-frequency acquisition, perception data from the obstacle avoidance area in front of the vehicle needs to be processed first to ensure timely decision-making. However, insufficient priority can lead to delays in obstacle avoidance decisions. For example, when a vehicle is traveling at high speed and an obstacle suddenly appears, the system may be unable to process the forward video stream quickly due to uneven distribution of computing power, delaying the generation of obstacle avoidance commands. Therefore, how to dynamically reduce the video resolution of the distant areas on both sides of the vehicle and reallocate computing power to emergency obstacle avoidance tasks, in order to ensure the real-time perception needs under high-frequency video acquisition and the integrity of information during network interruptions, has become a key issue in the dynamic allocation of environmental perception tasks for autonomous vehicles. Summary of the Invention

[0003] This invention provides a method for dynamic task allocation for edge computing, mainly including:

[0004] Obtain environmental parameters and data processing requirements for vehicle video acquisition, analyze the mapping relationship between the amount of forward-looking perception data and the instantaneous surge in video stream, and determine the computing power allocation requirements for edge nodes;

[0005] Based on the computing power allocation requirements, the priority of the obstacle avoidance task is determined, edge computing power is allocated to the obstacle avoidance area in front of the vehicle, and computing power is reallocated according to the bandwidth occupancy of the forward-looking sensor and the instantaneous surge of the video stream.

[0006] After the computing power is redistributed, the perception data of the obstacle avoidance area in front of the vehicle is obtained, the difference between the data processing time and the transmission delay is calculated to obtain the data processing delay, the bandwidth of the forward-looking sensor is adjusted according to the data processing delay, and a real-time obstacle avoidance decision is generated.

[0007] If a network interruption is detected, the perception data of the obstacle avoidance area in front of the vehicle is cached first. The relationship between the local cache capacity and the storage space usage is analyzed to generate a storage usage analysis result. Based on the storage usage analysis result and the real-time obstacle avoidance decision, the storage priority is adjusted to generate a cache allocation result.

[0008] The storage usage is compressed, the data drop rate of the distant areas on both sides of the vehicle is calculated, the video resolution of the distant areas on both sides of the vehicle is reduced according to the cache allocation result, and the storage space allocation is optimized in combination with the storage usage analysis result to generate an optimized storage allocation.

[0009] The remaining computing power resources are calculated based on the optimized storage allocation and reallocated to the obstacle avoidance task. It is then determined whether the remaining computing power resources meet the preset real-time decision threshold. Based on the determination result, the processing efficiency of the obstacle avoidance task is optimized, and obstacle avoidance processing computing power is generated.

[0010] Based on the obstacle avoidance computing power, the video acquisition data is processed to generate a final obstacle avoidance decision, and the task allocation status for continuous optimization is determined based on the final obstacle avoidance decision.

[0011] Furthermore, the acquisition of environmental parameters and data processing requirements for vehicle video capture, analysis of the mapping relationship between the amount of forward-looking perception data and the instantaneous surge in video stream, and determination of edge node computing power allocation requirements include:

[0012] The system acquires the raw video stream captured by the vehicle-mounted camera, extracts the brightness distribution features of the current frame image, and determines environmental parameters based on the brightness distribution features and image texture features. It then queries a preset encoding parameter mapping table based on the environmental parameters to determine the video encoding format. Based on the video encoding format and target detection accuracy requirements, it determines the video resolution and frame rate parameters, calculates the theoretical data stream rate, and generates a baseline data stream rate. It statistically analyzes the actual video stream data volume within a preset time window, calculates the actual data stream rate, and determines the instantaneous surge state of the video stream based on the ratio of the actual data stream rate to the baseline data stream rate. Based on the instantaneous surge state and duration of the video stream, it calculates the required additional computing power value, obtains the comprehensive resource occupancy rate of edge nodes, calculates the remaining available computing power, and determines the edge node computing power allocation requirements based on the remaining available computing power and the required additional computing power value.

[0013] Furthermore, the step of determining the obstacle avoidance task priority based on computing power allocation requirements, allocating edge computing power to the obstacle avoidance area in front of the vehicle, and reallocating computing power based on the bandwidth occupancy of the forward-looking sensor and instantaneous surges in the video stream includes:

[0014] The system obtains the required computing power allocation value, identifies obstacles in multiple obstacle avoidance areas in front of the vehicle, acquires the distance and speed data of each obstacle, calculates the risk coefficient of each area, and generates an obstacle avoidance task priority sequence based on the risk coefficient. It then allocates edge computing power to the corresponding obstacle avoidance areas according to the obstacle avoidance task priority sequence and records the allocated computing power value of the obstacle avoidance area in front of the vehicle. Finally, it acquires the current data transmission rate of the forward-looking sensor, calculates the bandwidth occupancy ratio, and reallocates computing power to the obstacle avoidance area in front of the vehicle based on the bandwidth occupancy ratio and the instantaneous surge in the video stream.

[0015] Furthermore, the process of acquiring perception data of the obstacle avoidance area in front of the vehicle after the redistribution of computing power, calculating the difference between the data processing time and the transmission delay to obtain the data processing delay, adjusting the bandwidth of the forward-looking sensor based on the data processing delay, and generating a real-time obstacle avoidance decision includes:

[0016] The system acquires radar point cloud data and camera image data of the obstacle avoidance area in front of the vehicle, calculates the data processing time difference, and generates a data processing delay. Based on the data processing delay, it adjusts the sampling frequency and image resolution of the forward-looking sensor to generate a high-frequency perception data stream. It performs state estimation on the high-frequency perception data stream, extracts obstacle position and speed parameters, calculates the collision time, and generates a real-time obstacle avoidance decision that includes obstacle position, speed, and risk level.

[0017] Furthermore, if the detection network is interrupted, the perception data of the obstacle avoidance area in front of the vehicle is cached first, the relationship between local cache capacity and storage space usage is analyzed, a storage usage analysis result is generated, and the storage priority is adjusted according to the storage usage analysis result and the real-time obstacle avoidance decision, generating a cache allocation result, including:

[0018] The system detects network interruptions, acquires radar and image data of the obstacle avoidance area in front of the vehicle, writes them to a local cache queue, calculates the cached data volume and storage device capacity, and calculates cache utilization. It extracts features from adjacent frame images, calculates similarity, and determines data redundancy. It calculates the ratio of data write rate to remaining storage space, generates a storage pressure index, and combines these to form a storage occupancy analysis result. Based on the storage occupancy analysis result and the collision risk level of the real-time obstacle avoidance decision, it calculates the cached data priority value. Based on the priority value, it adjusts the cache queue, deletes low-priority data, allocates and releases space for high-priority data, and generates a cache allocation result.

[0019] Furthermore, if the detection network is interrupted, the perception data of the obstacle avoidance area in front of the vehicle is cached first, the relationship between local cache capacity and storage space usage is analyzed, and storage usage analysis results are generated, including:

[0020] Read cached data from the obstacle avoidance area in front of the vehicle, calculate the distribution density of radar scan points, calculate the point cloud coverage, check data continuity, and generate an area perception accuracy value. Based on the area perception accuracy value and the cached data duration, calculate the data effectiveness and space occupancy rate to generate a cache resource utilization efficiency. Based on the cache resource utilization efficiency, identify and delete duplicate data, and calculate the deduplication rate. Calculate the data growth rate after deduplication, calculate the storage pressure index, and generate a storage occupancy analysis result.

[0021] Furthermore, the process of compressing storage usage, calculating the data drop rate in the distant areas on both sides of the vehicle body, reducing the video resolution of the distant areas on both sides of the vehicle body based on the cache allocation results, and optimizing storage space allocation in conjunction with the storage usage analysis results to generate an optimized storage allocation includes:

[0022] The system acquires the data storage occupancy of the distant areas on both sides of the vehicle body, calculates the ratio of data block size to storage capacity, and determines compression processing parameters; performs frequency domain transformation on the video data, retains low-frequency components, and calculates the data discard rate; adjusts the video resolution according to the data discard rate to generate low-resolution video data and calculates the space release amount; and reallocates storage space to other areas according to the space release amount and storage occupancy ratio to generate an optimized storage allocation.

[0023] Furthermore, the step of calculating the remaining computing resources based on the optimized storage allocation, reallocating them to obstacle avoidance tasks, determining whether the remaining computing resources meet a preset real-time decision threshold, optimizing the obstacle avoidance task processing efficiency based on the determination result, and generating obstacle avoidance processing computing power includes:

[0024] Based on the optimized storage allocation, calculate the remaining computing power resources; obtain the number and distance of obstacles, and calculate the total computing power requirement for the obstacle avoidance task; based on the ratio of the remaining computing power resources to the total computing power requirement, allocate computing time slices to each obstacle to generate obstacle avoidance processing computing power.

[0025] Furthermore, the step of processing the video acquisition data based on the obstacle avoidance computing power to generate a final obstacle avoidance decision, and determining the continuously optimized task allocation state based on the final obstacle avoidance decision, includes:

[0026] Based on the obstacle avoidance processing computing power, the video data stream is processed to identify obstacle outlines and types, extract distance, angle, and speed parameters, and generate preliminary obstacle avoidance data. Based on the preliminary obstacle avoidance data, the collision time and threat level are calculated, and obstacle avoidance action instructions are generated. Based on the execution results of the obstacle avoidance action instructions, the computing power quota and task priority are adjusted to generate a continuously optimized task allocation status.

[0027] The technical solutions provided by the embodiments of the present invention may include the following beneficial effects:

[0028] This invention discloses a dynamic task allocation method for edge computing, addressing the challenges of surging vehicle forward-looking perception data and insufficient computing power at edge nodes. By dynamically adjusting sensor bandwidth and computing power allocation, combined with local caching optimization and data deduplication, it resolves the conflict between data processing latency and storage pressure. First, by analyzing the mapping relationship between instantaneous surges in video streams and environmental parameters, the invention determines computing power allocation requirements, prioritizing real-time decision-making in the obstacle avoidance area in front of the vehicle. When the network is interrupted, forward-looking perception data is cached first, and storage pressure is reduced through data integrity analysis and deduplication. Further, storage space is freed up by compressing and optimizing the resolution of distant view data on both sides of the vehicle. Finally, the remaining computing power is used to optimize obstacle avoidance task processing efficiency, achieving low-latency, high-precision real-time obstacle avoidance decision-making. This invention significantly improves the obstacle avoidance stability and resource utilization efficiency of intelligent vehicles in complex environments. Attached Figure Description

[0029] Figure 1 This is a flowchart of a task dynamic allocation method for edge computing according to the present invention.

[0030] Figure 2 This is a schematic diagram of a dynamic task allocation method for edge computing according to the present invention. Detailed Implementation

[0031] The technical solution of the present invention will be clearly and completely described below with reference to the embodiments. 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.

[0032] like Figure 1-2 This embodiment of a dynamic task allocation method for edge computing may specifically include:

[0033] Step S101: Obtain the environmental parameters and data processing requirements changes during vehicle video acquisition, analyze the mapping relationship between the amount of forward-looking perception data and the instantaneous surge in video stream, and if the instantaneous surge in video stream exceeds a preset threshold, determine that the edge node computing power is insufficient and obtain the computing power allocation requirements.

[0034] The system acquires the raw video stream from the vehicle-mounted camera, extracts the brightness histogram distribution of the current frame, and calculates the illumination intensity value based on the peak position and distribution width of the histogram. Simultaneously, it determines the weather condition type (rain, snow, fog, etc.) based on image texture features and color saturation. Then, it queries a pre-defined encoding parameter mapping table based on the illumination intensity value and weather condition type to determine the corresponding H.264 or H.265 encoding format. For the determined encoding format, based on the target detection accuracy requirements of the vehicle's forward-facing camera, it selects an appropriate resolution from three levels: 1920×1080, 1280×720, and 640×480. Furthermore, based on vehicle speed sensor data and weather condition type, it dynamically adjusts the frame rate parameter within the range of 15fps to 30fps. The theoretical data stream rate is calculated by multiplying the resolution width by the height, the frame rate by the bit depth per pixel, and thus obtaining the baseline data stream rate. A sliding window method is used to count the actual video stream data volume within the last 5 seconds. The actual data stream rate is calculated by dividing the total actual data volume within the current window by the time window length. The ratio of the actual data stream rate to the baseline data stream rate is calculated. If the ratio exceeds 1.5 for three consecutive windows, it is considered a sudden surge in video stream data. The required additional computing power is calculated based on the product of the ratio and the duration. The overall resource utilization rate is calculated by multiplying the CPU utilization rate of the edge node by 0.6 and the GPU utilization rate by 0.4. The remaining available computing power is obtained by subtracting the overall resource utilization rate from 1 and then multiplying it by the total computing power of the edge node. If the remaining available computing power is less than the required additional computing power, the edge node is considered to have insufficient computing power, and a computing power allocation request is generated. The allocation request amount is equal to the difference between the required additional computing power and the remaining available computing power.

[0035] Specifically, in the process of vehicle-mounted video acquisition, obtaining environmental parameters is the foundation for realizing intelligent computing power allocation. The brightness histogram, as an important representation of image statistical features, directly reflects the overall brightness level of the image through its peak position.

[0036] In one possible implementation, when the histogram peak is concentrated in the 0-50 range, it indicates insufficient ambient light, and the system identifies it as nighttime or a dark environment; peaks in the 200-255 range indicate strong light or overexposure. By analyzing the histogram distribution width, a narrow distribution indicates uniform lighting, while a wide distribution indicates complex lighting with strong contrast between light and dark areas. This quantification method provides a reliable basis for subsequent coding parameter selection. Weather condition recognition relies on a comprehensive analysis of image texture and color features. In rainy scenes, images exhibit specific texture patterns, such as striped textures caused by raindrops, while color saturation is generally reduced. Foggy weather is characterized by decreased contrast and blurred distant views. By extracting these features and matching them with preset patterns, the system can accurately identify the current weather type. The coding parameter mapping table is designed to consider the data transmission needs under different environments. H.265 encoding is selected for higher compression ratios in low-light nighttime environments, while H.264 is prioritized for high-definition scenes in sunny weather to ensure real-time performance.

[0037] It should be noted that the baseline data stream rate was calculated using direct mathematical relationships. Taking a 1920×1080 resolution and 30fps frame rate as an example, each frame contains 2,073,600 pixels. If the YUV420 format is used, with each pixel occupying 12 bits, the theoretical data volume per second is 746.496 megabits. This baseline value becomes an important reference standard for subsequently judging data surges. The sliding window mechanism plays a crucial role in data stream monitoring.

[0038] Specifically, the system maintains a 5-second buffer and updates the window content every second. When a vehicle exits the tunnel, the sudden change in light intensity requires the encoder to use more bits to represent the image change, potentially causing the actual data flow rate to instantly reach more than twice the baseline value. Monitoring three consecutive windows ensures the stability of surge detection and avoids misjudgments caused by occasional fluctuations. Edge node computing power is evaluated using a weighted average, with CPU and GPU weights of 0.6 and 0.4 respectively, reflecting the characteristics of the video processing task. The CPU is primarily responsible for video decoding and system scheduling, while the GPU handles deep learning tasks such as object detection. When the overall resource utilization reaches 80%, the remaining 20% ​​of computing power is often insufficient to handle sudden data surges. The generated computing power allocation requirement at this point includes a specific resource gap value.

[0039] Step S102: Determine the priority of obstacle avoidance tasks based on computing power allocation requirements, and prioritize the allocation of edge computing power to the obstacle avoidance area in front of the vehicle. If the bandwidth usage of the forward-looking sensor is lower than the preset threshold, the computing power is redistributed based on the instantaneous surge in the video stream.

[0040] The system acquires the required computing power allocation values, identifies obstacles in three obstacle avoidance zones (front, left front, and right front) of the vehicle, measures the distance to each obstacle using millimeter-wave radar, calculates the collision time using vehicle speed sensor data, and obtains the risk coefficient for each zone by dividing the preset baseline time by the actual collision time. These risk coefficients are then sorted in descending order to generate an obstacle avoidance task priority sequence. Based on this priority sequence, computing resources are allocated sequentially from the available computing power pool at the edge nodes to the corresponding obstacle avoidance zones. The computing power allocated to the forward obstacle avoidance zone is recorded, and the current data transmission rate of the forward-looking sensor is obtained. The ratio of this transmission rate to the maximum bandwidth of the link is calculated. If the ratio is lower than a preset threshold, it indicates available bandwidth. The latest instantaneous surge value of the video stream is read, multiplied by the basic computing power requirement to obtain the supplementary computing power requirement, and some allocated computing power is recovered from other obstacle avoidance zones and prioritized for allocation to the forward obstacle avoidance zone, completing the reallocation of computing power based on the instantaneous surge in the video stream.

[0041] Specifically, the priority of obstacle avoidance tasks is determined based on a quantitative assessment of risk coefficients.

[0042] In one possible implementation, millimeter-wave radar continuously scans a 180-degree area in front of the vehicle, dividing detected obstacles into three zones based on their azimuth: front, left front, and right front. When the vehicle is traveling at 60 km / h, an obstacle 20 meters in front corresponds to a collision time of 1.2 seconds, while obstacles at the same distance to the side front pose a lower actual threat because they are not on the straight path. The risk factor is calculated as the ratio of a baseline time to the actual collision time, with the baseline time typically set at 3 seconds. This means that the risk factor for obstacles in front reaches 2.5, significantly higher than that for the side front zone.

[0043] It should be noted that this risk assessment mechanism fully considers the dynamic characteristics of the vehicle. In sharp turns, obstacles that were originally in front or to the side may quickly enter the main driving path. At this time, the system will dynamically update the risk coefficient of each area. Through continuous monitoring and real-time calculation, the generated obstacle avoidance task priority sequence can accurately reflect the most urgent safety needs. The allocation process of edge computing power follows a priority-driven principle.

[0044] Specifically, when the system identifies the area ahead as having the highest risk factor, the computing resources in the available computing power pool are prioritized for the perception processing tasks in that area. This allocation is not a simple resource division, but rather a dynamic adjustment based on actual processing needs. The computing power support provided to high-risk areas enables higher-frequency target detection and trajectory prediction, increasing the processing capacity from 30 frames per second to 60 frames per second, significantly improving obstacle avoidance response speed. Monitoring the bandwidth utilization ratio provides an important basis for computing power optimization. The data stream generated by the forward-looking sensor under normal operating conditions is relatively stable. When the actual transmission rate only occupies 40% of the link bandwidth, it indicates that there is a large amount of idle transmission capacity. This redundancy is not wasteful, but rather a buffer space reserved to cope with emergencies.

[0045] Understandably, sudden surges in video streams often occur during moments of dramatic environmental change. Scenarios such as a vehicle moving from an underground parking lot into direct sunlight or a sudden downpour can cause a sharp increase in image encoding complexity. The surge magnitude reflects the ratio of the actual data volume to the normal level; when a 2x surge is detected, the system immediately activates a computing power reallocation mechanism. This dynamic allocation of computing power demonstrates the flexibility of resource utilization.

[0046] In one embodiment, the system reclaims 50% of the allocated computing power from a low-risk rearward area, and this resource is quickly replenished in the forward obstacle avoidance area. This allocation is not a simple resource transfer, but an intelligent balancing based on real-time demand. The reclaimed computing power can support the forward area in handling the additional processing burden caused by data surges, ensuring that critical safety functions are not affected.

[0047] Step S103: Obtain the perception data of the obstacle avoidance area in front of the vehicle after the computing power is redistributed. The data processing delay is obtained by calculating the difference between the data processing time and the transmission delay. If the delay is lower than the preset delay threshold, the bandwidth of the forward-looking sensor is adjusted, and real-time obstacle avoidance decision is obtained by optimizing the data processing.

[0048] After reallocating computing power, millimeter-wave radar point cloud data and camera image data of the obstacle avoidance area in front of the vehicle are acquired. The timestamp of the data arriving at the edge node is recorded as the start time, and the timestamp of the output result after target detection processing is recorded as the end time. The difference between the end time and the start time is calculated to obtain the data processing delay. At the same time, the transmission delay value of the data on the transmission link is recorded. If the data processing delay is lower than the preset delay threshold, it indicates that the edge node has additional processing capability. The amount of data that can be increased is calculated based on the difference between the preset delay threshold and the actual processing delay. The sensing bandwidth is adjusted by increasing the sampling frequency of the forward-looking sensor from 15 frames per second to 30 frames per second and increasing the image resolution from 640×480 to 1280×720 to obtain the adjusted high-frequency perception data stream. For high-frequency sensing data streams, Kalman filtering is used to estimate and predict the state of obstacle positions. The filter input is the original detection position and the output is smooth trajectory data. The horizontal and vertical coordinate sequences of obstacles are extracted from the smooth trajectory data. The instantaneous velocity is obtained by dividing the difference between adjacent frame coordinates by the inter-frame time interval. The estimated collision time is obtained by dividing the relative distance between the obstacle and the vehicle by the relative velocity. The risk level is determined based on the collision time, and a real-time obstacle avoidance decision is generated.

[0049] Specifically, accurate measurement of data processing latency is a key foundation for optimizing the entire obstacle avoidance system.

[0050] In one possible implementation, when the point cloud data from the millimeter-wave radar and the image data from the camera arrive at the edge node, the system automatically records a timestamp of the reception time. This timestamp uses a high-precision clock with an accuracy at the microsecond level. After target detection processing, the output includes the coordinates of the identified obstacle bounding boxes and category labels, at which point the timestamp is recorded again. The difference between the two timestamps directly reflects the actual processing capability of the edge node. When the processing latency is only 20 milliseconds, compared to the preset threshold of 100 milliseconds, it indicates that the system has sufficient processing margin. The dynamic adjustment mechanism of the sensing bandwidth fully utilizes this processing margin.

[0051] Specifically, the system calculates the additional data that can be processed based on an 80-millisecond available time window. Increasing the frame rate of the forward-facing camera from 15 frames per second to 30 frames per second means doubling the number of images that need to be processed per second. Simultaneously, the resolution increases from 640×480 to 1280×720, tripling the data volume per frame. This adjustment is not a blind increase, but rather a precise match based on actual processing capabilities, ensuring that the increased data volume can be processed precisely within the available time.

[0052] It's important to note that Kalman filtering plays a crucial role in smoothing obstacle trajectories. This filter comprises two core steps: prediction and update. The prediction step estimates the obstacle's position at the next moment based on its historical motion state. For example, if a millimeter-wave radar detects an obstacle 30 meters in front of the vehicle in the first frame and 28 meters in the second, the filter incorporates this motion trend to predict its position in the third frame. The update step then corrects the predicted value with the actual detection results, eliminating the influence of measurement noise. The filtered trajectory data exhibits smooth and continuous characteristics, avoiding position jumps caused by sensor noise. Instantaneous velocity is calculated based on the position changes between consecutive frames.

[0053] For example, if an obstacle moves from 30 meters to 28 meters within a 0.033-second frame interval, its approach speed is 60 meters per second. This speed value is added to the vehicle's own speed to obtain the true relative motion. The assessment of the collision time directly determines the risk level.

[0054] In one embodiment, the system classifies a collision time of less than 2 seconds as high risk; 2 to 5 seconds as medium risk; and more than 5 seconds as low risk. This tiered assessment provides clear priority guidance for subsequent obstacle avoidance decisions.

[0055] Understandably, the entire processing flow forms a complete data optimization chain. From initial latency measurement to intelligent bandwidth adjustment, and then to precise trajectory tracking, it generates a comprehensive obstacle avoidance decision that includes location, speed, and risk level.

[0056] In step S104, when the network is interrupted, the perception data of the obstacle avoidance area in front of the vehicle is cached first, the relationship between the local cache capacity and the storage space occupation is analyzed, and the storage occupation analysis result is obtained. If the storage space occupation exceeds the preset storage threshold, the storage priority is adjusted based on the real-time obstacle avoidance decision data to obtain the cache allocation result.

[0057] When a network connection interruption is detected, millimeter-wave radar data and camera image data of the obstacle avoidance area in front of the vehicle are acquired and written to the local cache queue in timestamp order. The total number of bytes of data stored in the cache queue is counted, the total capacity and remaining available space of the storage device are read, and the ratio of used space to total capacity is calculated to obtain the cache utilization rate. Feature vectors of adjacent frame images are extracted and cosine similarity is calculated. Data with a similarity exceeding 0.9 is considered redundant. The proportion of redundant data is calculated to obtain the data redundancy. The storage pressure index is obtained by dividing the current data write rate by the remaining available space. The cache utilization rate, data redundancy, and storage pressure index are combined to form the storage occupancy analysis result. If the cache utilization rate exceeds the preset storage threshold, the obstacle avoidance decision information containing collision risk levels is used to assign a weight value of 3 to high-risk data, a weight value of 2 to medium-risk data, and a weight value of 1 to low-risk data. The storage priority value corresponding to each cached data is calculated. The cache queue is sorted in descending order according to the storage priority value. The priority retention threshold is set to 2. Data items with priority values ​​lower than the retention threshold are deleted. The amount of storage space released by the deleted data is calculated. The released space is allocated to newly collected high-priority forward obstacle avoidance area perception data. The amount of data retained, deleted, and remaining available space are recorded to form the cache allocation result.

[0058] Specifically, the data caching mechanism in network outage scenarios reflects the resilient design philosophy of edge computing.

[0059] In one possible implementation, when a vehicle enters a tunnel or underground parking lot causing network signal loss, the system immediately switches to offline caching mode. At this time, the millimeter-wave radar in the forward obstacle avoidance area generates approximately 2 megabytes of point cloud data per second, and the camera generates a video stream at a rate of 30 frames per second, with each compressed frame being approximately 200 kilobytes. This data is strictly ordered according to microsecond-level timestamps to ensure accurate reconstruction of the temporal sequence of events during subsequent playback. The calculation of cache utilization directly reflects the usage status of storage resources.

[0060] Specifically, when an edge node is equipped with 8GB of storage space and 6GB is already in use, the cache utilization rate reaches 75%. This value is not just a simple ratio, but a key indicator that triggers subsequent storage optimization.

[0061] It should be noted that the data redundancy assessment is based on feature vector similarity analysis. The system extracts SIFT feature points from each frame of image, forming a 128-dimensional feature vector. The feature vectors of adjacent frames are used to calculate the similarity value using cosine similarity. When the vehicle is stationary or moving slowly, the content of consecutive frames is almost identical, with a cosine similarity of over 0.95. This highly similar data is marked as redundant, providing a basis for deletion in subsequent storage optimization. The storage pressure index is designed to consider the dynamically changing data flow characteristics. Taking a write rate of 6 megabytes per second and 2GB of remaining available space as an example, the system can support approximately 340 seconds of continuous storage. This time window determines that the system must take optimization measures before exhausting storage. The mapping relationship between collision risk level and storage priority ensures the retention of critical safety data.

[0062] In one embodiment, a rapidly approaching vehicle 20 meters ahead is assessed as high-risk, and its corresponding perception data receives the highest weight value of 3. This means that even with limited storage space, this data will be prioritized for retention. In contrast, data on stationary objects 100 meters to the side or rear only receives a weight value of 1 and may be deleted under storage pressure.

[0063] Understandably, setting priority thresholds requires balancing data integrity and storage capacity. A retention threshold of 2 means that only data of medium to high risk will be retained. This mechanism performs well in practice: when a network outage lasts for 10 minutes, the system frees up 60% of storage space by deleting low-priority data, ensuring the complete recording of all high-risk event data. The generation of cache allocation results not only includes quantity statistics, but more importantly, it forms a traceable data management record. The record shows that 850MB of high-priority data was retained, 1.2GB of low-priority data was deleted, and the remaining available space increased to 3GB.

[0064] Based on the locally cached perception data, the data integrity and coverage of the obstacle avoidance area in front of the vehicle are analyzed, the area perception accuracy is evaluated, the cache utilization rate is calculated based on the area perception accuracy, and the used cache space is compared with the total available space to determine the cache resource utilization efficiency. Based on the utilization efficiency, duplicate perception data is identified and deduplicated. The storage pressure is analyzed using the distribution of the deduplicated data, and the stability of the system under high load is evaluated based on the storage pressure.

[0065] The system reads perception data of the obstacle avoidance area in front of the vehicle from the local cache, calculates the spatial distribution density of millimeter-wave radar scan points in each time period, calculates the ratio of the angle range of point cloud coverage to the preset scanning range, checks for missing data segments in the time series, and evaluates the area perception accuracy value based on the coverage ratio and continuity index. Based on the area perception accuracy value, the ratio of the duration of valid data in the cache to the total recording duration is used as the data effectiveness rate. The system reads the size of the used cache space and the total available space capacity, calculates the ratio of the used space to the total space as the space occupancy rate, and multiplies the data effectiveness rate by the space occupancy rate to obtain the cache resource utilization efficiency. The system judges the resource utilization status based on the cache resource utilization efficiency. If the utilization efficiency is lower than a preset threshold, the cached data is sorted by timestamp, the feature similarity of adjacent data frames is calculated, data with similarity exceeding the threshold is marked as duplicates and deleted, and the difference in data volume before and after deduplication is calculated to obtain the deduplication rate. Using deduplicated data, the amount of data growth per unit time is statistically analyzed, and the ratio of data growth rate to remaining storage space is calculated as a storage pressure indicator. The consumption trend of storage space under the current growth rate is simulated. If the prediction shows that the storage space will be exhausted within a preset time threshold, the system is determined to be unstable under high load. The system operation risk level is assessed based on the storage pressure indicator and the stability judgment results.

[0066] Specifically, assessing the accuracy of regional perception involves multi-dimensional data quality analysis.

[0067] In one possible implementation, the scanning points of the millimeter-wave radar exhibit a specific spatial distribution pattern. When the radar is operating normally, the scanning points should uniformly cover a 180-degree fan-shaped area in front of the vehicle. Spatial distribution density is calculated by dividing the scanning area into several grids and counting the number of point clouds in each grid. Dense areas indicate strong perception capability in that direction, while sparse areas may have perception blind spots. The coverage ratio directly reflects the completeness of perception; when the actual coverage is 170 degrees while the preset requirement is 180 degrees, the coverage ratio is 0.94, indicating a 10-degree perception gap. Checking data continuity reveals the perception quality in the temporal dimension.

[0068] Specifically, the system checks the continuity of data timestamps with millisecond-level precision. Normally, the radar generates a data frame every 50 milliseconds. If no data is found within a 100-millisecond timeframe, it is marked as a missing segment. This missing data may originate from sensor malfunction or data transmission interruption, directly impacting the real-time performance of obstacle avoidance decisions.

[0069] It should be noted that the calculation of cache resource utilization efficiency takes into account both data validity and storage economy. Data validity reflects the proportion of truly valuable data in the cache.

[0070] For example, if the total recording duration is 600 seconds, but only 480 seconds contain valid sensory data, with the remainder being blank or erroneous data, then the data validity rate is 0.8. Space utilization rate assesses resource usage from a physical storage perspective. When 6GB of 8GB cache space is used, the space utilization rate is 0.75. Multiplying the two yields a utilization efficiency of 0.6, meaning that only 60% of storage resources are effectively utilized. Feature similarity calculation provides a scientific basis for data deduplication.

[0071] In one embodiment, the system extracts key features from each frame of data, including the number, location distribution, and speed characteristics of detected obstacles. When a vehicle is stationary at a red light, the perception data from consecutive frames is highly similar, with a similarity of up to 0.98. This duplicate data occupies valuable storage space without providing new information value. By setting a similarity threshold of 0.95, the system can accurately identify and delete redundant data. The storage pressure index is designed to take into account the dynamically changing characteristics of the data flow.

[0072] Understandably, the data growth rate is not constant; in complex traffic scenarios, the amount of data generated per second can reach three times the normal level. When the remaining 2GB of storage space faces a data growth of 10MB per second, the storage pressure reaches a dangerous level. By constructing a consumption trend curve, it is predicted that at the current growth rate, the storage space will be exhausted within 200 seconds. Stability assessment is based on simulation analysis of various extreme scenarios. High load conditions not only include a surge in data volume but also involve competition for computing resources and storage access conflicts. When multiple sensing modules write data simultaneously, storage bandwidth becomes a bottleneck. The system assesses stability by monitoring write latency and queue length. If the prediction shows that the system will reach its resource limit within 300 seconds, it is determined to be in a high-risk state, triggering the corresponding resource allocation mechanism to maintain stable system operation.

[0073] Step S105: Compress the storage usage to obtain the data drop rate of the distant areas on both sides of the vehicle. Reduce the video resolution of the distant areas on both sides of the vehicle according to the cache allocation result. Optimize the storage space release by combining the compression ratio determined by the storage usage analysis result. If the data drop rate is lower than the preset drop threshold, adjust the storage space allocation according to the cache allocation result to obtain the optimized storage allocation.

[0074] The system acquires the video data storage and occupancy status of the distant areas on both sides of the vehicle, calculates the ratio of the current data block size to the preset storage capacity, and determines compression parameters when the ratio exceeds a preset compression threshold. It then performs frequency domain transformation on the distant area video frames using discrete cosine transform, retaining low-frequency components and discarding high-frequency details. The ratio of the number of discarded high-frequency components to the total number of frequency domain components is used as the data discard rate. A resolution adjustment coefficient is determined based on the difference between the data discard rate and a preset standard value. When the data discard rate is higher than the standard value, this adjustment coefficient is used to downsample the original video resolution. A weighted average of adjacent pixels is calculated using bilinear interpolation to obtain low-resolution video data, and the difference in data volume before and after downsampling is calculated to obtain the initial space release amount. Combining the initial space release amount with the storage occupancy ratio of each monitoring area, the available storage quota for the distant area is calculated. If the data discard rate is lower than the preset discard threshold, the released storage quota for the distant area is redistributed to the near-field and mid-field areas according to the weighted proportions of each area's importance, resulting in an optimized storage allocation scheme.

[0075] Specifically, the core challenge in storage management for vehicle monitoring systems lies in how to preserve the most valuable video data within limited storage space. Although distant areas contain environmental information, their importance is generally lower than that of near-field areas, thus making them the focus of storage optimization.

[0076] It should be noted that obtaining video data storage occupancy involves real-time monitoring of data streams in different monitoring areas. A storage occupancy distribution map is created by statistically analyzing the amount of data generated in each area per unit time. When the data block size in a distant area reaches 80% of its allocated capacity, a compression mechanism is triggered. The compression parameters are determined based on the current occupancy ratio; the higher the ratio, the greater the compression intensity.

[0077] In one possible implementation, Discrete Cosine Transform (DCT) serves as the core technology for image compression. Its principle is to transform image data from the spatial domain to the frequency domain. After transformation, each 8x8 pixel block in a video frame has its low-frequency components concentrated in the upper left corner, representing the main structural information of the image; high-frequency components are distributed in the lower right corner, mainly containing details and noise. By setting a frequency threshold, the system retains important low-frequency information while discarding less important high-frequency details. The data discard rate reflects the degree of compression; when the discard rate reaches 40%, it means that 40% of the frequency domain components have been removed, but the visual quality remains acceptable.

[0078] Specifically, determining the resolution adjustment factor requires balancing storage efficiency and image quality. When the data discard rate exceeds a preset standard value of 30%, it indicates that compression is approaching the quality baseline, and further space is saved by reducing the resolution. Bilinear interpolation plays a crucial role in the downsampling process, generating a new pixel value by calculating the weighted average of the four original pixels surrounding the target pixel. The weights are determined based on an inverse distance ratio to ensure a smooth image transition. During the reduction from 1920x1080 to 1280x720, the data volume decreases by approximately 55%, and this difference represents the initial space release. The reallocation of storage quotas reflects the concept of dynamic resource management. The system weighs the actual usage and importance of each area. Near-field areas, containing critical security information, are typically allocated 60% of storage space; mid-field areas, responsible for auxiliary monitoring functions, are allocated 30%; and far-field areas account for 10%. When the far-field area releases 5GB of space through compression, this resource is redistributed to the near-field and mid-field areas in a 7:3 ratio.

[0079] Step S106: Calculate the remaining computing power resources based on the optimized storage allocation, reallocate them to the obstacle avoidance task, determine whether the remaining computing power meets the preset real-time decision threshold, and if so, optimize the processing efficiency of the obstacle avoidance task to obtain the obstacle avoidance processing computing power.

[0080] Based on the processor resources released by the optimized storage allocation scheme, the total computing power capacity of the current vehicle computing platform is calculated. This is subtracted from the computing power already allocated to data processing in each monitoring area to calculate the remaining computing power, which represents the computing capacity available for other tasks. The number of obstacles detected by the vehicle sensors and their relative distances to the vehicle are obtained. The computing power required to avoid each obstacle is calculated based on its distance. All avoidance requirements are summed to obtain the total obstacle avoidance task requirement. The ratio of the remaining computing power to the total obstacle avoidance task requirement is calculated, and it is determined whether this ratio is greater than a preset real-time decision threshold. If the ratio is greater than the preset real-time decision threshold, the proportion of computing power that can be allocated to the obstacle avoidance task is determined based on this ratio. A round-robin approach is used to allocate computing time slices according to the obstacle distance from nearest to farthest, with closer obstacles receiving longer time slices and farther obstacles receiving shorter time slices. The sum of the computing power corresponding to each time slice is the obstacle avoidance processing computing power.

[0081] Specifically, the management of computing resources on in-vehicle computing platforms is directly related to vehicle driving safety, especially in ensuring the real-time performance of obstacle avoidance functions. Once storage optimization frees up some processor resources, this valuable computing power needs to be rationally allocated to more urgent tasks.

[0082] It's important to note that the release of processor resources stems from the ripple effect of optimized storage management. After the video data from distant areas is compressed, the computing units originally used for processing high-resolution video streams are freed up. In-vehicle computing platforms are typically equipped with multi-core processors, with total computing power measured in GFLOPS (Floating-Point Operations Per Second). Data processing in different monitoring areas requires a varying number of computing cores; by monitoring the utilization rate of each core in real time, the current computing power occupancy can be accurately calculated. The remaining computing power is obtained through simple subtraction, representing the system's available computing potential.

[0083] In one possible implementation, the computational requirements for obstacle detection need to consider multiple dimensions. Onboard sensors, including millimeter-wave radar, lidar, and cameras, continuously scan the vehicle's surroundings. When an obstacle is detected, the system records its relative position and motion state. Obstacles within 5 meters of the vehicle are marked as high priority, requiring 30 position updates per second; medium-range obstacles (5 to 20 meters) require 15 updates per second; and long-range obstacles (more than 20 meters) require 5 updates per second. Each processing step includes target recognition, trajectory prediction, and avoidance path calculation, with the computational power consumed proportional to the processing frequency.

[0084] Specifically, the real-time decision threshold is set based on the principle of safety redundancy. This threshold is typically set to 1.5, meaning that the remaining computing power must be at least 1.5 times the total obstacle avoidance task requirement to ensure sufficient processing capacity in case of emergencies. The ratio is calculated by dividing the remaining computing power by the total obstacle avoidance requirement. When 3 near-range, 5 medium-range, and 8 far-range obstacles are detected at a certain moment, the system quickly accumulates the computing power requirements for each type of obstacle to obtain the total requirement. This quantitative method makes computing power allocation decisions based on evidence. The polling scheduling mechanism demonstrates a balance between fairness and real-time performance in computing power allocation. A time slice is allocated to each obstacle processing task, and the length of the time slice is related to the obstacle's priority. Near-range obstacles receive a 10-millisecond computing time slice, during which the processor focuses on obstacle avoidance calculations; medium-range obstacles receive 5 milliseconds; and far-range obstacles receive 2 milliseconds. Within one polling cycle, all obstacles receive a corresponding processing opportunity. The computing power corresponding to the time slice is calculated by multiplying the processor frequency by the time; the sum of the computing power of all time slices is the obstacle avoidance processing computing power.

[0085] Step S107: Based on the obstacle avoidance computing power, process the video acquisition data, obtain the final obstacle avoidance decision through the collaborative optimization of real-time obstacle avoidance decision data and computing power scheduling analysis results, and determine the task allocation status for continuous optimization based on the final obstacle avoidance decision.

[0086] Based on the obstacle avoidance computing power, the video data stream captured by the vehicle-mounted camera is processed in real time. A convolutional neural network is used to identify the outlines and types of obstacles in the video frames, extracting the distance, angle, and speed parameters of the obstacles relative to the vehicle, and outputting preliminary obstacle avoidance data containing obstacle attribute information. The distance and speed parameters of each obstacle in the preliminary obstacle avoidance data are obtained, and the estimated collision time is calculated. The threat level is determined based on the ratio of the collision time to a preset safe time threshold. Obstacle avoidance action commands are generated based on the current computing power utilization rate. If the threat level exceeds the emergency threshold, a braking command is output; otherwise, a steering command is output, resulting in the final obstacle avoidance decision. Based on the actual distance change between the vehicle and the obstacle after the final obstacle avoidance decision is executed, the completion time and computing power resources consumed by the obstacle avoidance action are recorded. By comparing the difference rate between the pre-allocated computing power and the actual consumed computing power, the computing power allocation ratio for subsequent obstacle avoidance tasks is adjusted, the priority ranking of various obstacle avoidance tasks is updated, and the task allocation status for continuous optimization is determined.

[0087] Specifically, the obstacle avoidance function in vehicle-mounted video processing relies on the rational allocation of computing resources and real-time data processing capabilities. Obstacle avoidance computing power, as a dedicated resource for obstacle detection and avoidance tasks, directly determines the system's reaction speed and decision-making accuracy.

[0088] It's important to note that the application of convolutional neural networks (CNNs) in obstacle recognition is based on their powerful feature extraction capabilities. Video data streams are input at 30 frames per second, and each frame undergoes multiple convolutional operations to extract features such as edges, textures, and shapes. The first few layers of the network identify basic features like straight lines and curves, the middle layers combine these features to form contours, and the deeper layers identify the complete obstacle shape. Obstacle attribute information includes type identifiers such as pedestrians, vehicles, and static objects; position coordinates are represented by pixel positions in the image and converted to actual distances using camera calibration parameters; and movement speed is calculated through positional changes between consecutive frames.

[0089] In one possible implementation, the calculation of the estimated collision time involves the principles of relative kinematics. The system acquires the relative distance and relative velocity between the obstacle and the vehicle, and obtains the estimated collision time by dividing the distance by the approaching velocity. When the obstacle approaches at 15 meters per second and is 30 meters away, the collision time is 2 seconds. A preset safe time threshold is typically set to 3 seconds; a ratio less than 1 indicates a collision risk. Threat levels are divided into three levels: a ratio less than 0.3 is the emergency level, 0.3 to 0.7 is the warning level, and 0.7 to 1 is the alert level. Computational utilization reflects the current system load; under high load, emergency threats are prioritized to ensure that critical tasks receive sufficient computing resources.

[0090] Specifically, obstacle avoidance commands are generated following a safety-first principle. Emergency levels trigger braking commands, with the system calculating the required braking force and outputting corresponding control signals. Warning levels generate steering commands, using path planning algorithms to calculate the optimal avoidance trajectory. The steering angle is determined based on the obstacle's location and drivable area, ensuring smooth obstacle avoidance without deviating from the lane. This tiered processing mechanism avoids overreaction and improves driving comfort. Continuous optimization of task allocation reflects the system's adaptive capabilities. Actual distance changes after execution record the obstacle avoidance effect; successful avoidance is indicated by an increased distance or maintaining a safe distance. The system tracks the total time from detection to completion of each obstacle avoidance action, as well as the amount of computing resources consumed. When actual computing power consumption exceeds the pre-allocated amount by 20%, it indicates insufficient quota, and the system automatically increases the computing power quota for that type of task. Conversely, if actual consumption is only 60% of the pre-allocated amount, the quota is appropriately reduced, freeing up resources for other tasks. Priority ranking is dynamically adjusted based on task completion timeliness and resource utilization efficiency, with emergency obstacle avoidance always maintaining the highest priority, while routine monitoring tasks fluctuate according to actual needs.

[0091] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for dynamic task allocation for edge computing, characterized in that, The method includes: Obtain environmental parameters and data processing requirements for vehicle video acquisition, analyze the mapping relationship between the amount of forward-looking perception data and the instantaneous surge in video stream, and determine the computing power allocation requirements for edge nodes; Based on the computing power allocation requirements, the priority of the obstacle avoidance task is determined, edge computing power is allocated to the obstacle avoidance area in front of the vehicle, and computing power is reallocated according to the bandwidth occupancy of the forward-looking sensor and the instantaneous surge of the video stream. After the computing power is redistributed, the perception data of the obstacle avoidance area in front of the vehicle is obtained, the difference between the data processing time and the transmission delay is calculated to obtain the data processing delay, the bandwidth of the forward-looking sensor is adjusted according to the data processing delay, and a real-time obstacle avoidance decision is generated. If a network interruption is detected, the perception data of the obstacle avoidance area in front of the vehicle is cached first. The relationship between the local cache capacity and the storage space usage is analyzed to generate a storage usage analysis result. Based on the storage usage analysis result and the real-time obstacle avoidance decision, the storage priority is adjusted to generate a cache allocation result. The storage usage is compressed, the data drop rate of the distant areas on both sides of the vehicle is calculated, the video resolution of the distant areas on both sides of the vehicle is reduced according to the cache allocation result, and the storage space allocation is optimized in combination with the storage usage analysis result to generate an optimized storage allocation. The remaining computing power resources are calculated based on the optimized storage allocation and reallocated to the obstacle avoidance task. It is then determined whether the remaining computing power resources meet the preset real-time decision threshold. Based on the determination result, the processing efficiency of the obstacle avoidance task is optimized, and obstacle avoidance processing computing power is generated. Based on the obstacle avoidance computing power, the video acquisition data is processed to generate a final obstacle avoidance decision, and the task allocation status for continuous optimization is determined based on the final obstacle avoidance decision.

2. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The process of acquiring environmental parameters and data processing requirements for vehicle video capture, analyzing the mapping relationship between the amount of forward-looking perception data and the instantaneous surge in video stream, and determining the computing power allocation requirements for edge nodes includes: The system acquires the raw video stream captured by the vehicle-mounted camera, extracts the brightness distribution features of the current frame image, and determines environmental parameters based on the brightness distribution features and image texture features. It then queries a preset encoding parameter mapping table based on the environmental parameters to determine the video encoding format. Based on the video encoding format and target detection accuracy requirements, it determines the video resolution and frame rate parameters, calculates the theoretical data stream rate, and generates a baseline data stream rate. It statistically analyzes the actual video stream data volume within a preset time window, calculates the actual data stream rate, and determines the instantaneous surge state of the video stream based on the ratio of the actual data stream rate to the baseline data stream rate. Based on the instantaneous surge state and duration of the video stream, it calculates the required additional computing power value, obtains the comprehensive resource occupancy rate of edge nodes, calculates the remaining available computing power, and determines the edge node computing power allocation requirements based on the remaining available computing power and the required additional computing power value.

3. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The process of determining obstacle avoidance task priorities based on computing power allocation requirements, allocating edge computing power to the obstacle avoidance area in front of the vehicle, and reallocating computing power based on the bandwidth occupancy of the forward-looking sensors and instantaneous surges in the video stream includes: The system obtains the required computing power allocation value, identifies obstacles in multiple obstacle avoidance areas in front of the vehicle, acquires the distance and speed data of each obstacle, calculates the risk coefficient of each area, and generates an obstacle avoidance task priority sequence based on the risk coefficient. It then allocates edge computing power to the corresponding obstacle avoidance areas according to the obstacle avoidance task priority sequence and records the allocated computing power value of the obstacle avoidance area in front of the vehicle. Finally, it acquires the current data transmission rate of the forward-looking sensor, calculates the bandwidth occupancy ratio, and reallocates computing power to the obstacle avoidance area in front of the vehicle based on the bandwidth occupancy ratio and the instantaneous surge in the video stream.

4. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The process of acquiring perception data of the obstacle avoidance area in front of the vehicle after redistributing computing power, calculating the difference between data processing time and transmission delay to obtain the data processing delay, adjusting the bandwidth of the forward-looking sensor based on the data processing delay, and generating real-time obstacle avoidance decisions includes: The system acquires radar point cloud data and camera image data of the obstacle avoidance area in front of the vehicle, calculates the data processing time difference, and generates a data processing delay. Based on the data processing delay, it adjusts the sampling frequency and image resolution of the forward-looking sensor to generate a high-frequency perception data stream. It performs state estimation on the high-frequency perception data stream, extracts obstacle position and speed parameters, calculates the collision time, and generates a real-time obstacle avoidance decision that includes obstacle position, speed, and risk level.

5. The task dynamic allocation method for edge computing according to claim 1, characterized in that, When the detection network is interrupted, the perception data of the obstacle avoidance area in front of the vehicle is cached first. The relationship between local cache capacity and storage space usage is analyzed to generate storage usage analysis results. Based on the storage usage analysis results and the real-time obstacle avoidance decision, the storage priority is adjusted to generate cache allocation results, including: The system detects network interruptions, acquires radar and image data of the obstacle avoidance area in front of the vehicle, writes them to a local cache queue, calculates the cached data volume and storage device capacity, and calculates cache utilization. It extracts features from adjacent frame images, calculates similarity, and determines data redundancy. It calculates the ratio of data write rate to remaining storage space, generates a storage pressure index, and combines these to form a storage occupancy analysis result. Based on the storage occupancy analysis result and the collision risk level of the real-time obstacle avoidance decision, it calculates the cached data priority value. Based on the priority value, it adjusts the cache queue, deletes low-priority data, allocates and releases space for high-priority data, and generates a cache allocation result.

6. The task dynamic allocation method for edge computing according to claim 1, characterized in that, If the detection network is interrupted, the perception data of the obstacle avoidance area in front of the vehicle will be cached first. The relationship between the local cache capacity and storage space usage will be analyzed to generate storage usage analysis results, including: Read cached data from the obstacle avoidance area in front of the vehicle, calculate the distribution density of radar scan points, calculate the point cloud coverage, check data continuity, and generate an area perception accuracy value. Based on the area perception accuracy value and the cached data duration, calculate the data effectiveness and space occupancy rate to generate a cache resource utilization efficiency. Based on the cache resource utilization efficiency, identify and delete duplicate data, and calculate the deduplication rate. Calculate the data growth rate after deduplication, calculate the storage pressure index, and generate a storage occupancy analysis result.

7. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The process of compressing storage usage, calculating the data drop rate in the distant areas on both sides of the vehicle, reducing the video resolution in the distant areas on both sides of the vehicle based on the cache allocation results, and optimizing storage space allocation based on the storage usage analysis results to generate an optimized storage allocation includes: The system acquires the data storage occupancy of the distant areas on both sides of the vehicle body, calculates the ratio of data block size to storage capacity, and determines compression processing parameters; performs frequency domain transformation on the video data, retains low-frequency components, and calculates the data discard rate; adjusts the video resolution according to the data discard rate to generate low-resolution video data and calculates the space release amount; and reallocates storage space to other areas according to the space release amount and storage occupancy ratio to generate an optimized storage allocation.

8. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The process of calculating the remaining computing resources based on the optimized storage allocation, reallocating them to obstacle avoidance tasks, determining whether the remaining computing resources meet a preset real-time decision threshold, optimizing the processing efficiency of obstacle avoidance tasks based on the determination result, and generating obstacle avoidance processing computing power includes: Based on the optimized storage allocation, calculate the remaining computing power resources; obtain the number and distance of obstacles, and calculate the total computing power requirement for the obstacle avoidance task; based on the ratio of the remaining computing power resources to the total computing power requirement, allocate computing time slices to each obstacle to generate obstacle avoidance processing computing power.

9. The task dynamic allocation method for edge computing according to claim 1, characterized in that, The step of processing video acquisition data based on the obstacle avoidance computing power, generating a final obstacle avoidance decision, and determining a continuously optimized task allocation state based on the final obstacle avoidance decision includes: Based on the obstacle avoidance processing computing power, the video data stream is processed to identify obstacle outlines and types, extract distance, angle, and speed parameters, and generate preliminary obstacle avoidance data. Based on the preliminary obstacle avoidance data, the collision time and threat level are calculated, and obstacle avoidance action instructions are generated. Based on the execution results of the obstacle avoidance action instructions, the computing power quota and task priority are adjusted to generate a continuously optimized task allocation status.

Images