A two-stage task cooperation offloading method in Internet of Vehicles

Abstract

The application discloses a two-stage task cooperation unloading method in Internet of Vehicles, designs a task cooperation unloading framework fusing LTE-V direct connection mode and LTE-V cellular mode, and arranges a mobile edge computing server through a wired connection beside a roadside unit. The mobile edge computing server at the roadside unit side cooperates with an idle service vehicle to provide a computing service. Meanwhile, a two-stage unloading decision scheme is provided. A link quality evaluation model is used to evaluate the link quality so as to improve the data transmission stability between a predicted computing node and a request vehicle. Then, an unloading decision model is used to predict a response time, and an unloading decision is made according to the predicted response time. Compared with an existing scheme, the unloading decision scheme is more effective, can reduce the response time and save energy consumption.

Description

TECHNICAL FIELD

[0001] The present application relates to a cooperative computing offloading method, in particular to a two-stage task cooperation offloading method in Internet of Vehicles. BACKGROUND

[0002] With the popularity of Internet of Things (IoT), the configuration of vehicles has become more complex, and the applications supported by Vehicle Terminals (VTs) have become more and more intensive, such as high-precision navigation, high-definition video streaming and live streaming. However, the resources and processing power are still limited, and cannot meet the needs of computationally intensive and time-sensitive tasks.

[0003] The central cloud has abundant resources, and vehicle tasks can be offloaded to the cloud center, but remote transmission brings greater latency and higher transmission cost. In addition, the performance of cloud-based offloading technology is limited by the real-time changing wireless environment and communication requirements, and the Quality of Service (QoS) of Non-Line of Sight (NLOS) is easily affected by external environmental factors (reference [1]). Vehicle Cloud Computing (VCC) integrates vehicle ad hoc networks and cloud computing (reference [2]), which can alleviate these problems. VCC also realizes the computing cooperation between the cloud center and VTs through Vehicle-to-Vehicle (V2V) and Vehicle-to-Infrastructure (V2I) technologies, where VTs serve as computing carriers, which can improve resource utilization and computing efficiency. However, due to the limitations of backhaul and backbone network transmission rate and time fluctuations, VCC cannot meet the high requirements of time-sensitive and computationally intensive tasks (reference [3]). In recent years, Mobile Edge Computing (MEC) has received extensive attention in the 5 / 6G field as a potential solution, which deploys communication, computing and storage capabilities close to VTs. Distributed edge servers reduce network congestion and reduce computing time (reference [4]). In Vehicle Edge Computing (VEC), tasks are offloaded to MEC servers through Roadside Units (RSUs), and MEC servers execute tasks and return computing results (reference [5]). Compared with VCC, VEC greatly reduces the transmission time. In addition, VEC expands the computing capacity of vehicles and saves CPU cycles, further ensuring the quality of service requirements.

[0004] Internet of Vehicles (IoV) is a key technology that can meet both comfortable driving experience and high-quality entertainment services. In recent years, the increase in the number of vehicles has brought challenges to scarce wireless resources, limited computing resources and the processing power of VTs, increasing the response time. Therefore, it is crucial to reasonably allocate MEC computing to reduce the response time. SUMMARY

[0005] The technical problem solved by the present application is to provide a two-stage task cooperation offloading method in vehicle networking, which can reduce the response time of task offloading and save energy consumption.

[0006] The technical solution adopted by the present application to solve the above technical problem is: a two-stage task cooperation offloading method in vehicle networking, all vehicles are equipped with cellular interfaces, and each vehicle is provided with sensors for sensing position, speed and direction, a task cooperation offloading framework integrating LTE-V direct connection mode and LTE-V cellular mode is established, in the task cooperation offloading framework, computing nodes are composed of roadside unit side mobile edge computing servers and service vehicles, when in the LTE-V direct connection mode, the computing node is the service vehicle, and when in the LTE-V cellular mode, the computing node is the roadside unit side mobile edge computing server, and the specific offloading method is:

[0007] a link quality evaluation model for predicting the position of the vehicle in the next time slot and evaluating the link quality, and an offloading decision model for predicting the response time, in the data request stage, the request vehicle sends a HELLO information packet to the adjacent computing node, the link quality is obtained by the link quality evaluation model, when the task needs to be offloaded to the computing node, the offloading decision model obtains the computing node with the shortest response time according to the link quality, the task input data volume, the maximum tolerable time, the available CPU resources of the computing node and the four prediction variables of the wireless channel, the request information is sent to the computing node with the shortest response time by the request vehicle, when the computing node is the roadside unit side mobile edge computing server, the roadside unit side mobile edge computing server directly provides computing services to the request vehicle, when the computing node is the service vehicle, after obtaining the request information, the service vehicle judges whether it has sufficient computing resources, and feeds back the resource status to the request vehicle, and the request vehicle decides whether to offload the task.

[0008] Compared with the prior art, the present application has the following advantages:

[0009] 1) a task cooperation offloading framework integrating LTE-V direct connection mode and LTE-V cellular mode is designed, the mobile edge computing server is deployed beside the roadside unit through a wired connection, and the roadside unit side mobile edge computing server and the idle service vehicle cooperate to provide computing services;

[0010] 2) a two-stage offloading decision scheme is proposed, the link quality is evaluated by the link quality evaluation model to improve the data transmission stability between the predicted computing node and the request vehicle, and then the response time is predicted by the offloading decision model, and the offloading decision is made according to the predicted response time.

[0011] 3) Simulation results show that the proposed offloading decision scheme is more effective than existing schemes such as linear regression prediction algorithm and combination prediction offloading algorithm, which can reduce response time and save energy consumption.

[0012] Preferably, in the place of turning or obstacle shielding, the request vehicle is in the LTE-V cellular mode, the request vehicle sends a HELLO information packet to the roadside unit through the cellular network, and the mobile edge computing server on the roadside unit side directly provides computing services for the request vehicle.

[0013] Preferably, when the request vehicle is in the LTE-V direct connection mode, the information of the adjacent service vehicle is obtained in the alternative network topology scenario.

[0014] Preferably, the content of the HELLO information packet includes broadcast ID, location information, requested task information, and list information of adjacent computing nodes.

[0015] Preferably, the link quality evaluation model is:

[0016] wherein E ij represents the link quality, T s represents a fixed time interval, T u represents an update period, represents the proportion of T s in the update period T u , p M represents the number of HELLO information packets in time T s , R ij represents the distance between the request vehicle and the computing node

[0017]

[0018] wherein (a i , b i ) represents the longitude and latitude of the current position of the request vehicle n i , (a′ i , b′ i ) represents the longitude and latitude of the predicted position of the request vehicle n i , v i represents the speed of the request vehicle n i , v im represents the maximum allowed speed of the request vehicle n i , θ i represents the angle of the request vehicle n i , (a j , b j ) represents the computing node nj longitude and latitude of the current position, (a' j ,b' j ) represents the longitude and latitude of the predicted position of the computing node n j . v j represents the speed of the computing node n j , θ j represents the angle of the computing node n j , v jm represents the maximum speed allowed for the computing node n j .

[0019] Preferably, the offloading decision model is a deep neural network composed of an input layer, a hidden layer and an output layer, the input layer is a task feature matrix composed of link quality E ij , task input data volume maximum tolerable time available CPU resources of the computing node and wireless channel b ij . The number of hidden layers is 10. BRIEF DESCRIPTION OF DRAWINGS

[0020] Fig. 1 It is the task coordination offloading framework of the fusion of LTE-V direct connection mode and LTE-V cellular mode proposed in the application;

[0021] Fig. 2 It is a two-stage offloading decision scheme diagram proposed in the application, wherein (a) is a link quality evaluation model, and (b) is an offloading decision model;

[0022] Fig. 3 It is a comparison diagram of position prediction results of the application and prior art under different iteration numbers, wherein (a) is the predicted longitude, and (b) is the predicted latitude;

[0023] Fig. 4 It is a comparison diagram of root mean square errors of the application and prior art under different iteration numbers, wherein (a) is the predicted longitude, and (b) is the predicted latitude;

[0024] Fig. 5 It is a comparison diagram of response times of the application and prior art under different iteration numbers;

[0025] Fig. 6 It is a comparison diagram of average delays of the application and prior art under different iteration numbers;

[0026] Fig. 7 It is a comparison diagram of average delays of the application and prior art under different vehicle numbers;

[0027] Fig. 8 The average latency comparison chart of the present application and the prior art under different task data volumes;

[0028] Fig. 9 The energy consumption comparison chart of the present application and the prior art under different task data volumes. DETAILED DESCRIPTION

[0029] The present application will be further described in detail below with reference to the embodiments of the drawings.

[0030] Embodiment: A two-stage task cooperation offloading method in Internet of Vehicles, all vehicles are equipped with cellular interfaces, and each vehicle is provided with sensors for sensing position, speed and direction, a task cooperation offloading framework integrating LTE-V direct mode and LTE-V cellular mode is established, in the task cooperation offloading framework, the computing nodes are composed of mobile edge computing servers on the roadside unit side and service vehicles, when in LTE-V direct mode, the computing nodes are service vehicles, in the alternative network topology scenario, the information of all service vehicles in the vicinity is obtained, in the place of turning or obstacle shielding, the vehicle is requested to be in LTE-V cellular mode, the vehicle is requested to send data information to the roadside unit through the cellular network, the mobile edge computing server on the roadside unit side is directly used as the computing node to provide computing services for the requesting vehicle, and the specific offloading method is as follows:

[0031] A link quality evaluation model for predicting the position of the vehicle in the next time slot and evaluating the link quality, and an offloading decision model for predicting the response time are established, the link quality evaluation model is:

[0032] wherein, E ij represents the link quality, T s represents a fixed time interval, T u represents an update period, represents the proportion of T s in the update period T u , p M represents the number of HELLO information packets in the time T s , and R ij represents the distance between the requesting vehicle and the computing node

[0033]

[0034] wherein, (a i , b i ) represents the longitude and latitude of the current position of the requesting vehicle n i , (a′ i , b′ i ) represents the longitude and latitude of the current position of the requesting vehicle n iPredict the longitude and latitude of the location. v i Indicates the request for vehicle n i speed, v im Indicates the request for vehicle n i Maximum permissible speed, θ i Indicates the request for vehicle n i Angle, (a j ,b j ) represents the computation node n j The longitude and latitude of the current location, (a' j ,b' j ) represents the computation node n j Predict the longitude and latitude of the location. v j Represents the computation node n j velocity, θ j Represents the computation node n j Angle, v jm Represents the computation node n j The maximum allowed speed, when in LTE-V cellular mode, is zero due to fixed location; therefore, the distance between the requesting vehicle and the computing node is:

[0035] The longitude and latitude of the requested vehicle and computing node's predicted locations are obtained from an evolutionary neural network. In this network, a weight-adaptive particle swarm optimization algorithm is used to initialize the feedback neural network, and the weight update formula is expressed as:

[0036]

[0037] Where w(t) is the inertia weight, which lies between the maximum value w max With minimum value w min Between d1 and d2, d1 and d2 are control factors, and t is the current response time. max This is the maximum tolerable response time. Using the loss function of a weight-adaptive particle swarm optimization algorithm as the training error, the optimal weights and threshold of the feedback neural network are determined through a search. The loss function f... i Defined as:

[0038]

[0039] Where D is the particle dimension. It is the expected output value of the training data, x dThis is the actual output value of the model. In this way, the optimal weights and thresholds that minimize training error can be effectively found, thereby optimizing the performance of the feedback neural network. Therefore, by inputting the vehicle's current position into the optimized feedback neural network, the requesting vehicle and the computing node can be predicted in update period T. u The position after (a′) i b′ i ) and (a' j ,b' j The offloading decision model is a deep neural network consisting of an input layer, hidden layers, and an output layer. The input layer is determined by the link quality E. ij Task input data volume Maximum tolerable time Available CPU resources of compute nodes and wireless channel b ij The constructed task feature matrix The number of hidden layers is 10.

[0040] During the data request phase, the requesting vehicle sends a HELLO message packet to nearby computing nodes. The HELLO message packet includes a broadcast ID, location information, requested task information, and a list of nearby computing nodes. Link quality is determined by a link quality assessment model. When a task needs to be offloaded to a computing node, the offloading decision model determines the offloading based on the link quality E. ij and the amount of input data for the task Maximum tolerable time Available CPU resources of compute nodes and wireless channel b ij The four predictor variables determine the computing node with the shortest response time. The requesting vehicle sends the request information to the computing node with the shortest response time. When the computing node with the shortest response time is the mobile edge computing server on the roadside unit side, the mobile edge computing server on the roadside unit side directly provides computing services to the requesting vehicle. When the computing node with the shortest response time is the service vehicle, after receiving the request information, the service vehicle determines whether it has sufficient computing resources and feeds back the resource status to the requesting vehicle, which then decides whether to unload the task.

[0041] The method of the present invention was simulated, and the results and performance analysis are as follows.

[0042] A simulation environment is established on the MATLAB platform to evaluate the performance of the offloading mechanism. It is assumed that the configuration of the vehicle conforms to the LTE-V standard, and the parameter settings of the vehicle are consistent with a medium-sized SUV. The computing power of each vehicle comes from reference [6]. The environment deployment of the MEC server conforms to the provisions of the MEC white paper. The data set of the vehicle path prediction model (references [7]-[8]) comes from the Microsoft T-Drive project. This data set contains trajectory data of 10,357 taxis in Beijing in 2008. The data set contains 15 million coordinate points, and the total trajectory distance exceeds 9 million kilometers. The task data set referred to in the response time prediction comes from reference [9]. The basic simulation parameters are listed in Table 1.

[0043] Table 1

[0044]

[0045]

[0046] 5.2 Prediction performance evaluation

[0047] The present application proposes a two-stage task offloading decision algorithm, including a link quality evaluation model based on ENN and a predicted response time offloading decision model. In the first stage, ENN predicts the future position of the vehicle, and the link quality evaluation model calculates the link quality between the node and the vehicle after a certain time; in the second stage, the link quality and the available resource size of the candidate node are input into the deep neural network model to predict the response time. Fig. 3 And Fig. 4 The accuracy of the evolutionary neural network (ENN) in predicting the position is evaluated, Fig. 5 The accuracy of the deep neural network in predicting the response time is evaluated.

[0048] Fig. 3 The position prediction performance of the three models, i.e. the ENN model, the classic BP network model and the AMRIA model (reference

[10] ), is compared. It can be seen that the ENN model has high precision in predicting longitude and latitude by adjusting the initial weight and threshold, and has good fitting effect. The classic BP neural network has better nonlinear processing ability, but due to the random initialization of the value, the algorithm is easy to fall into local optimum.

[0049] Fig. 4 The root mean square error (RMSE) of the three models is compared, and the RMSE is the inconsistency between the actual value and the predicted value. The closer the RMSE value is to zero, the better the performance is, which can be calculated by the following formula:

[0050]

[0051] where N is the total number of sampling data, O i is the true value, P i is the predicted value.

[0052] From Fig. 4 it can be seen that the RMSE values of ENN, classic BP and AMRIA algorithms become smaller as the number of iterations increases. Compared with other models, ENN has the best effect and faster convergence speed.

[0053] 2) Response time prediction

[0054] As Fig. 5 shown, we compared the prediction performance of the proposed two-stage prediction model with the reference

[11] . Our solution has a better fitting effect on the actual response time. The reason is that by fusing the computing power and communication status, our solution adopts a two-stage response time to learn information. In addition, when dealing with multiple variables, the nonlinear mapping characteristics of ENN are superior to linear regression. Therefore, the two-stage prediction model can make more accurate offloading decisions.

[0055] 5.3 Offloading decision performance analysis

[0056] Task offloading is an effective way to alleviate the limited computing power of vehicles, and time and energy consumption are key factors to evaluate the performance of offloading. In this subsection, the proposed two-stage offloading decision algorithm (TSODA) is compared with the offloading performance of the following schemes:

[0057] Linear regression prediction algorithm (LRP): According to the relationship between historical state and response time, a linear regression equation is established for statistical analysis.

[0058] Combined mode offloading algorithm (CMR) reference [4]: By directly offloading through V2I mode or predicting relay transmission through V2V mode, the task is adaptively offloaded to the MEC server.

[0059] GPSR algorithm (GPSR) reference

[10] : Only according to the direct neighbor information of the router to make offloading decisions.

[0060] In Fig. 6In Fig. 6, we show the average time for a task size of a certain value. It can be seen that the average time of task offloading is relatively stable except for GPSR. TSODA performs better than the other three algorithms because it first considers the communication link quality between nodes to ensure higher reliability, thereby effectively reducing the number of retransmissions and transmission time; then selects the optimal offloading node based on the response time, further reducing the computation time. GPSR extends an arbitrary routing mechanism based on the right-hand rule, resulting in more routing hops and leading to greater transmission time, so GPSR has the largest average time. CMR can find the next-hop node to transmit the task forward, taking into account the available computing resources and the distance between adjacent vehicles. However, due to the limitations of environmental changes, the optimal relay node cannot be obtained, resulting in increased transmission time. In addition, the prediction accuracy of LRP is lower than ENN, so the average task execution time is higher.

[0061] In Fig. 7 Fig. 7, we show the average time as the number of VTs increases. The same trend verifies that the increase in the number of vehicles makes computing resources scarce and enhances interference between vehicles. The curve of ENN is much lower than that of CMR, indicating the effectiveness of the two-stage response time prediction. When the candidate computing nodes are sparsely distributed, there are fewer available adjacent nodes, and the movement of vehicles further exacerbates routing failure. Therefore, in the GPSR algorithm, the time is the longest because the vehicles are assigned to imperfect routing relays. The interaction between variables directly affects the prediction results of LRP. In addition, LRP cannot handle the nonlinear causal relationship between variables.

[0062] Fig. 8 and Fig. 9 show the average time and energy consumption of VTs when the task size is a certain value, respectively. The performance improvement of the proposed ENN scheme is obvious. In Fig. 8 our mechanism adopts a collaborative offloading scheme, and fine-grained subtasks are executed in parallel on different computing nodes, greatly reducing transmission time and execution time. GPSR performs the worst because the packet loss rate and timeout rate between routing packets increase.

[0063] The energy consumption formula is as follows: where z represents the energy consumption per unit CPU cycle, a j represents the computing capacity of the vehicle. Energy consumption is related to the available resources of the computing node and the available resources required by the task. Through the response time prediction model, the computing node with the minimum energy consumption can be selected while meeting the time requirement.

[0064] In Fig. 9In contrast, for the comparative algorithm, the available resource update of the computing node leads to the change of its computing capacity, which indirectly leads to the prediction error greater than the proposed algorithm. In addition, the energy consumption formula reflects the instability of LRP when dealing with nonlinear variables.

[0065] Summary:

[0066] In this paper, cooperative communication and task offloading in VEC systems are studied to minimize response time and improve vehicle quality of service (QoS). The proposed TSODA includes two stages: first, a link quality model between vehicles is established to enhance transmission stability. The ENN scheme is used to predict the position of the vehicle. Second, the link quality, task size, and available resources of the computing node are input into the DNN model to predict the response time and make task offloading decisions. In the simulation results section, the accuracy of the two-stage prediction model is compared with the classic BP and AMRIA. It can be seen that the prediction results of the model have strong fitting characteristics with the actual values. In addition, compared with the prior art, TSODA can reduce time and energy consumption.

[0067] References

[0068] [1] S. Olariu. A Survey of Vehicular Cloud Research: Trends, Applications and Challenges, IEEE Trans. Intell. Transp. Syst. 21(6) (June 2020) 2648-2663, https: / / doi.org / 10.1109 / TITS.2019.2959743.

[0069] [2] A. Alamer, Y. Deng, G. Wei, X. Lin, Collaborative Security in Vehicular Cloud Computing: A Game Theoretic View, IEEE Netw. 32(3) (June 2018) 72-77, https: / / doi.org / 10.1109 / MNET.2018.1700329.

[0070] [3] K. Zhang, Y. Mao, S. Leng, A. Vinel, Y. Zhang, Delay constrained offloading for Mobile Edge Computing in cloud-enabled vehicular networks, in: 2016 8th International Workshop on Resilient Networks Design and Modeling (RNDM), 2016, pp. 288-294, https: / / doi.org / 10.1109 / RNDM.2016.7608300.

[0071] [4] K. Zhang, Y. Mao, S. Leng, Y. He, Y. ZHANG, Mobile-Edge Computing for Vehicular Networks: A Promising Network Paradigm with Predictive Off-Loading, IEEE Veh. Technol. Mag, 12(2) (June 2017) 36-44, https: / / doi.org / 10.1109 / MVT.2017.2668838.

[0072] [5] H. Li, T. Zhang, J. Zhang, K. Shi, P. Zeng. Time deterministic routing algorithm and protocol based on time-varying graph over the space-ground integrated network. Journal on Communications, 41(10) (2020) 116-129. https: / / doi.org / 10.11959 / j.issn.1000-436x.2020188.

[0073] [6] W. Jung, J. Yim, Y. Ko, S. Singh, Acods: adaptive computation offloading for drone surveillance system, in: In 2017 16th Annual Mediterranean Ad Hoc Networking Workshop (Med-Hoc-Net), 2017, pp. 1-6. https: / / doi.org / 10.1109 / MedHocNet.2017.8001647.

[0074] [7] J. Yuan, Y. Zheng, X. Xie, G. Sun. Driving with Knowledge from the Physical World, (2011) 316-324. https: / / doi.org / 10.1145 / 2020408.2020462.

[0075] [8] X. Hou, Y. Li, M. Chen, D. Wu, D. Jin, S. Chen, Vehcular Fog Computing: A Viewpoint of Vehicles as the Infrastructures, IEEE Trans. Veh. Technol, 65(6) (2016) 3860-3873, https: / / doi.org / 10.1109 / TVT.2016.2532863 .

[0076] [9] H. Zhang, Q. Zhang, X. Du, Toward Vehicle-Assisted Cloud Computing for Smartphones, IEEE Trans. Veh. Technol, 64(12) (2015) 5610-5618, https: / / doi.org / 10.1109 / TVT.2015.2480004 .

[0077]

[10] B. Karp, H. Kung, GPSR: Greedy perimeter stateless routing for wireless networks, in: Proceedings of the 6th annual international conference on Mobile computing and networking, 2000, pp. 243-254, https: / / doi.org / 10.1145 / 345910.345953.

[0078]

[11] R. Valentino, W. S. Jung, Y. B. Ko, A design and simulation of the opportunistic computation offloading with learning-based prediction for unmanned aerial vehicle (uav) clustering networks. Sensors, 18(11) (2018) 3751. https: / / doi.org / 10.3390 / s18113751.

Claims

1.A two-stage task cooperation offloading method in Internet of Vehicles, characterized in that, All vehicles are equipped with cellular interface and each vehicle is provided with sensors for sensing position, speed and direction, a task cooperative offloading framework is established which fuses LTE-V direct connection mode and LTE-V cellular mode, in which computing nodes are composed of mobile edge computing servers on roadside units and service vehicles, when in the LTE-V direct connection mode, the computing nodes are the service vehicles, when in the LTE-V cellular mode, the computing nodes are the mobile edge computing servers on roadside units, and the specific offloading method is: A link quality evaluation model for predicting the position of the vehicle in the next time slot and evaluating the link quality, and an offloading decision model for predicting the response time, in the data request stage, the request vehicle sends a HELLO information packet to the adjacent computing nodes, the link quality is obtained by the link quality evaluation model, when the task needs to be offloaded to the computing node, the offloading decision model obtains the computing node with the shortest response time according to the link quality, the amount of task input data, the maximum tolerable time, the available CPU resources of the computing node and the four prediction variables of wireless channel, the request vehicle sends the request information to the computing node with the shortest response time, when the computing node is the mobile edge computing server on the roadside unit, the mobile edge computing server on the roadside unit directly provides computing services for the request vehicle, when the computing node is the service vehicle, after obtaining the request information, the service vehicle judges whether it has enough computing resources, and feeds back the resource status to the request vehicle, whether to offload the task is determined by the request vehicle. 2.The two-stage task cooperation offloading method in the Internet of Vehicles according to claim 1, wherein, In the place where the turning or the obstacle blocks, the request vehicle is in the LTE-V cellular mode, the request vehicle sends the HELLO information packet to the roadside unit through the cellular network, and the mobile edge computing server on the roadside unit directly provides computing services for the request vehicle. 3.The two-stage task cooperation offloading method in the Internet of Vehicles according to claim 1, wherein, When the request vehicle is in the LTE-V direct connection mode, the information of the adjacent service vehicles is obtained in the alternative network topology scenario. 4.The two-stage task cooperation offloading method in the Internet of Vehicles according to claim 1, wherein, The content of the HELLO information packet includes broadcast ID, position information, requested task information and list information of adjacent computing nodes. 5.The two-stage task cooperation offloading method in the Internet of Vehicles according to claim 1, wherein, The link quality evaluation model is: wherein E ij represents the link quality, T s represents a fixed time interval, T u represents an update period, represents T s the proportion in the update period T u , p M represents the number of HELLO information packets in the time T s , R ij represents the distance between the requesting vehicle and the computing node where (a i ,b i ) represents the longitude and latitude of the current position of the request vehicle n i , (a i ',b i ) represents the longitude and latitude of the predicted position of the request vehicle n i , v i represents the speed of the request vehicle n i , v im represents the maximum speed allowed for the request vehicle n i , θ i represents the angle of the request vehicle n i , (a j ,b j ) represents the longitude and latitude of the current position of the computing node n j , (a' j ,b' j ) represents the longitude and latitude of the predicted position of the computing node n j , v j represents the speed of the computing node n j , θ j represents the angle of the computing node n j , and v jm represents the maximum speed allowed for the computing node n j . 6.The two-stage task cooperation offloading method in the Internet of Vehicles according to claim 1, wherein, The unloading decision model is a deep neural network composed of an input layer, a hidden layer and an output layer ij , task input data volume Maximum tolerable time Available CPU resources of the computing node And wireless channel b ij The task feature matrix is composed of The number of hidden layers is 10.

Images