CN115100623A - Unmanned aerial vehicle auxiliary vehicle networking blind area pedestrian detection system based on end-edge-cloud cooperation - Google Patents

Abstract

The invention discloses a blind spot pedestrian detection system based on terminal-edge-cloud collaboration based on unmanned aerial vehicle assisted vehicle networking. The system at least includes a UAV terminal, a mobile edge node, a static edge node and a cloud node, and each node has computing and transmission capabilities. The method includes: 1. UAVs are deployed in the air to monitor the blind spots of vehicles in real time, collect real-time video stream information, and inform the static edge nodes on the roadside of the blind spot pedestrian detection service in time; 2. The static edge nodes need information according to the task , combined with the available computing resources and communication bandwidth of heterogeneous nodes, according to a certain scheduling algorithm to determine which node the task will be offloaded to; 3. The static edge node will notify the UAV of the scheduling result, and the UAV will monitor it through V2X communication. The video stream is sent to the corresponding node; 4. The corresponding node performs the pedestrian detection task according to the deployed model, and sends a warning message to the vehicle when a potential collision danger is detected. The present invention provides a solution for assisting the Internet of Vehicles through unmanned aerial vehicles, and realizes the prevention of dangerous collisions in blind spots of vehicles in the traffic system.

Description

A UAV-assisted Vehicle Networking Blind Spot Pedestrian Detection Based on Device-Edge-Cloud Collaboration system

technical field

The invention relates to unmanned aerial vehicles, Internet of Vehicles, edge computing, task offloading, V2X communication and intelligent transportation systems, and in particular relates to an unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration and a processing method thereof .

Background technique

In recent years, the development of wireless communication and intelligent network technology has promoted the development of vehicular ad hoc networks (VANETs), which are the basis of emerging intelligent transportation systems (ITS). On the other hand, with the popularization of UAV-based applications such as environmental monitoring, intelligent monitoring, wireless communication, and aerial photography, it is a new paradigm for VANET to cooperate with UAVs to realize innovative and powerful information technology services. vision.

The integration of UAVs with VANETs has been extensively studied, including data forwarding, traffic monitoring, computational offloading, and trajectory optimization. In the data forwarding scenario, the UAV is used as the relay node to study data caching and trajectory planning to maximize the network throughput. In traffic monitoring scenarios, UAVs make full use of their mobility to monitor traffic network conditions in real time and report abnormal conditions to the control center. In the computational offloading scenario, the research focus is to consider the communication, storage, and computing resource allocation costs between UAVs and VANETs, reduce computational processing delays, and save UAV energy consumption. In the trajectory optimization scenario, by planning the trajectory of the UAV, the maximum system throughput and service delay reduction when the UAV is integrated with VANET are studied.

In the current traffic system, blind spots in front of vehicles are the main cause of traffic accidents in many cases, and blind spots in vehicles on the road can be fixed and randomly generated. For example, if there is a large passenger car or an obstacle in front of the vehicle at a certain moment and at a certain position, a blind spot will be randomly generated in front of the vehicle. In this case, the vehicle is very likely to cause traffic accidents. At present, ITS mainly uses roadside surveillance cameras to broadcast real-time road conditions. , aims to eliminate this hidden danger, but the blind spot is randomly generated, this method cannot help the vehicle to perform real-time detection and early warning of the blind spot in such a scenario, and the prevention effect of traffic accidents is not good. In addition, the traditional cloud computing architecture, It also cannot provide real-time and efficient services for road vehicles. There are still huge challenges in how to solve the randomness generated by the blind spot of the vehicle's field of vision, and how to solve the problems such as the untimely response of cloud computing architecture tasks.

SUMMARY OF THE INVENTION

In order to solve the above problems, the present invention mainly provides an unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration. In the air, the possible blind spots of vehicles on the road are monitored, and pedestrians can be detected based on a certain target detection model. Meanwhile, drones can communicate with vehicles, roadside infrastructure and cloud servers through V2V (vehicle-to-vehicle), V2I (vehicle-to-infrastructure), and V2C (vehicle-to-cloud) communications, respectively. Therefore, UAVs can offload target detection tasks to moving vehicles, static edge nodes deployed on the roadside, and cloud servers by transmitting the monitored video to the corresponding nodes. Due to the high flexibility of UAVs, the system On the one hand, it solves the problem of randomness in the blind spot of vehicle vision, on the other hand, the use of device-edge-cloud collaboration overcomes the problem of slow response of cloud computing.

In order to solve the above technical problems, the main steps of the unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on the terminal-edge-cloud collaboration provided by the present invention are as follows:

Step 1: Offline model training for real-time monitoring of pedestrians in the blind area. Pedestrian datasets captured by drones are collected for model training, and the trained models are deployed on terminals, mobile edges, static edges, and cloud nodes with heterogeneous computing and communication capabilities. The present invention considers using two classical target detection models, SSD and YOLO. Compared with SSD, YOLO is a larger-scale network, which requires higher computational overhead and has better detection effect; SSD is a single-stage network. Simultaneous implementation of object detection and classification. The lightweight and fast inference properties of SSDs make them more suitable for mobile devices.

The training data set of the system comes from a video of a university campus square shot by a drone. 2000 frames were randomly selected from the video, 90% of which were used for training and the remaining 10% for model validation. The training process set the batch size to 32, the IOU threshold to 0.98, and the training iterations to 5000. Both models are tested in traffic scenarios. The UAV video frames are collected, and the scene simulates the potential blind spot of the vehicle. The final experimental results show that the detection accuracy of YOLO and SSD can be maintained at 95% and 90%, respectively.

Step 2: Deploy the drone in the air to monitor the blind spot of the vehicle in the traffic road. At the same time, the blind spot pedestrian detection service is enabled and the static edge node on the roadside is notified. After the static edge node receives the notification, it starts to update the task offloading strategy online. Specifically, for the method of obtaining video stream data for the UAV, the UAV uses FFmpeg to realize the decoding operation of the video, and converts the obtained decoding result into an image in the YUVImage format, so as to realize the transmission and storage of the image. The final YUV video stream is sent to each node in the system through V2X communication.

Step 3: Provide a method for determining the offloading of the task to a node for execution in combination with the computing resources available to the heterogeneous nodes and the communication bandwidth of the network, and the method includes:

Step 3.1: Under the condition that the node computing, storage and communication resources are constrained, the objective function defined by minimizing the average task delay as the optimization goal:

The node parameters include: definition S={s ₁ ,s ₂ ,…,s _|S| }, M={m ₁ ,m ₂ ,…,m _|M| }, U={u ₁ ,u ₂ ,… , u _|U| } and {c} are the sets of static edge nodes, mobile edge nodes, terminal nodes and cloud servers, respectively;

N={s ₁ ,s ₂ ,…,s _|S| ,m ₁ ,m ₂ ,…,m _|M| ,u ₁ ,u ₂ ,…,u _|U| ,c};

表示；UAV perception for task collection

express;

分别表示输入数据大小、所需CPU周期数、精度要求、生成时间和截止日期；Each task can be represented as a 5-tuple

Respectively represent the input data size, the number of CPU cycles required, the accuracy requirement, the generation time and the deadline;

Each offload node n∈N and 5 tuples <C _n ,B _n ,ρ _n ,S _n ,R _n >representing the radius of communication coverage, total wireless bandwidth, number of channels, maximum storage function, and computing power (for example, the number of CPU cycles per unit time);

Denoting con _u,n,t = 1 means that u∈N can exchange messages with n∈N in single-hop communication at time t;

Use H={h ₁ , h ₂ ,...,h _|H| } to represent the weight of the neural network model;

Suppose the neural network model h _k is used to offload the task w _ji to node n for inference, and a binary variable is defined

Communication delay between nodes:

The sum of the total transmission delay includes the UAV sending video data to the offloading node, and the offloading node sending the calculation results to the two parts of the vehicle. The definition of the total communication delay of task offloading is as follows:

The calculation delay of the task:

The waiting delay of the task:

表示，它是总传输时延、计算时延和等待时延的计算得到:Finally, the task latency of offloading w _ji to node n is given by

means that it is calculated from the total transmission delay, calculation delay and latency delay:

The goal of the system is to minimize the average task delay by determining the offloading strategy, and the objective function is:

The constraints of the objective function include:

Constraints C1 and C2 indicate that each task must be offloaded and computed to a node:

C1: _xj,i,n, h∈{0,1}

C2:∑ _n∈N x _j,i,n,h =1

Constraint C3 represents a storage constraint:

Constraint C4 means that the number of tasks transmitted simultaneously cannot exceed the number of channels of the offloading node:

Constraint C5 means that the selected neural network model meets the accuracy requirements:

Constraint C6 means that the task must be completed by the deadline:

Step 3.2: According to the optimization objective of the objective function, a task offloading algorithm is designed:

The task delay is estimated based on the currently applied uninstall policy. If a predefined threshold is exceeded, a greedy method is used to search for a new offloading strategy. It consists of two strategies, one is a delay-driven strategy, the goal is to minimize the overall task delay of the system; the other is a resource-driven strategy, the goal is to maximize resource utilization.

然后，选择

最小的策略。Delay-driven strategy: Given an offload strategy, the task delay is t _cur . When a new task with data θ _ji and computational requirement c _ji arrives, the offloading delay t _aver is estimated according to the delay model derived above. When the difference between _tcur and _taver exceeds the predefined threshold _tdiff , the algorithm starts to find a new offloading strategy. Specifically, it traverses each computing node and each neural network model to calculate the total task latency

Then, select

minimal strategy.

当系统中节点的R_n增加时，算法开始寻找新的卸载策略。具体更新策略类似于任务驱动时延策略。Resource-driven strategy: Consider maximizing resource utilization, taking computing resource availability as an example. Specifically, given an offloading strategy, the computation delay is

When the R _n of the nodes in the system increases, the algorithm starts to find new offloading strategies. The specific update strategy is similar to the task-driven delay strategy.

Step 4: The drone communicates with other nodes through V2X. Specifically, the drone and the static edge node and the mobile edge node are located in a car networking at the same time, and communicate through methods such as WiFi and DSRC. By communicating with cloud nodes through 4G, 5G or cellular networks, the drone can receive the task offloading node selection results from static edge nodes, and send the collected blind spot video streams to the corresponding nodes for task calculation.

Step 5: According to the pedestrian detection model obtained in step 1, the node of the computing task will deploy the offline trained pedestrian detection model in advance, and the node will select the corresponding deep learning model for inference calculation according to the received message. The two task offloading strategies mentioned in step 3.2 are obtained. After the node obtains the calculation results, considering the limited communication bandwidth and the validity of the results, it will not send all the calculation results to the vehicles on the road, but will only send a warning signal to the vehicle when a collision danger is detected.

Description of drawings

The accompanying drawings of the present invention are described as follows:

Fig. 1 is the system schematic diagram of the present invention;

Fig. 2 is the flow chart of the present invention;

Detailed ways

FIG. 1 is a schematic diagram of the system of the present invention. The figure mainly shows that the terminal-edge-cloud collaborative architecture consists of four basic elements: UAV terminals, vehicles as mobile edge nodes, roadside units as static edge nodes, and cloud servers. Each element has certain computing and communication capabilities, enabling the UAV terminal to communicate with vehicles, roadside units, and the cloud through V2X communication. Under this architecture, the general application scenario is described as follows: The drone is deployed in the air and can monitor the blind spot of the driving vehicle, such as the pedestrian surrounded by the circle in the figure. The monitored video stream can be processed according to a specific object detection model such as YOLO. The model can be deployed on drones, vehicles, roadside units, and the cloud for pedestrian detection. The roadside unit obtains the unloading node of the current task according to a certain unloading algorithm and notifies it to the UAV. The UAV side transmits the video stream to the corresponding node through V2X communication for task calculation. Finally, if a potential pedestrian collision is detected, the warning information is transmitted to the driving vehicle through the task computing node.

FIG. 2 is a flow chart of the specific implementation of the present invention. The following will describe the steps of the system in detail:

In step 101, the pedestrian detection model is trained offline, and two classic models, SSD and YOLO, are used for training. The lightweight characteristics of SSD are more suitable for mobile devices, and YOLO has higher detection accuracy. A large number of traffic roads are collected by drones in For the pedestrian datasets captured in the air, the trained models are deployed on UAV terminals, roadside static edge nodes, vehicle-mounted mobile edge nodes, and remote cloud nodes.

In step 102, the UAV is deployed in the air of the traffic road. Since most UAVs do not have computing power, a Raspberry Pi can be mounted on the UAV as a processor to monitor the presence of blind spots of vehicle vision in the traffic road. .

In step 103, the drone monitors the blind spot conditions in real time, and captures information such as the location of pedestrians, and then uses FFmpeg to decode the video, and converts the decoded result into a YUVImage format image, so as to realize the transmission and storage of the video stream, and at the same time A notification is sent to the roadside unit at the same time as the monitoring of the blind spot is started.

In step 104, the drone communicates with the roadside unit through V2I. There are many channel interference signals in the traffic road, and the phenomenon of packet loss during the transmission process is serious. Using the reliable transmission TCP protocol will greatly improve the transmission delay. In the actual traffic road, the task response rate is required. Therefore, the present invention mainly designs data transmission in the form of UDP broadcast. In the case of ignoring the loss of a small number of packets, the message or data is sent to the roadside unit. Here, the blind spot monitoring is performed. A service-on notification is sent to the roadside unit.

In step 105, the roadside unit receives the notification information sent by the UAV, and the roadside unit starts to update the task unloading strategy in real time. Since the computing power of the roadside unit is stronger than that of the vehicle end and the UAV end, and at the same time it is closer to the actual traffic road than the cloud node, the present invention deploys the task offloading algorithm to the roadside unit in advance.

In step 106, the roadside unit integrates the computing, storage and bandwidth resources of all heterogeneous nodes in the system, determines the objective function that minimizes the average task delay as the optimization goal, and defines the objective function, wherein the definition and constraints are as follows:

Relevant parameters include: define S={s ₁ ,s ₂ ,…,s _|S| }, M={m ₁ ,m ₂ ,…,m _|M| }, U={u ₁ ,u ₂ ,…, u _|U| } and {c} are the sets of static edge nodes, mobile edge nodes, terminal nodes and cloud servers, respectively;

N={s ₁ ,s ₂ ,…,s _|S| ,m ₁ ,m ₂ ,…,m _|M| ,u ₁ ,u ₂ ,…,u _|U| ,c};

表示；UAV perception for task collection

express;

分别表示输入数据大小、所需CPU周期数、精度要求、生成时间和截止日期；Each task can be represented as a 5-tuple

Respectively represent the input data size, the number of CPU cycles required, the accuracy requirement, the generation time and the deadline;

Each offload node n∈N and 5 tuples <C _n ,B _n ,ρ _n ,S _n ,R _n >representing the radius of communication coverage, total wireless bandwidth, number of channels, maximum storage function, and computing power (for example, the number of CPU cycles per unit time);

Denoting con _u,n,t = 1 means that u∈N can exchange messages with n∈N in single-hop communication at time t;

Use H={h ₁ , h ₂ ,...,h _|H| } to represent the weight of the neural network model;

Suppose the neural network model h _k is used to offload the task w _ji to node n for inference, and a binary variable is defined

Communication delay between nodes:

The sum of the total transmission delay includes the UAV sending video data to the offloading node, and the offloading node sending the calculation results to the two parts of the vehicle. The definition of the total communication delay of task offloading is as follows:

The calculation delay of the task:

The waiting delay of the task:

表示，它是总传输时延、计算时延和等待时延的计算得到:Finally, the task latency of offloading w _ji to node n is given by

means that it is calculated from the total transmission delay, calculation delay and latency delay:

The goal of the system is to minimize the average task delay by determining the offloading strategy, and the objective function is:

The constraints of the objective function include:

Constraints C1 and C2 indicate that each task must be offloaded and computed to a node:

C1: _xj,i,n, h∈{0,1}

C2:∑ _n∈N x _j,i,n,h =1

Constraint C3 represents a storage constraint:

Constraint C4 means that the number of tasks transmitted simultaneously cannot exceed the number of channels of the offloading node:

Constraint C5 means that the selected neural network model meets the accuracy requirements:

Constraint C6 means that the task must be completed by the deadline:

In step 107, the system estimates the task delay according to the currently applied uninstall policy. If a predefined threshold is exceeded, a greedy method is used to search for a new offloading strategy. It consists of two strategies, one is a delay-driven strategy, the goal is to minimize the overall task delay of the system; the other is a resource-driven strategy, the goal is to maximize resource utilization.

当系统中节点的R_n增加时，计算当前的卸载时延t_aver，具体更新策略类似于任务驱动时延策略。Given an unloading strategy, the system initially unloads tasks according to the unloading strategy, indicating that the task delay is t _cur . When a new task with data θ _ji and computational requirement c _ji arrives, the offloading delay t _aver is estimated according to the delay model derived above. The same resource-driven strategy considers maximizing resource utilization, taking computing resource availability as an example. Specifically, given an offloading strategy, the computation delay is

When the R _n of the nodes in the system increases, the current unloading delay _taver is calculated, and the specific update strategy is similar to the task-driven delay strategy.

At step 108, when the difference between _tcur and _taver exceeds the predefined threshold _tdiff , the algorithm starts to find a new offloading strategy.

然后，选择贪心的选择

最小的策略，并记录该卸载策略发送给无人机端。In step 109, the roadside unit starts to search for the optimal task offloading strategy in the current environment. Specifically, it traverses each computing node and each neural network model to calculate the total task delay.

Then, choose the greedy choice

Minimal strategy, and record the unloading strategy and send it to the drone.

At step 110, when the difference between t _cur and t _aver does not exceed the threshold, the algorithm continues to use the previous unloading strategy.

In step 111, the UAV receives the unloading strategy notification from the roadside unit, and then sends the collected blind spot video stream to the corresponding node for task calculation. In this process, the UAV communicates with other nodes through V2X. Specifically, the UAV and the static edge node and the mobile edge node are located in the same Internet of Vehicles at the same time, and communicate through methods such as WiFi, DSRC, etc. The computer communicates with the cloud node through 4G, 5G or cellular network.

In step 112, the node receives the video stream from the drone, the node of the computing task will call the pedestrian detection model that has been deployed offline and trained in advance, and the node will select the corresponding deep learning model according to the received message for inference calculation, and the node obtains After calculating the results, considering the limited communication bandwidth and the validity of the results, not all the calculation results will be sent to the vehicles on the road, and only a warning signal will be sent to the vehicle when the danger of pedestrian collision is detected.

In step 113, the computing node detects that there is a danger of pedestrian collision, and sends the warning information to the vehicle. The driver in the vehicle makes corresponding operations according to the warning information, and chooses whether to end the blind spot pedestrian detection service. If the service is not ended, the system will Proceed to step 103.

In step 114, the computing node does not detect that there is a danger of pedestrian collision, and does not send early warning information, and the system will continue to perform the blind spot pedestrian detection service.

Claims (8)

1. A blind spot pedestrian detection system based on end-side-cloud collaboration based on unmanned aerial vehicle-assisted vehicle networking, is characterized in that, comprises the following steps: Step 1. In the offline preparation stage, collect data sets containing pedestrians in dangerous sections of traffic roads, which are used to train pedestrian detection models such as YOLO and SSD. The trained models can be deployed on terminals and mobile edges with heterogeneous computing and communication capabilities. , static edge and cloud nodes. Step 2. Deploy the drone in the air to monitor the blind spot of the vehicle in the traffic road, enable the pedestrian detection service in the blind spot, and notify the static edge nodes on the roadside. Step 3. The roadside static edge node, combined with the available computing resources of all heterogeneous nodes in the current system and the communication bandwidth of the network, determines to offload the task to a certain node to perform the corresponding reasoning task according to the proposed task offloading algorithm. Step 4. The roadside static edge node transmits the unloading result of the task to the UAV, and the UAV sends the monitored video stream to the corresponding node through V2X communication. Step 5. The node that obtains the calculation task performs detection and calculation on pedestrians according to the pedestrian detection model obtained in step 1. If a potential collision danger is detected in the blind spot, it sends a warning message to the vehicle on the road, and obtains the vehicle with the warning signal. , and carry out the corresponding evasion operation according to the actual driving situation. 2. A kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 1, is characterized in that: in step 1, according to the actual situation of the traffic system, simultaneously improve the detection of blind spots. For the accuracy of pedestrian detection, this system considers two classic target detection models, SSD and YOLO. Compared with SSD, YOLO is a larger network, which requires higher computational overhead and has better detection effect; Lightweight and fast inference features make it more suitable for mobile devices. The training data set of the system comes from a video stream of a university campus plaza captured by a drone. 2000 frames were randomly selected from the video, 90% of which were used for training and the remaining 10% for model validation. The training process set the batch size to 32, the IOU threshold to 0.98, and the training iterations to 5000. Finally, according to the experimental test, the detection accuracy of YOLO and SSD can be maintained at 95% and 90%, respectively. 3. A kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 1, the unmanned aerial vehicle in the described step 2 monitors the blind spot in real time, it is characterized in that: to The method of UAV for video stream data acquisition, UAV uses FFmpeg to realize video decoding operation, and converts the obtained decoding results into images in YUVImage format, so as to realize the transmission and storage of images. The final YUV video stream is sent to each node in the system through V2X communication. 4. a kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 1, the computing and bandwidth resources available for integrating heterogeneous nodes in the described step 3, it is characterized in that: in: The node parameters include: define S={s ₁ ,s ₂ ,...,s _|S| }, M={m ₁ ,m ₂ ,...,m _|M| }, U={u ₁ ,u ₂ , ..., u _|U| } and {c} are a set of static edge nodes, mobile edge nodes, terminal nodes and cloud servers, respectively; N={s ₁ ,s ₂ ,…,s _|S| ,m ₁ ,m ₂ ,…,m _|M| ,u ₁ ,u ₂ ,…,u _|U| ,c}; UAV perception for task collection

express; Each task can be represented as a 5-tuple

Respectively represent the input data size, the number of CPU cycles required, the accuracy requirement, the generation time and the deadline; Each offload node n∈N and 5 tuples <C _n ,B _n ,ρ _n ,S _n ,R _n >representing the radius of communication coverage, total wireless bandwidth, number of channels, maximum storage function, and computing power (for example, the number of CPU cycles per unit time); Denoting con _u,n,t = 1 means that u∈N can exchange messages with n∈N in single-hop communication at time t; Use H={h ₁ , h ₂ ,...,h _|H| } to represent the weight of the neural network model; Suppose the neural network model h _k is used to offload the task w _ji to node n for inference, and a binary variable is defined

Communication delay between nodes:

Total communication latency for task offloading:

The calculation delay of the task:

The waiting delay of the task:

Finally, the task latency of offloading w _ji to node n is given by

means that it is calculated from the total transmission delay, calculation delay and latency delay:

The goal of the system is to minimize the average task delay by determining the offloading strategy, and the objective function is:

5. a kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 4, the described objective function defined by the average task delay as the optimization target, is characterized in that: The constraints of the objective function include: Constraints C1 and C2 indicate that each task must be offloaded and computed to a node: C1: _xj,i,n, h∈{0,1} C2:∑ _n∈N x _j,i,n,h =1 Constraint C3 represents a storage constraint: C3:

Constraint C4 means that the number of tasks transmitted simultaneously cannot exceed the number of channels of the offloading node: C4:

Constraint C5 means that the selected neural network model meets the accuracy requirements: C5:

Constraint C6 means that the task must be completed by the deadline: C6:

6. A kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 1, the task unloading algorithm of step 3 is characterized in that: the system is based on the currently applied unloading strategy. Estimate task latency. If a predefined threshold is exceeded, a greedy method is used to search for a new offloading strategy. It consists of two strategies, one is a delay-driven strategy, the goal is to minimize the overall task delay of the system; the other is a resource-driven strategy, the goal is to maximize resource utilization. Delay-driven strategy: Given an offload strategy, the task delay is t _cur . When a new task with data θ _ji and computational requirement c _ji arrives, the offloading delay t _aver is estimated according to the delay model derived above. When the difference between _tcur and _taver exceeds the predefined threshold _tdiff , the algorithm starts to find a new offloading strategy. Specifically, it traverses each computing node and each neural network model to calculate the total task latency

Then, select

minimal strategy. Resource-driven strategy: Consider maximizing resource utilization, taking computing resource availability as an example. Specifically, given an offloading strategy, the computation delay is

When the R _n of the nodes in the system increases, the algorithm starts to find new offloading strategies. The specific update strategy is similar to the task-driven delay strategy. 7. a kind of unmanned aerial vehicle-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud coordination according to claim 1, the unmanned aerial vehicle of described step 4 communicates with other nodes in the system, it is characterized in that : The UAV communicates with other nodes through V2X. Specifically, the UAV and the static edge node and the mobile edge node are located in the same Internet of Vehicles, and communicate through methods such as WiFi, DSRC, etc., and the UAV communicates with the cloud. The nodes communicate through 4G, 5G or cellular networks, and the UAV can receive the task offloading node selection result from the static edge node, and send the collected video stream of the blind spot to the corresponding node for task calculation. 8. A UAV-assisted vehicle networking blind spot pedestrian detection system based on end-edge-cloud collaboration according to claim 1, the node in step 5 detects pedestrians in the blind spot, characterized in that: the node of the calculation task The offline-trained pedestrian detection model will be deployed in advance, and the node will select the corresponding deep learning model for inference calculation according to the received message. The selection result can be obtained from the two task offloading strategies mentioned in claim 6. After the node obtains the calculation results, considering the limited communication bandwidth and the validity of the results, it will not send all the calculation results to the vehicles on the road, but will only send a warning signal to the vehicle when a collision danger is detected.

Images