Dynamic environment adaptive control method, device, equipment, medium and program product

Abstract

The application discloses a dynamic environment self-adaptive control method, device, equipment, medium and program product, and relates to the field of edge computing. The method comprises the following steps: determining a prediction result inference mode of the next behavior according to environment perception data in a shared cache and running state data of a to-be-controlled device. If the behavior prediction stops updating, a behavior prediction result stop updating identifier parameter is cached in the shared cache as the behavior prediction result. If the behavior prediction mode is adopted, different data elements of historical behavior data and behavior prediction data read from the shared cache are determined, and the final behavior prediction result is determined according to the behavior difference data and the historical behavior data. The application can solve the problems of poor real-time performance and large resource demand in related technologies, effectively reduce the response delay of behavior control of intelligent entity devices or virtual objects, and reduce the required computing resources and storage resources in the control process.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of edge computing, and in particular to a dynamic environment adaptive control method and device, electronic equipment, nonvolatile storage medium and computer program product. BACKGROUND

[0002] Dynamic environment adaptive control refers to adjusting the behavior of a controlled device in real time according to real-time and unpredictable changes in the physical world or virtual space environment. In the related art, the data communication efficiency between modules is low during the environment adaptive control of the controlled device, repeated calculation and frequent I / O (Input / Output) are frequent, real-time performance is poor and resource demand is large, which cannot meet the deployment requirements of edge scenarios. SUMMARY

[0003] The present application provides a dynamic environment adaptive control method and device, electronic equipment, nonvolatile storage medium and computer program product, which effectively reduces the response delay of behavior control of intelligent entity devices or virtual objects, reduces the required computing resources and storage resources of the control process, and can meet the deployment requirements of edge scenarios of users.

[0004] To solve the above technical problems, the present application provides the following technical solutions:

[0005] In one aspect, the present application provides a dynamic environment adaptive control method, comprising:

[0006] According to the environment perception data and the running state data of the to-be-controlled device within a preset time period, a prediction result inference method of the to-be-controlled device is determined; wherein the environment perception data is cached in a shared cache constructed in advance after being perceived.

[0007] If the prediction result inference method is a behavior prediction stop updating method, the behavior prediction result is cached in the shared cache as a behavior prediction result stop updating identification parameter.

[0008] If the prediction result inference method is a behavior prediction method, the historical behavior data of the last frame of the to-be-controlled device is read from the shared cache, and if the historical behavior data and the behavior prediction data of the current data frame are different in part, the behavior prediction result corresponding to the current data frame is determined according to the behavior difference data and the historical behavior data.

[0009] In another aspect, the present application provides a dynamic environment adaptive control device, comprising:

[0010] The prediction updating mode determination module is configured to determine a prediction result inference mode of the to-be-controlled device according to environment perception data and operation state data of the to-be-controlled device in a preset time period, wherein the environment perception data is cached in a shared cache after being perceived.

[0011] The stop prediction module is configured to cache a behavior prediction result stop update identification parameter as the behavior prediction result into the shared cache if the prediction result inference mode is the behavior prediction stop update mode.

[0012] The local prediction updating module is configured to read historical behavior data of a previous frame of the to-be-controlled device from the shared cache if the prediction result inference mode is the behavior prediction mode, and determine a behavior prediction result corresponding to a current data frame according to behavior difference data and the historical behavior data if the historical behavior data is different from part of data elements of the behavior prediction data of the current data frame.

[0013] The application further provides an electronic device comprising a memory and a processor, wherein the processor is configured to implement the steps of any of the above dynamic environment adaptive control methods when executing a computer program stored in the memory.

[0014] The application further provides a non-volatile storage medium, wherein a computer program is stored on the non-volatile storage medium, and the computer program is configured to implement the steps of any of the above dynamic environment adaptive control methods when executed by a processor.

[0015] Finally, the application further provides a computer program product comprising a computer program / instruction, wherein the computer program / instruction is configured to implement the steps of any of the above dynamic environment adaptive control methods when executed by a processor.

[0016] The technical scheme provided by the application has the advantages that the environmental perception data and the historical behavior data are stored in the shared cache, the perception module and the prediction module efficiently reuse the environmental data and the historical data, the same data does not need to be repeatedly loaded and calculated, the overhead of repeated I / O calls and inference calculation is significantly reduced, redundant calculation is reduced, the overall data channel efficiency is improved, the real-time and low-delay requirements of the edge application scenario can be met, the prediction data cache mode and the prediction data generation mode of the to-be-controlled device in the next moment are determined through dynamic environmental transformation and the state of the to-be-controlled device itself, the output data consistency can be maintained in the task pause or controlled state through the mode of stopping updating the prediction behavior and locally updating the prediction behavior, unnecessary behavior recalculation caused by slight perception fluctuations is avoided, the dynamic control update granularity and frequency are realized, redundant trajectory recalculation is avoided, high-frequency full-data rewriting is avoided, not only the I / O bandwidth pressure of the edge device is reduced and the service life of the storage medium is prolonged, but also the required computing resources and computing pressure for adaptive control of the to-be-controlled device are effectively reduced, the waste of computing power resources is effectively avoided, the system robustness is significantly improved, the low-delay and low-resource requirement scenario can be met, and the method is suitable for the resource-constrained and low-delay edge computing scenario.

[0017] In addition, the application also provides a corresponding implementation device, electronic equipment, non-volatile storage medium and computer program product for the dynamic environment adaptive control method, which further makes the method more practical, and the device, electronic equipment, non-volatile storage medium and computer program product have corresponding advantages. BRIEF DESCRIPTION OF DRAWINGS

[0018] In order to more clearly illustrate the technical scheme of the application or the related art, the following will briefly introduce the drawings needed to be used in the embodiment or the related art description. Obviously, the drawings in the following description are only some embodiments of the application, and those skilled in the art can obtain other drawings according to these drawings without creating any creative labor.

[0019] Figure 1 The hardware composition framework schematic diagram applicable to the dynamic environment adaptive control method provided by the application is shown in the figure.

[0020] Figure 2 The flowchart of the dynamic environment adaptive control method provided by the application is shown in the figure.

[0021] Figure 3 The framework schematic diagram of the dynamic environment adaptive control method provided by the application in an exemplary application scenario is shown in the figure.

[0022] Figure 4 The flowchart of another dynamic environment adaptive control method provided by the application is shown in the figure.

[0023] Figure 5 A structural framework diagram of an exemplary embodiment of the dynamic environment self-adaptive control device provided by the present application is shown in the figure.

[0024] Figure 6 A structural diagram of an exemplary embodiment of the electronic device provided by the present application is shown in the figure. DETAILED DESCRIPTION

[0025] In order to make the person skilled in the art better understand the technical solutions of the present application, the present application will be further described in detail below in combination with the drawings and specific embodiments. Among them, the terms "first", "second", "third", "fourth" and the like in the specification and the above-mentioned drawings are used to distinguish different objects, and are not used to describe a specific order. In addition, the terms "include" and "have" and any variations of the two are intended to cover non-exclusive inclusion. The term "exemplary" means "as an example, embodiment or illustrative". Any embodiment described as "exemplary" here is not necessarily interpreted as superior or better than other embodiments.

[0026] With the wide application of edge computing in the fields of industrial internet, intelligent manufacturing, smart logistics, intelligent transportation, etc., unmanned aerial vehicles, logistics fleets, mobile robots and other edge intelligent agents will perform task path reasoning and dynamic cache management on resource-constrained edge computing nodes, realize edge-end collaborative decision-making, and users' demand for multi-task collaboration, efficient data flow and low-latency response is also increasing. In actual deployment, edge computing nodes need to perform high-frequency data generation tasks, environment perception tasks and real-time decision-making tasks. However, due to the constraints of computing resources, network bandwidth and energy consumption of edge computing nodes, in order to improve the response ability and business operation stability of the edge computing system, efficient data flow and efficient cache update management between modules need to be realized.

[0027] In trajectory planning and control intensive scenarios such as autonomous driving, UAV path control, mobile robot scheduling, intelligent transport fleet, etc., the perception module, prediction module and decision module need to exchange historical trajectory data and environmental state data at a high frequency. The current edge computing system lacks perception and response to state stability and task dynamics, adopts a high-frequency whole package caching strategy, leading to system load soaring, cache repeated writing explosion and a large amount of repeated calculation, especially on complex intersections, high-frequency trajectory iteration and resource-constrained edge nodes. The communication between modules is inefficient, trajectory calculation is redundant, and the overall response is lagging. Each functional module lacks a unified data semantic interface and can only pull data and do repeated calculations, resulting in uncontrolled energy consumption and time delay. Ultimately, due to the use of full-quantity prediction and full-quantity refreshing in trajectory prediction and trajectory caching in related technologies, there is a lack of perception of low-dynamic scenes such as red light stops and slow-moving congestion, and the characteristics of short-term invariance of trajectories cannot be used to save resources. This coarse-grained homogeneous update not only wastes computing power and storage, but also slows down the system and reduces stability.

[0028] Currently, in the process of behavior control of the controlled device, multiple functional modules in the edge computing node, such as the perception module, the prediction module, and the control module, need to exchange a large amount of structured data in real time, such as the state of the controlled device or the environmental target, historical data, and intentions. However, due to the lack of a unified shared caching mechanism, each module often needs to load and calculate the same data repeatedly, resulting in redundant reasoning and frequent I / O calls, low communication efficiency of structured data across modules, and significant reduction in system computing efficiency and response speed. Especially in edge deployment environments, system resources are limited, and this redundancy is even more severe, directly restricting the scalability and robustness of the system. In addition, related technologies update behavior prediction data with fixed periods and single thresholds. In stationary or waiting states, trajectory / path prediction and cache updates continue, lacking a state stability identification and freezing mechanism. In low-speed running or slowly changing scenarios, the difference between trajectory data frames is small, but the high-frequency update rhythm is maintained, and frames with no significant changes are not skipped, causing repeated calculation and energy waste. Due to the inability to flexibly adjust the cache update granularity and frequency according to the system running state, on-demand updates cannot be achieved, resulting in unnecessary cache rewriting and trajectory recalculation, increasing the pressure on central processing units / graphics processing units, and leading to system resource waste and reduced running efficiency. Therefore, the edge computing system of related technologies is prone to system resource shortages and increased latency during behavior control of the controlled device, affecting the stability and reliability of the decision link, and failing to meet the requirements of edge computing resources, storage resources and real-time performance.

[0029] For example, one related technology is applied to the process of solving the computing offloading and resource allocation of integrated space-ground network, and jointly optimizes computing and communication resources under bandwidth, energy consumption, and information timeliness constraints through federated learning and multi-agent decision-making in a multi-UAV cloud-edge-end collaborative architecture. This method can improve the overall communication scheduling capability for global optimization of airspace trajectories, but does not involve fine-grained control and dynamic update strategies for local trajectory caching at edge nodes, making it difficult to meet the needs of low-latency prediction and local update in rapidly changing scenarios. Another related technology optimizes data transmission paths in a vehicle-road-cloud collaborative scenario, uses a spatiotemporal correlation model and cooperative transmission mechanism to optimize communication links, and introduces a feedback confirmation mechanism to ensure transmission robustness. Although it has certain adaptive ability of transmission path, it is limited to data link optimization between cloud-road-vehicle nodes, and does not involve caching strategy scheduling and efficient information multiplexing mechanism among modules at the edge, resulting in insufficient low-redundancy operation capability at the edge. Another related technology combines traffic light signals and vehicle dynamics state, designs time-varying speed constraints, and uses model predictive control to achieve path tracking. Although this method has closed-loop control capability, its modeling process relies on fixed prior rules and lacks flexibility in adapting to changes in environmental dynamics and driving intent; there is a lack of linkage mechanism between trajectory planning and perception results, which can easily result in response distortion in complex scenarios. Another related technology is based on historical trajectories and high-precision maps to establish an intersection path guidance mechanism, and enhances stability through trajectory pairing. This method emphasizes structured path modeling, but does not introduce traffic light state as a constraint, lacks decision-making strategies such as passage opportunity judgment and red light waiting, resulting in response defects at intersections. Another related technology involves trajectory smoothing and traffic control optimization under edge computing architecture, suitable for road side units or macro scheduling levels. Although it can improve the overall system scheduling efficiency, it does not establish a linkage mechanism between light state perception and trajectory reasoning frequency between perception modules and trajectory modules, still having problems of system resource waste and response lag.

[0030] Through comprehensive analysis of the above related technologies, it can be found that these related technologies have the following common deficiencies: relying on fixed prior conditions, lacking flexible adaptive mechanisms; lack of efficient information sharing interface among perception, prediction, and control modules, leading to repeated computation and resource waste; trajectory management strategies are mostly coarse-grained updates, which cannot achieve intelligent freezing or local trajectory replacement in low-frequency stable states such as red light waiting, congestion slow driving, and stationary operation; system module design is fragmented, making it difficult to adaptively select caching update mechanisms according to scene context; using full-frequency prediction and full-quantity caching update mode, lacking the ability to perceive and respond to scene state, resulting in waste of computing and storage resources and reducing the running efficiency and stability in edge computing environment.

[0031] In view of this, in order to solve the problem that the related technical perception, prediction, control and other functional modules usually lack a structured data sharing mechanism, each module needs to repeatedly load similar data and perform redundant calculations, resulting in high overall system load and high energy consumption, which is difficult to meet the requirements of high-frequency trajectory update and low-latency response in complex traffic environments, the present application supports cross-module information sharing through shared caching, thereby adapting to efficient data caching update in edge scenarios, improving the reuse efficiency of key data such as historical state, task intent, and behavior sequence between different modules, and significantly reducing redundant calculation and communication overhead. In combination with the running state of the to-be-controlled device and the environmental perception state, a fine-grained caching update strategy with state perception capability is determined, including at least a predicted behavior stop update mode and a local predicted behavior update mode, which has cross-module collaboration, lightweight scheduling and scene perception capabilities, dynamically controls the update granularity and frequency according to the scene state, can intelligently adjust the update frequency and content during stable or slowly changing state, minimizes redundant processing, improves system operation efficiency and resource utilization, and improves the computing efficiency, response real-time and running stability of the edge system.

[0032] In combination with the specific application environment architecture or specific hardware architecture on which the dynamic environment adaptive control method depends, the specific application environment architecture or specific hardware architecture is described herein. The following describes the technical solutions of the present application in combination with Figure 1 Some possible application scenarios related to the technical solutions of the present application are exemplarily introduced, which can include the following content:

[0033] The edge computing platform includes multiple edge computing nodes supporting the provision of multiple types of artificial intelligence tasks, and has low latency and high environmental adaptability. Each edge computing node can be a server or a small cluster composed of a group of servers, such as a server on a rack as an edge computing node. Each edge computing node is divided according to the type of edge computing service provided, and at least includes a trajectory control node of an intelligent driving vehicle, a robot behavior control node, an augmented reality user navigation node, a logistics control node, an industrial automation node and a traffic control node.

[0034] The edge computing node comprises at least a plurality of function modules, and the function modules cooperatively realize environment perception and control the behavior of the agent according to real-time changes of the environment. The function modules can include a perception module, a prediction module, a dynamic environment adaptive control module and a shared cache. The agent runs on the edge computing node, which is an autonomous or semi-autonomous entity with autonomous perception, calculation, communication and execution capabilities, and can complete local task decision, cache scheduling, trajectory planning, behavior prediction and other operations, such as an autonomous vehicle, a drone, a warehouse robot and a smart mobile device. The perception module is used for detecting the environment in which the agent is located and the target of interest in the environment and mapping the environment. After completing the environment and multi-target identification in the environment, the relevant environment information such as environment perception data and the running state data of the agent is written into the shared cache in real time, and the structured state information is written into the shared cache in real time. The prediction module reads the existing historical data information in the shared cache and the state of the agent itself, combines the current perception result, and of course if other entity interaction is involved, the corresponding intention data also needs to be obtained, to generate the behavior prediction data of the agent at the next moment. The dynamic environment adaptive control module determines the prediction result inference mode of the next frame behavior of the agent according to the environment perception data and the running state data of the agent in a preset time period. If the prediction result inference mode is the behavior prediction stop updating mode, the behavior prediction result is cached in the shared cache as a behavior prediction result stop updating identification parameter. If the prediction result inference mode is the behavior prediction mode, the historical behavior data of the last frame of the to-be-controlled device is read from the shared cache. If the historical behavior data and the behavior prediction data of the current data frame are different in part, the behavior prediction result corresponding to the current data frame is determined according to the behavior difference data and the historical behavior data. The shared cache can read and write data for the perception module, the prediction module and the dynamic environment adaptive control module, and provides high consistency, low delay state synchronization capability for each module as a cross-module data hub.

[0035] Each edge computing node can adopt a modular architecture design, and through the cycle of perception-prediction-decision-execution, the environmental uncertainty is converted into a deterministic optimization problem, and real-time conversion from data-policy-action is realized. When providing edge services to the outside, the shared cache mechanism and the behavior prediction and update strategy of the agent can be decoupled and deployed in the corresponding local controller, such as a vehicle terminal of an autonomous vehicle, a drone platform, a robot controller and an industrial collaboration device, to form a lightweight data management system with state perception capability, calculation scheduling capability and resource adaptive capability.

[0036] In an example scenario, when receiving the robot behavior control service, the robot behavior control node decouples and deploys the above-mentioned functional modules in the controller of the robot. Through the shared cache, the data reading delay can be reduced, the task execution efficiency is improved by caching the task state and the environment information, and the target tracking, path adjustment and real-time decision-making in a dynamic environment have obvious advantages. Through the dynamic environment adaptive control module, the resource utilization can be effectively optimized, the redundant calculation is reduced, the robot can more stably and efficiently complete the task in a complex and variable environment, and thus the response speed and the stability of task execution of the robot system can be significantly improved.

[0037] In an example scenario, when receiving the intelligent driving trajectory control service, the trajectory control node of the intelligent driving vehicle decouples and deploys the above-mentioned functional modules in the vehicle terminal of the intelligent driving vehicle. The perception module perceives the traffic light state, road condition data and driving state of surrounding vehicles. The dynamic environment adaptive control module can accurately judge whether to perform trajectory update in the current frame by perceiving the traffic light state, environment congestion and vehicle operation trend, and realizes trajectory prediction stop or local update through the state sequence, effectively avoiding invalid calculation. Combined with the shared cache mechanism, the structured traffic information and historical trajectory can be reused between the perception module and the prediction module, greatly improving the data interaction efficiency, and being particularly suitable for various typical automatic driving scenarios such as urban intersections, highway cruising and intelligent parking.

[0038] In an example scenario, when receiving the AR (Augmented Reality, augmented reality) user navigation task, the augmented reality user navigation node decouples and deploys the above-mentioned functional modules in the controller of the AR device. In the application scenario of high-precision identification of small targets (such as navigation signs, direction arrows, etc.), the shared cache mechanism can greatly reduce the data reading delay and improve the real-time responsiveness of the task. Through the combination of SLAM (Simultaneous Localization and Mapping, simultaneous localization and mapping) or visual input, through the environmental perception of the user interface prompt or navigation icon, the user's navigation experience is greatly enhanced. In addition, the user navigation trajectory supports behavior prediction stop update mode, which can optimize system resources by reducing unnecessary calculation, ensure the stability and efficiency of the AR system when processing complex data, and thus improve the real-time navigation experience of the user.

[0039] In an example scenario, when receiving a warehouse automation task or a cargo tracking task or an intelligent transportation control task, the logistics control node deploys the above-mentioned functional modules in the logistics / supply chain automation control system in a decoupled manner, which can greatly improve the response speed and scheduling efficiency of the system. The shared cache mechanism enables real-time data (such as cargo location, transportation path, inventory status, etc.) to be quickly shared and updated, thereby optimizing inventory management, route planning, and resource scheduling. The dynamic environment adaptive control module can dynamically adjust the path prediction frequency according to the transportation state (such as high-speed transportation, congestion, parking, etc.), reducing unnecessary computation and resource waste. In dynamic transportation and cargo tracking, by supporting behavior prediction stop update mode, unnecessary trajectory recalculation when stationary or low-speed driving can also be avoided, reducing the computational cost and improving the efficiency and real-time performance of the entire logistics system.

[0040] In an example scenario, when receiving an industrial automation control service such as an intelligent manufacturing task, a production line scheduling task, and a robot production line collaboration task, the industrial automation node deploys the above-mentioned functional modules in the industrial automation control system in a decoupled manner, which can optimize the device scheduling and production efficiency of the production line. Through the shared cache mechanism, it helps to improve the data flow efficiency between devices, reduce the delay and redundancy overhead caused by repeated calculation and data loading. In intelligent scheduling of production lines, the dynamic environment adaptive control module supports the prediction stop update mode and local behavior prediction mode, which can dynamically adjust the task scheduling frequency according to the working state of the device (such as standby, running, shutdown, etc.), avoid inefficient operation and resource waste, and significantly reduce the computational burden of the device control system in the scenario of device stationary or low-speed operation in the production process through the prediction stop update mode, optimizing energy use and device operation efficiency.

[0041] In an example scenario, when receiving an intelligent traffic management service, the traffic control node deploys the above-mentioned functional modules in the traffic signal control system in a decoupled manner, and through the shared cache mechanism, it shares traffic state, road condition information, and traffic light signals in real time, optimizes traffic signal control, traffic flow monitoring, and road network optimization, realizes dynamic control of intelligent traffic signals, reduces traffic congestion, and improves traffic efficiency. By supporting the prediction stop update mode, the response frequency and strategy of the traffic light can be adjusted according to the traffic volume and traffic conditions of different road sections, further improving the response speed of the traffic signal control system and the smoothness of traffic flow.

[0042] From the above, the edge computing platform of the embodiment decouples the shared cache mechanism and the intelligent trajectory updating strategy through modular design and deploys them in the local controller of the controlled intelligent agent. The edge computing platform can be flexibly deployed in the local control end of an automatic driving system, a warehouse robot, a UAV platform, and an industrial collaborative device, has strong edge deployment friendliness, flexibility, universality, and portability, is suitable for various edge perception and decision-making tasks, and is particularly suitable for embedding a lightweight edge platform for operation. Through the shared cache mechanism and various behavior prediction and updating strategies, the processing efficiency and real-time response capability can be improved in a multi-task environment, cache on-demand updating and resource intelligent scheduling are achieved, and the edge computing platform is particularly suitable for application scenarios that require efficient path planning, decision-making control, and resource scheduling. The edge computing platform not only optimizes system resource utilization, improves cache access and updating efficiency, and enhances the adaptability to changes in the running state, but also significantly reduces repeated calls of inference modules in the edge computing node and saves computing resources.

[0043] It should be noted that the above application scenarios are only shown for the purpose of facilitating understanding of the ideas and principles of the present application, and the embodiments of the present application are not limited in this respect. On the contrary, the embodiments of the present application can be applied to any applicable scenario. After introducing the technical solutions of the present application, various non-limiting embodiments of the present application will be described in detail in combination with the drawings and specific embodiments.

[0044] First, please refer to Figure 2 , Figure 2 A flowchart of a dynamic environment adaptive control method provided by the embodiment, which can include the following content:

[0045] S201: determining a prediction result inference manner of a to-be-controlled device according to environment perception data and running state data of the to-be-controlled device in a preset time period.

[0046] The to-be-controlled device is an entity device or a virtual object that needs to adaptively control its behavior or output according to environmental changes. The environmental changes refer to real-time and unpredictable changes in the physical world or virtual space environment, such as moving targets, road condition fluctuations, user displacement, inventory changes, and device state changes. The to-be-controlled device or the computer control system in which it is located has a computer program with autonomous perception and response capability. Accordingly, it can be an autonomous or semi-autonomous intelligent agent with perception, computation, communication, and execution capabilities. Control refers to applying specific and executable operation instructions to physical entity devices or virtual objects, such as adjusting robot joint torque / speed, generating vehicle steering / throttle instructions, updating AR display content, sending replenishment / vehicle dispatch instructions, adjusting production line tempo / robot action, and making the controlled object (robot, vehicle, AR interface, inventory system, production line) achieve and maintain the desired state or performance indicators such as accurate tracking, safe driving, accurate navigation, inventory balancing, and efficient production.

[0047] The environment perception data is obtained by the perception module of the agent to be controlled and is cached in a pre-constructed shared cache after being perceived, and the shared cache supports the data reading and writing or data sharing of each functional module of the agent. The environment perception data at least includes relevant data of the surrounding environment of the device to be controlled, which includes visual data of the target object required to be concerned in the surrounding environment and the current scene (two-dimensional image data providing rich color, texture, shape and context information, three-dimensional point cloud data and infrared radiation data providing accurate three-dimensional geometric information of the environment), distance and space data, auditory data, environmental physical data measuring physical and chemical properties of the environment, position and location data. Of course, according to the current application scenario, the relevant running state data and intention data of other surrounding agents in the surrounding environment of the device to be controlled, which affect the behavior of the device to be controlled, can also be included, and the running state data of the surrounding agents and the device to be controlled at least includes the attitude and motion data of itself, such as the acceleration measured by the accelerometer, the angular velocity measured by the gyroscope and the magnetic field direction measured by the magnetometer, the global geographic position (latitude and longitude) of the device to be controlled and the current accurate time, the distance of the agent movement and the current speed. The prediction result inference manner refers to determining the behavior prediction result to be output by the device to be controlled and the behavior prediction data to be cached in the shared cache according to the current input data frame (environment perception data (task rhythm (waiting / execution) and environment input (congestion / abnormality)) and running state data (stationary / creeping)), through the fusion data recognition result of context information such as speed, change rate and dynamic environment, adjusting the behavior prediction and updating strategy in real time, and improving the intelligent level of response. The prediction result inference manner at least includes the behavior prediction stop updating manner of S202 and the behavior prediction manner of S203, and the behavior representation data of the device to be controlled includes a plurality of data elements.

[0048] S202: If the prediction result inference manner is the behavior prediction stop updating manner, the behavior prediction result stop updating identification parameter is cached as the behavior prediction result in the shared cache.

[0049] The behavior prediction stop updating manner refers to stopping the behavior prediction operation on the current input data frame, that is, the prediction module of the agent stops predicting the behavior, and on the basis that the prediction module stops predicting the behavior, the current output behavior prediction result is a behavior prediction result stop updating identifier parameter, which is stored in the shared cache to be used by other modules of the agent. For example, in the signal control scene of traffic lights, job instruction lights and the like, the to-be-controlled device can be in a static state or a slow waiting state. In order to reduce the redundant prediction calculation overhead in such a state, the behavior prediction frequency can be reduced or the prediction data cache updating operation can be temporarily frozen on the premise of maintaining the prediction accuracy, so as to guarantee the state stability and improve the system efficiency. In the state that the to-be-controlled device is in a slow movement, local resource congestion and the like, the overall path trend is stable, and the change between the continuous frames of the trajectory is very small. In such a case, the behavior prediction, cache updating and intention vector encoding of the current frame can be skipped. The intention vector encoding can include the encoding of the next behavior intention of the to-be-controlled device, and can also include the encoding data of the behavior intention of other entity objects or virtual objects around the to-be-controlled device. The historical data of the last frame is reused, the redundant calculation is reduced, and the running efficiency of the edge node is improved. The behavior prediction result stop updating identifier parameter can be a stop updating time and a stop updating frame number. The stop updating time refers to the time corresponding to the behavior prediction result, which can be represented by a time stamp. The stop updating frame number refers to the statistical result of how many frames of input data frames do not perform the behavior prediction operation.

[0050] The controlled device is in a static state (such as a vehicle waiting for a red light, an industrial robot on standby, a mobile robot paused, a drone hovering), a low-speed moving state or a slowly changing state (such as a low-speed following state, a traffic jam state), a stable behavior (such as a stable trajectory), and the like. The behavior prediction results corresponding to adjacent data frames are similar or the same, so the behavior prediction operation on the current input data frame and the high-frequency update rhythm of the data frame with small differences can be stopped, thereby fully utilizing the behavior stability in the static or low dynamic state, reducing the invalid overhead in such static or quasi-static scenarios, thereby solving the problem of serious redundancy of computing resources and storage bandwidth caused by repeatedly performing behavior prediction and cache update operations at each frame in the controlled waiting or static state (such as a red light signal, a work pause, etc.). For example, the environmental signal state (such as a traffic light or a work indicator light), the speed change trend of the controlled device, the distance between the current position and the stop area, the dynamic density of the surrounding intelligent agents (area congestion factor), and other factors can be considered comprehensively to effectively avoid the controlled device performing redundant behavior reasoning and cache writing operations in the controlled waiting state, improve the resource utilization efficiency of the edge computing system, and reduce the problem of trajectory jitter caused by minor disturbances, thereby enhancing the stability and safety of the edge computing system operation. In addition, considering the overall trend of behavior prediction data, the behavior prediction stopping updating mode does not perform trajectory prediction and does not write to the cache, but only caches the behavior prediction result stopping updating identification parameter, instead of periodic forced updating, effectively reducing the redundant computing load and cache writing frequency, saving system resources. The trajectory is not repeatedly updated when the scene is unchanged for a short time, and the system operation efficiency is improved.

[0051] S203: If the prediction result inference mode is a behavior prediction mode, read the historical behavior data of the last frame of the controlled device from the shared cache, and if the historical behavior data and the behavior prediction data of the current data frame are different in part, determine the behavior prediction result corresponding to the current data frame according to the behavior difference data and the historical behavior data.

[0052] The behavior prediction mode is a mode in which behavior prediction needs to be performed, that is, a prediction module of the to-be-controlled device is in normal operation. The behavior prediction mode can include local prediction and full-quantity prediction. The full-quantity prediction is a prediction mode and a data updating mode in the related art, that is, an entire behavior prediction data replacement method is used. The local prediction is a mode in which, in a scenario in which overall behavior change is not significant but local details deviate, resource waste caused by full-quantity trajectory recalculation and a large amount of additional storage resources are avoided. When it is identified through the above steps that the behavior prediction stop updating mode does not apply, the behavior prediction mode is used. Further, whether local updating or full-quantity updating is performed is determined according to a difference degree between current behavior prediction data (that is, behavior prediction data of a current data frame) and behavior data corresponding to a previous input data frame (that is, historical behavior data of a previous frame). The behavior prediction data in this embodiment includes a plurality of data elements. Taking trajectory prediction data as an example, the data elements are trajectory points. Taking cargo transportation data as an example, the data elements are cargo positions. The amount of data of some data elements can be determined according to actual conditions. The behavior difference data is data that is different from the historical behavior data and the behavior prediction data of the current data frame, that is, data that deviates from the historical behavior data. After the behavior prediction result is determined, the behavior prediction result is also stored in the shared cache.

[0053] In the technical solution provided in this embodiment, the environment perception data and the historical behavior data are stored in the shared cache, efficient reuse of environment data and historical data is implemented for the perception module and the prediction module, repeated loading and repeated calculation of the same data are not needed, overheads of repeated I / O calls and inference calculation are significantly reduced, redundant calculation is reduced, the overall data channel efficiency is improved, real-time performance and low-delay requirements of edge application scenarios can be met. Through dynamic environment transformation and a state of the to-be-controlled device itself, a prediction data caching mode and a prediction data generation mode of the to-be-controlled device in a next moment are determined. Through support for the stop updating prediction behavior and the local updating prediction behavior, output data consistency can be maintained in a task suspension or controlled state, unnecessary behavior recalculation caused by slight perception fluctuations is avoided, dynamic control updating granularity and frequency are implemented, redundant trajectory recalculation is avoided, high-frequency full-quantity data rewriting is avoided, not only I / O bandwidth pressure of the edge device is reduced and a service life of a storage medium is prolonged, but also calculation resources and calculation pressure required for adaptive control of the to-be-controlled device are effectively reduced, waste of computing power resources is effectively avoided, system robustness is significantly improved, low-delay and low-resource scenarios can be met, and the solution is suitable for resource-constrained and low-delay edge computing scenarios.

[0054] In the above embodiments, how to determine the prediction result inference mode is not limited. An exemplary implementation of the prediction result inference mode in this embodiment can include the following content.

[0055] The target running state data at the target moment, the historical running state data and the historical environment perception data in a preset time period are acquired; if it is determined according to the target running state data that the action amplitude of the to-be-controlled device is less than a preset action amplitude threshold, it is determined according to the historical running state data that the action change amount of the to-be-controlled device in the preset time period meets a preset quasi-zero acceleration state condition, and it is determined according to the historical environment perception data that a plurality of target objects existing in the behavior front space of the to-be-controlled device meet a preset action maintenance condition, the prediction result inference mode is a behavior prediction stop updating mode; if the prediction result inference mode is not the behavior prediction stop updating mode, the prediction result inference mode is a behavior prediction mode.

[0056] In the embodiment, the target moment is a perception time corresponding to the current data frame, the target running state data is running state data of the to-be-controlled device corresponding to the target moment, the preset time period is a period of time before the target moment, the historical running state data is corresponding running state data in the preset time period, and the historical environment perception data is environment perception data in the preset time period. These data can be read from a shared cache. The preset action amplitude threshold can be determined according to actual conditions. The preset quasi-zero acceleration state condition is a static state or a quasi-static state. The number of target objects can be determined according to actual conditions. The preset action maintenance condition indicates that the target objects are in a static state or a slow behavior state.

[0057] For example, the behavior prediction stop updating mode can include a prediction data cache freezing mode and a frame skipping updating mode. In a signal control scenario such as a traffic light or an operation instruction light, the to-be-controlled device can be in a static state or a slow waiting state. In order to reduce the redundant prediction calculation overhead in such a state, the trajectory prediction frequency can be reduced by using the frame skipping updating mode, or the trajectory cache updating operation can be temporarily frozen to ensure state stability and improve system efficiency while maintaining prediction accuracy by using the prediction data cache freezing mode. The state judgment process is as follows: obtaining the environment indication signal data (such as a traffic signal light or an indication signal serving as an instruction light) in the environment perception data at the target time; obtaining the behavior stop area (such as a zebra crossing or an entry prohibited area or a non-operable area) in the environment perception data at the target time and the target spatial position of the to-be-controlled device at the target time, and calculating the forward travel distance (i.e., the distance between the to-be-controlled device and the behavior stop area) between the to-be-controlled device and the behavior stop area along the forward action direction; if the environment indication signal data remains unchanged within a preset future time period, and the forward travel distance is less than a preset distance threshold value set in advance according to the actual scene, the behavior prediction stop updating mode is the prediction data cache freezing mode; if the environment indication signal data changes within the preset future time period, or the forward travel distance is greater than or equal to the preset distance threshold value, and there is no continuous multiple frames using the behavior prediction stop updating mode before the current data frame, the behavior prediction stop updating mode is the frame skipping updating mode. Correspondingly, if the behavior prediction stop updating mode is the prediction data cache freezing mode, the generation of the next frame of behavior prediction data of the to-be-controlled device is inhibited, the behavior prediction state of the current data frame is set to a first identification value indicating the prediction data cache freezing mode, and the first identification value and the corresponding timestamp are taken as behavior prediction result stop updating identification parameters; if the behavior prediction stop updating mode is the frame skipping updating mode, the historical behavior data of the previous frame is taken as the behavior prediction data of the next frame of the to-be-controlled device, the behavior prediction state of the current data frame is set to a second identification value indicating the frame skipping updating mode, and the second identification value and the corresponding timestamp are taken as behavior prediction result stop updating identification parameters.

[0058] Further, the embodiment also ensures that the frozen or frame-skipping state has the ability to recover in time by removing the conditions for the prediction stop updating mode: if the new environment perception data or the new running state data does not meet the conditions corresponding to the behavior prediction stop updating mode, such as a change in the environment signal (such as a traffic permission recovery), a surrounding state disturbance (such as the appearance of a new obstacle), a significant displacement of the body position, etc., the behavior prediction state of the current data frame is set to a third identification value, and the third identification value indicates the resumption of normal behavior prediction operation and cache updating operation.

[0059] In the embodiment, the maximum number of frame skipping can be set in advance. By setting the maximum number of frame skipping, the continuous frame skipping is avoided to cause the slow dynamic response of the to-be-controlled device to the scene, and the safety risk is avoided. In the embodiment, when determining whether to perform behavior prediction, not only the geometric difference of adjacent data frames of the behavior is considered, but also rich context information such as the speed, speed change, and surrounding intelligent agent of the to-be-controlled device is fully considered, which is a fine-grained judgment.

[0060] For example, the satisfaction condition of the prediction data cache freezing mode can include: (1) an environmental indication signal such as a red traffic signal light, a prohibition instruction, and the like will maintain a prohibition of passing or a waiting state for a future time period such as 8 seconds or more: ; wherein, represents a time period in the future, represents a time in the time period, and the signal light state is a three-classification vector such as red, yellow, and green or permission, prohibition, and warning. (2) The current prediction behavior end point is close to a preset stop point, that is, the forward travel distance of the to-be-controlled device is less than a prediction distance threshold value (such as 3 meters): ; wherein, represents the current prediction behavior end point, represents the preset stop point, . (3) The action amplitude of the to-be-controlled device is small and is lower than an action amplitude threshold value , such as = 0.2 m / s, the speed of the to-be-controlled device is extremely slow, and can be represented as: . (4) The speed change rate of a plurality of time steps (such as 5 frames) in succession tends to 0 (that is, a preset quasi-zero acceleration state condition is satisfied), and the speed change rate is defined as: , which can be expressed as , is a minimum value. (5) The proportion of low-speed moving intelligent agents in the behavior forward space is high, such as the blocking of the behavior forward space or the high traffic density, which can be represented by , wherein, is a regional congestion factor, is a congestion determination threshold value, , i represents the i th target object in the behavior forward space, represents the action amplitude such as the behavior speed of the i th target object in the behavior forward space, is the number of target objects.

[0061] The satisfaction condition of the frame skipping update mode can include: (1) the action amplitude of the to-be-controlled device is small and is lower than an action amplitude threshold value , such as = 0.2 m / s, speed of the device to be controlled very slow, which can be expressed as: (2) the rate of change of speed in a plurality of consecutive time steps (e.g., 5 frames) tends to 0 (i.e., meets the preset quasi-zero acceleration state condition), the rate of change of speed is defined as: , which can be expressed as , is a minimum value. (3) the proportion of low-speed moving agents in the behavior forward space is high, such as congestion or high traffic density in the behavior forward space, which can be expressed by , wherein is a regional congestion factor, is a congestion determination threshold, , i represents the ith target object in the behavior forward space, represents the action amplitude (e.g., behavior speed) of the ith target object in the behavior forward space, is the number of target objects.

[0062] Further, the state sequence of the identification prediction data cache freezing mode, the frame skipping update mode, and the behavior prediction mode can be pre-set , the state sequence is a sequence composed of 0 (the third identification value), 1 (the first identification value), and 2 (the second identification value), and each value represents a frame. By default , the number of frames of behavior freezing and frame skipping update can be recorded to prevent the number of frames of continuous non-updated trajectories from being too large, thereby ensuring the safety of system operation, such as ensuring driving safety in the self-driving field. indicates that the behavior is to be predicted and the cache is to be updated; indicates that the prediction data cache freezing mode is triggered, the behavior prediction and the behavior prediction data cache update are suspended, and only the timestamp and of the cached behavior are updated; indicates that the frame skipping update mode is triggered, the behavior prediction and the behavior prediction result cache update are suspended, and only the timestamp and of the cached trajectory are updated. In the embodiment, when the joint conditions of the prediction data cache freezing mode are met, the prediction data cache freezing mode is triggered, the frame is set to the freezing state in the state sequence, that is, . When any of the following release conditions is detected, the prediction data cache freezing mode is automatically released, and normal prediction and update are resumed: the signal state is about to change (such as ending waiting and becoming passing) ( ); or the current predicted behavior is far away from the stopping area; a new surrounding agent behavior change or a sudden dynamic obstacle is detected, or a new task instruction or control signal is received, and then is set, and the behavior prediction is continued. When the joint conditions of the frame skipping update mode are met, the frame skipping update mode is triggered, The frame skipping behavior predicts and writes cache, multiplexes the historical behavior data and intention vector of the last frame, only updates the timestamp and cache index, and reduces the I / O and inference load. If it is detected that the data frame continuously updated by skipping frames exceeds the preset frame skipping threshold, indicating that the behavior difference exceeds the set error range, or does not meet one of the other conditions of the frame skipping update mode, the frame skipping update mode is immediately exited, and , the behavior prediction update is restored.

[0063] As can be seen from the above, in the embodiment, when the to-be-controlled device is in a static or controlled waiting state, behavior prediction and cache writing are suspended, only the state timestamp and stop frame number statistics are retained, and redundant calculation is reduced. When in a slow moving, local resource congestion or other slowly changing state, the overall behavior trend is stable or the behavior changes little between consecutive frames, the behavior prediction, cache update and intention vector encoding of the current frame can be skipped, the behavior prediction data of the last frame is reused, redundant calculation is reduced, and the running efficiency of the edge computing node is improved.

[0064] To reduce the redundant prediction calculation overhead of the to-be-controlled device that may be in a static state or a slow waiting state, the system can reduce the trajectory prediction frequency, or temporarily freeze the trajectory cache update operation while maintaining the prediction accuracy, to ensure the stability of the state and improve the system efficiency. The embodiment also avoids the waste of resources caused by full behavior recalculation in the scenario where the overall behavior of the to-be-controlled device changes little but the local details deviate, and supports four fine-grained update strategies for the prediction result inference mode: prediction data cache freezing mode, frame skipping update mode, local behavior data update mode and full behavior data update mode. According to the current system state (such as device running speed, target point proximity, local task congestion, etc.), the corresponding prediction result inference mode is dynamically selected to improve the behavior update efficiency and edge computing resource utilization. The identification method of the local behavior data update mode and the full behavior data update mode can include the following contents:

[0065] Calculate the difference between each data element of the historical behavior data and the data element corresponding to the behavior prediction data of the current data frame. If the total number of target data elements greater than or equal to the preset difference threshold is less than or equal to the preset local deviation threshold, the behavior prediction mode is the local behavior data update mode, the behavior prediction result corresponding to the current data frame is determined according to the value of each target data element in the behavior prediction data of the current data frame and the historical behavior data, and stored in the shared cache. If the total number of target data elements greater than or equal to the preset difference threshold is less than the preset local deviation threshold, the behavior prediction mode is the full behavior data update mode, the behavior prediction data of the current data frame is taken as the behavior prediction result corresponding to the current data frame, and stored in the shared cache.

[0066] In the embodiment, after the judgment of the prediction data cache freezing mode and the frame skipping update mode, if the output is 0, the behavior prediction model is entered, and the difference between the current frame and the previous frame is compared to judge the degree of change of the behavior, if the overall behavior change is not large, but the target data element (such as the previous 2-3 track points) offset exceeds the preset difference threshold, it is determined as the local behavior data update mode, otherwise, the full behavior data update mode is used.

[0067] As can be seen from the above, the embodiment adopts fine-grained behavior prediction and cache update, and comprehensively considers the driving state of the to-be-controlled device, environmental signals, and the degree of change of the predicted behavior, and dynamically selects whether to use the prediction data cache freezing mode, the frame skipping update mode, the local behavior data update mode, or the full behavior data update mode, effectively alleviates the redundant calculation pressure of the behavior prediction data cache, and at the same time guarantees the control stability and response efficiency of the to-be-controlled device. Further, by setting the state sequence auxiliary record update strategy activation state, the number and frequency of track freezing and frame skipping update are counted to assist in determining whether to adopt freezing and frame skipping cache optimization strategies in the future, and the real-time performance and safety of the edge computing system are improved.

[0068] Considering that the automatic driving vehicle, unmanned aerial vehicle, unmanned transportation device and other edge intelligent agents often need to make real-time judgments according to the state of the target signal (such as traffic signal, operation instruction light) in the surrounding environment when performing path prediction or task scheduling, in order to guarantee the safety and traffic efficiency in the complex dynamic environment, in order to further improve the efficient communication between the perception module and the prediction module, effectively reduce redundant calculation, avoid repeated I / O calls, and improve the overall calculation efficiency, meet the real-time performance requirements, the application also provides a cross-module structured information sharing cache implementation mode, by designing a structured, concurrent access supported, and history information sliding window supported sharing cache mode, the real-time performance and calculation resource utilization of the system as a whole can be significantly improved, the sharing cache of the embodiment can include two regions, which respectively store structured environmental state information (environmental state information region) and historical behavior data (historical behavior information region), for supporting data reuse between the perception module and the prediction module, and can include the following contents:

[0069] ​Exemplarily, whenever an environment perception data frame is generated, spatial position data and / or pose data and / or state data of each environment target is extracted from the environment perception data frame as environment target perception data of each environment target; a data frame number of the environment perception data frame is taken as an index, and environment cache structured data corresponding to the environment perception data frame is generated according to a timestamp of the environment perception data frame and the environment target perception data; the timestamp of the environment perception data frame is taken as a key, the environment cache structured data is taken as a value, and the environment cache structured data is written into an environment state information region of the shared cache as a cache entry in a key-value pair data structure manner, so as to read each environment target perception data from the shared cache through a read request command containing the key. Further, on the basis of the key-value structure storage, type constraint information can also be added, data types corresponding to state data of the environment cache structured data are obtained as first type description information, data types of the timestamp, spatial position data and pose data of the environment perception data frame are obtained as second type description information, the first type description information and the second type description information are taken as field type constraint description information, and the environment cache structured data, the first type description information and the second type description information are written into the environment state information region of the shared cache as a cache entry in a key-value pair data structure manner and a data cache format of the field type constraint description information. The state and spatial position information of the environment control signals such as traffic lights and work instruction lights are stored in the shared cache in a structured form, which is used to provide continuous and stable state input for a behavior prediction or behavior decision module of a to-be-controlled device, and supports multi-frame data multiplexing and decoupling access between modules.

[0070] For example, when generating the historical behavior data of the to-be-controlled device according to the last frame of input data, the data frame number of the last frame of input data is taken as an index, and the behavior prediction structured data of the last frame of input data is generated according to the historical behavior data, the time stamp of the last frame of input data, and the device identification information of the to-be-controlled device; the time stamp of the last frame of input data is taken as a key, the behavior prediction structured data is taken as a value, the behavior prediction structured data is written into the historical behavior information area of the shared cache as a cache entry in a key-value pair data structure mode, and each frame of historical behavior data is read from the shared cache through a read request command containing the key. In the scenario where the intention needs to be considered, the cross-module collaborative call between the perception module and the prediction module is facilitated, the consistency of the target detection of the perception module and the intention of the to-be-controlled device is improved, the historical behavior information area further includes the intention encoding result in addition to the historical behavior data, the historical intention encoding data corresponding to the last frame of input data can be obtained, the behavior prediction structured data of the last frame of input data is generated according to the historical trajectory data and the corresponding historical intention encoding data, the time stamp of the last frame of input data, and the device identification information of the to-be-controlled device. Further, on the basis of the key-value structure storage, type constraint information can be further added, the data types of the time stamp of the behavior prediction structured data, the position data of each trajectory point of the historical trajectory data, the historical intention encoding data, and the device identification information are obtained as field type constraint description information; the behavior prediction structured data and each data type are written into the historical behavior information area of the shared cache as a cache entry in the data cache format of the key-value pair data structure mode and the field type constraint description information.

[0071] In this embodiment, multiple functional modules in an edge computing system (such as a vehicle-mounted computing system) need to read structured environment information and historical information multiple times, the structured environment information is used for trajectory and behavior prediction of a prediction module, and the historical information is used for trajectory and behavior prediction and also helps to guide target detection. For repeated data reading between the perception module and the prediction module, this embodiment realizes efficient communication, information reuse, and delay optimization among multiple modules through structured data modeling and module decoupling access interfaces. The spatial positions, state categories (such as red light, green light, work allowed, prohibited, etc.), remaining time lengths, and other attribute information of multiple targets can be stored by taking the time stamp as an index, to be read by the prediction module in real time and used for generation of a behavior prediction result or strategy switching judgment. Meanwhile, the cache of historical information and intention vectors can provide continuous behavior context for a to-be-controlled device, support more smooth and stable generation of a prediction result, and enable collaborative scheduling of environment perception and behavior reasoning on a resource-limited edge computing platform, thereby improving overall system operation efficiency and safety robustness of task decision-making.

[0072] Taking the cache of environment indication signal (traffic light, indicator light) information as an example, each frame of cached data takes the current frame number For index, the structure is defined as follows: , . Wherein, is dedicated to caching traffic / traffic light structured information from the target detection head, is the timestamp of the frame, is the traffic / traffic light set, and the structured information of the ith traffic / traffic light includes its two-dimensional spatial position , orientation angle , state label , and height . Taking trajectory prediction as an example of behavior prediction results, the structure of each frame prediction result can be represented as: ; wherein, is used to cache the current trajectory information, historical trajectory information and intention encoding result of the controlled device (autonomous vehicle, unmanned aerial vehicle, unmanned vehicle), includes the controlled device identifier, trajectory point sequence , the ith trajectory point , r is the total number of trajectory points, represents the high-dimensional intention encoding vector data output by the intention encoding module, also bound with timestamp for alignment with the perception time, d is the dimension of the vector, is the timestamp of the frame.

[0073] To realize decoupling access between modules, the cache interface adopts a unified read_cache (read cache) (key (key)) / write_cache (write cache) (key (key), value (value)) abstraction. Exemplarily, the Key-Value pair data structure can store key-value pairs of data using a dictionary data structure, and the value in the dictionary can be read through read_cache (key) and new key-value pairs can be written through write_cache (key, value), thereby avoiding direct dependence between the perception module, the prediction module and the rule module. All cache entries use key-value structure + type constraint description, such as dict+enum+float, to avoid chained objects or cross-reference structures and ensure serialization and cross-platform migration. Among them, dict: the cache entry uses a dictionary structure, which is convenient for extension and indexing; enum: the state field uses an enumeration type (such as red light = 0, green light = 1, yellow light = 2); float: the trajectory point position, speed, timestamp, etc. are stored using floating-point numbers to ensure calculation accuracy.

[0074] Further, the embodiment also supports sliding window access to historical behavior data: the shared cache maintains the historical behavior information region in a circular array manner, when receiving a historical data read instruction containing a target timestamp, the data read start timestamp is determined according to the maximum window length and the target timestamp; from the shared cache, each target historical behavior data corresponding to the data read start timestamp to the target timestamp is read.

[0075] In the embodiment, the behavior data is taken as trajectory data for example, and the behavior prediction data is taken as trajectory prediction sequence, in order to ensure the consistency of the trajectory data in the near future Frame trajectory real-time query, the cache is maintained in a circular array manner, and all accesses are completed through unified interfaces read_cache() and write_cache(). Sliding window access is supported, and the maximum window length is set to Frame, indicating that the trajectory prediction result of m time steps at most is cached, each time step corresponds to a trajectory sequence, and a trajectory sequence contains r future trajectory points, that is, sequence is supported. Wherein, represents the trajectory point sequence of the cache up to . represents the trajectory prediction sequence of the time step , represents the timestamp corresponding to the current time step frame, represents the frame number index.

[0076] As can be seen from the above, the embodiment stores the structured information of the environmental data (such as traffic signal lights, work instruction lights and other static targets) and the historical behavior information cache, the historical behavior information region stores the historical behavior point sequence of the to-be-controlled device at multiple time steps, and also contains the high-dimensional intention vector corresponding thereto, which provides the continuous semantic expression of behavior generation for the prediction module. On the other hand, the environmental semantic information such as environmental signals is also stored in the shared cache in a structured form, which provides a stable input source for behavior prediction and behavior judgment. Further, by supporting sliding window management and timestamp alignment, the consistency of multiple frames of information in different modules is ensured, and cross-module bidirectional collaboration of target detection and behavior prediction is realized. Further, unlike the processing mode of “module independent running, local cache or repeated call of perception result” in the edge computing system, the embodiment realizes the structured information sharing and state collaboration between the perception module and the prediction module through the unified data structure and access interface, which significantly improves the overall running efficiency and consistency of the system, and provides efficient data support for intention guidance, target selection, behavior prediction and other system tasks.

[0077] Further, in order to optimize the accuracy of target detection of the perception module and behavior prediction of the behavior prediction module, after determining the behavior prediction result corresponding to the current data frame according to the behavior difference data and the historical behavior data, a corresponding intention encoding result can also be generated. For the local behavior data updating mode, the corresponding intention encoding result generation process can include: reading the historical intention encoding data of the last frame that is semantically associated with the behavior prediction result from the shared cache; encoding the target data elements that are different from the data elements of the historical behavior data in the behavior prediction data of the current data frame to generate local intention encoding data; and splicing the local intention encoding data to the historical intention encoding data to generate the intention encoding result corresponding to the current data frame.

[0078] In the embodiment, after the local behavior data updating mode is executed, the part of the behavior sequence corresponding to the changed behavior segment is re-encoded, the updated local intention vector is generated, and the updated local intention vector is spliced into the original intention encoding sequence. On the basis of ensuring the accuracy and real-time performance of intention vector encoding, the resource waste caused by full-amount intention vector re-encoding is avoided.

[0079] Further, the application also provides a more accurate and simple implementation mode for the identification mode of the local behavior data updating mode and the full-amount behavior data updating mode, which can include the following contents:

[0080] The data elements of the historical behavior data and the behavior prediction data of the current data frame are taken as data points, and the point-to-point difference values of the historical behavior data and the behavior prediction data of the current data frame are used to generate point-to-point error information; the point-to-point difference values in the point-to-point error information are sequentially compared with a preset difference value threshold, and the point-to-point difference values greater than or equal to the preset difference value threshold are assigned a first mask value (such as 0), and the point-to-point difference values less than the preset difference value threshold are assigned a second mask value (such as 1), so as to convert the point-to-point error information into a behavior difference mask; when the total number of the first mask values in the behavior difference mask is less than or equal to a preset local offset threshold, the first mask values of the behavior difference mask are replaced by the corresponding data elements of the behavior prediction data of the current data frame, and the second mask values are replaced by the corresponding data elements of the historical behavior data in sequence, so as to generate the behavior prediction result corresponding to the current data frame. When the total number of the first mask values in the behavior difference mask is greater than the preset local offset threshold, the behavior prediction data of the current data frame is output as the behavior prediction result corresponding to the current data frame and stored in the shared cache; and the intention encoding is performed according to the behavior prediction data of the current data frame, and the intention encoding result corresponding to the current data frame is output and stored in the shared cache.

[0081] In the embodiment, the difference between the historical behavior data and the current calculated behavior is determined by the point-to-point behavior difference, the local replacement is suitable for subtle changes, and the full update is suitable for large scene mutations. The performance and real-time performance are considered, and the calculation overhead is significantly reduced. The difference is converted into a behavior difference mask, and the point-level local behavior replacement is performed according to the behavior difference mask, so as to avoid indiscriminate rewriting of the entire behavior when the behavior changes slightly, improve the cache reuse rate, and also help to accurately control the synchronization update of the behavior data point and the intention vector, improve the effect and safety of the downstream decision module, facilitate the implementation of the local update process based on the mask control at the hardware layer, and have good deployability and scalability.

[0082] Taking trajectory prediction as an example, represents a data element (trajectory point) of the current behavior prediction, represents a data element in the historical behavior data corresponding to a data element of the current behavior prediction, and the historical trajectory sequence is , and the current predicted trajectory sequence is The point-to-point difference comparison is performed on the trajectory points of the current frame and the previous frame, and the point-to-point error information can be expressed as: , is the total number of trajectory points. A preset difference threshold is set, the point-to-point error information is converted into a behavior difference mask : . Among them, indicates that the data element (trajectory point) corresponding to the time step does not change significantly, indicates that the current prediction result and the historical data differ greatly, which can be dynamically adjusted according to the scene sensitivity. Among them, the number of data elements that need to be updated is the total number of first mask values, if , the full behavior data update mode is performed. If , which can be dynamically adjusted according to the scene sensitivity, the local behavior data update mode is triggered, and the system performs cache update according to the following strategy: when , the data element of the historical behavior data of the previous frame is directly reused; when , the data element in the behavior prediction data calculated again is used for replacement. The final behavior prediction result can be expressed as: .

[0083] In the embodiment, the bit-by-bit judgment supports behavior difference detection and point-level update in any form and any position, avoids indiscriminate replacement operation on the entire behavior, and significantly reduces the calculation overhead. At the same time, it is convenient to implement local update process based on mask control at the hardware layer, and has good deployability and scalability.

[0084] Further, based on the above embodiment, the generation process of the intention encoding result can include: reading the historical intention encoding data of the last frame having semantic association with the behavior prediction result from the shared cache; sequentially traversing each mask value of the behavior difference mask, if it is the second mask value, the data element corresponding to the second mask value in the behavior prediction result is taken as a data prediction point, the intention encoding is performed on the data prediction point, and the new point intention encoding value is used to replace the second mask value; if it is the first mask value, the historical point intention encoding value corresponding to the first mask value in the historical intention encoding data is obtained, and the historical point intention encoding value is used to replace the first mask value; the replaced behavior difference mask is output as the intention encoding result corresponding to the current data frame, and is stored in the shared cache.

[0085] In the embodiment, local behavior change will affect the intention expression, the embodiment only re-encodes the part of the data element sequence corresponding to the changed behavior segment, generates an updated local intention vector, and splices it into the historical intention encoding sequence, avoiding the resource waste caused by full-amount intention vector re-encoding. For example, the intention encoding vector corresponding to the i-th data element has semantic association , the mask re-encoding update can be performed according to the following relationship:

[0086] .

[0087] If the full-amount behavior data update method is used, the intention vector encoding is directly re-performed.

[0088] In order to make the person skilled in the art more clearly understand the implementation mode of the present application, the edge computing node performs an automatic driving task as an example in the embodiment, and the controlled device is a vehicle, as shown in Figure 3 and Figure 4 , the perception module is responsible for target detection and environment mapping, and after completing multi-target identification, the structured state information of the target (such as the position, state, direction, etc. of the traffic element) is written into the shared cache in real time; the prediction module reads the existing historical trajectory information and target state in the cache, combines the current perception result and the intention vector, and generates the trajectory prediction at the next moment; the trajectory cache update system dynamically judges whether to perform trajectory freezing, frame skipping update or local trajectory point replacement according to the edge agent state, traffic signal and historical prediction information, and controls the cache refresh frequency through the scheduling signal, which can include the following contents:

[0089] In the embodiment, a shared cache is constructed in advance, and an interface of the shared cache is abstracted as a read cache (key) and a write cache (key-value). The state of a traffic light, road condition information, vehicle speed and spatial position information, historical driving trajectory and intention encoding data of the vehicle are stored in the shared cache in a key-value structured form, to provide continuous and stable state input for a path prediction or behavior decision module of the vehicle, and support multiplexing of multiple frames of data and decoupled access between modules.

[0090] The shared cache of the embodiment stores structured environment state information of traffic lights and historical trajectory data, and supports four fine-grained update strategies (prediction data cache freezing mode, frame skipping update mode, local behavior data update mode, and full-amount behavior data update mode) for prediction result inference, which are respectively trajectory freezing strategy, frame skipping update strategy, local trajectory replacement strategy, and full-amount trajectory update strategy. The trajectory freezing strategy is applicable to the case of standby, static or task suspension state (for example, the vehicle is in a static state controlled by a traffic light or a signal light, and the waiting time is long), and skips trajectory prediction and cache update for a period of time, only updates the timestamp and state sequence, until the state change is detected, and the trajectory freezing is released. The frame skipping update strategy is applicable to the case of slow task advancement and small path change (for example, low-speed driving scenarios such as traffic congestion, the overall movement trend of the vehicle is extremely slow, and the trajectory changes extremely little between continuous frames), and the system can skip the trajectory prediction, cache update and intention vector encoding of the current frame, reuse the trajectory data of the last frame, and only update the timestamp and state sequence. The local trajectory replacement strategy is applicable to the scenario where the overall trajectory changes are not significant but there is a local offset. The strategy first compares the trajectory points of the current frame and the last frame point by point, obtains the trajectory error and converts it into a trajectory difference mask, and updates the trajectory points or the full-amount trajectory according to the trajectory difference mask.

[0091] In the process of vehicle driving, the traffic light state, the distance between the current predicted trajectory and the stop line, the vehicle driving speed, the speed change rate per unit time of the vehicle, and the regional congestion factor are read from the shared cache. If the vehicle driving speed is less than a preset vehicle speed threshold, the speed change rate is close to 0, and the regional congestion factor is greater than a preset regional congestion threshold, the trajectory prediction operation is not triggered, otherwise, the trajectory prediction is triggered, and is set. Further, it is judged whether the vehicle is in a red light state and the distance between the current predicted trajectory and the stop line is less than a preset distance threshold in a future time period. If yes, the trajectory freezing strategy is triggered, and is set. If no, the frame number of continuous use of the trajectory freezing strategy and the frame skipping update strategy is less than a preset threshold, the frame skipping update strategy is triggered, and is set. If the trajectory freezing strategy and the frame skipping update strategy are used, the trajectory cache update is paused, and the timestamp and or cache into the shared cache. In the trajectory freezing strategy and the frame skipping update strategy, once the corresponding condition is not met, the normal trajectory prediction is restored, that is, the trajectory prediction is set to be normal If the prediction module needs to perform trajectory prediction, the trajectory error of the trajectory prediction sequence and the last frame historical trajectory sequence in the shared cache is calculated, the trajectory error is converted into a trajectory difference mask, if the trajectory difference mask is identified as a local trajectory replacement strategy, the local trajectory update is performed, and after the local trajectory update strategy is executed, the intention encoding is performed according to the trajectory difference part, and is spliced with the last frame intention encoding as the current intention encoding result, finally the updated trajectory prediction data and intention encoding result are stored in the shared cache, and the full trajectory and full intention encoding update strategy is triggered.

[0092] As can be seen from the above, the embodiment avoids redundant calculation and data replication, reduces repeated I / O operations, and significantly improves the data processing efficiency in a resource-constrained environment through the cross-module shared cache mechanism. Considering the vehicle running state, environmental indication information, and the change degree of the predicted trajectory, etc. multi-dimensional input, the appropriate strategy is dynamically selected according to the trajectory freezing (stationary, signal controlled, task pause state), frame skipping update (low speed buffer), local trajectory point replacement or whole frame trajectory re-prediction, the on-demand update of the trajectory data is realized, the update strategy of the trajectory data is optimized, the cache and communication load is reduced, the cache pressure and energy consumption burden caused by frequent writing are reduced, the redundant calculation pressure of the trajectory cache in the edge computing system is effectively relieved, the system computing load is significantly reduced while maintaining the task continuity and prediction accuracy, the stability of the trajectory data and the system response efficiency in the key task scene are guaranteed, and the real-time running demand in the diversified complex scene is adapted, providing efficient and stable support ability for the intelligent system in the complex traffic or mobile scene.

[0093] Through the description of the above embodiments, those skilled in the art can clearly understand that the method according to the above embodiments can be realized by means of software and the necessary general hardware platform, of course, it can also be realized by hardware, but in many cases the former is a better embodiment.

[0094] The application also provides a corresponding device for the dynamic environment adaptive control method, which further makes the method more practical. The device can be described from the perspective of functional modules and the perspective of hardware. The dynamic environment adaptive control device provided by the application is introduced below. The device is used to implement the dynamic environment adaptive control method provided by the application. In this embodiment, the dynamic environment adaptive control device can include or be divided into one or more program modules, which are stored in a storage medium and executed by one or more processors to complete the dynamic environment adaptive control method disclosed in embodiment one. The program module referred to in this embodiment refers to a series of computer program instruction segments that can complete a specific function. It is more suitable to describe the execution process of the dynamic environment adaptive control device in the storage medium than the program itself. The following description will specifically introduce the functions of each program module in this embodiment. The dynamic environment adaptive control device described below can be referred to in conjunction with the dynamic environment adaptive control method described above.

[0095] Based on the perspective of functional modules, see Figure 5 , Figure 5 The structure diagram of the dynamic environment adaptive control device provided in this embodiment is shown in a specific embodiment. The device can include:

[0096] The prediction update mode determination module 501 is configured to determine the prediction result inference mode of the to-be-controlled device according to the environment perception data and the running state data of the to-be-controlled device in a preset time period. The environment perception data is cached in a pre-constructed shared cache after being perceived.

[0097] The stop prediction module 502 is configured to cache the behavior prediction result stop update identification parameter as the behavior prediction result in the shared cache if the prediction result inference mode is the behavior prediction stop update mode.

[0098] The local prediction update module 503 is configured to read the historical behavior data of the last frame of the to-be-controlled device from the shared cache if the prediction result inference mode is the behavior prediction mode. If the historical behavior data and the part of the data elements of the behavior prediction data of the current data frame are different, the behavior prediction result corresponding to the current data frame is determined according to the behavior difference data and the historical behavior data.

[0099] In some embodiments of the present embodiment, the prediction update mode determination module 501 can further include: obtaining target running state data at a target time, historical running state data and historical environment perception data in a preset time period; the target time is the perception time corresponding to the current data frame; if it is determined according to the target running state data that the action amplitude of the to-be-controlled device is less than a preset action amplitude threshold, it is determined according to the historical running state data that the action change amount of the to-be-controlled device in the preset time period satisfies a preset quasi-zero acceleration state condition, and it is determined according to the historical environment perception data that a plurality of target objects exist in the behavior front space of the to-be-controlled device and satisfy a preset action maintenance condition, then the prediction result inference mode is a behavior prediction stop update mode; if the prediction result inference mode is not the behavior prediction stop update mode, then the prediction result inference mode is a behavior prediction mode.

[0100] As an exemplary embodiment of the above embodiment, the prediction update mode determination module 501 can be further used to: obtain environment indication signal data in the environment perception data at the target time; obtain a behavior stop area in the environment perception data at the target time and a target space position of the to-be-controlled device at the target time, and calculate a forward travel distance between the to-be-controlled device and the behavior stop area along the forward action direction; if the environment indication signal data remains unchanged in a preset future time period, and the forward travel distance is less than a preset distance threshold, then the behavior prediction stop update mode is a prediction data cache freezing mode; if the environment indication signal data changes in a preset future time period, or the forward travel distance is greater than or equal to the preset distance threshold, and there is no continuous multiple frames using the behavior prediction stop update mode before the current data frame, then the behavior prediction stop update mode is a frame skipping update mode.

[0101] As an exemplary embodiment of the above embodiment, the prediction update mode determination module 501 can be further used to: if the behavior prediction stop update mode is the prediction data cache freezing mode, then the generation of the next frame of behavior prediction data of the to-be-controlled device is inhibited, at the same time, the behavior prediction state of the current data frame is set to a first identification value identifying the prediction data cache freezing mode, and the first identification value and the corresponding timestamp are taken as behavior prediction result stop update identification parameters; if the behavior prediction stop update mode is the frame skipping update mode, then the historical behavior data of the last frame is taken as the behavior prediction data of the next frame of the to-be-controlled device, at the same time, the behavior prediction state of the current data frame is set to a second identification value identifying the frame skipping update mode, and the second identification value and the corresponding timestamp are taken as behavior prediction result stop update identification parameters.

[0102] As another exemplary implementation of the above embodiment, the prediction update mode determination module 501 can be further configured to: if the new environment perception data or the new running state data does not satisfy the condition corresponding to the behavior prediction stop update mode, set the behavior prediction state of the current data frame to a third identification value, and the third identification value represents a resumption of the normal behavior prediction operation and the cache update operation.

[0103] Exemplarily, in some other implementations of the present embodiment, the partial prediction update module 503 can further include: calculating the difference between each data element of the historical behavior data and the data element corresponding to the behavior prediction data of the current data frame; if the total number of target data elements greater than or equal to the preset difference threshold value is less than or equal to the preset partial offset threshold value, the behavior prediction mode is the partial behavior data update mode, the behavior prediction result corresponding to the current data frame is determined according to the numerical value of each target data element in the behavior prediction data of the current data frame and the historical behavior data, and is stored in the shared cache; if the total number of target data elements greater than or equal to the preset difference threshold value is less than the preset partial offset threshold value, the behavior prediction mode is the full behavior data update mode, the behavior prediction data of the current data frame is taken as the behavior prediction result corresponding to the current data frame, and is stored in the shared cache.

[0104] Exemplarily, in some other implementations of the present embodiment, the device can further include a cache module, which can be configured to: whenever a perception generates an environment perception data frame, extract the spatial position data and / or the attitude data and / or the state data of each environment target from the environment perception data frame as the environment target perception data of each environment target; take the data frame number of the environment perception data frame as an index, generate the environment cache structured data corresponding to the environment perception data frame according to the timestamp of the environment perception data frame and the environment target perception data; take the timestamp of the environment perception data frame as a key and the environment cache structured data as a value, write the environment cache structured data as a cache entry in the environment state information area of the shared cache in a key-value pair data structure manner, so as to read each environment target perception data from the shared cache through a read request command containing the key.

[0105] As an exemplary implementation of the above embodiment, the cache module can be further configured to: obtain the data type corresponding to the state data of the environment cache structured data as first type description information; obtain the data types of the timestamp, the spatial position data and the attitude data of the environment perception data frame as second type description information; take the first type description information and the second type description information as field type constraint description information; write the environment cache structured data, the first type description information and the second type description information as a cache entry in the environment state information area of the shared cache in a data cache format of the key-value pair data structure manner and the field type constraint description information.

[0106] Exemplarily, in some other embodiments of the present embodiment, the cache module can further comprise: when the historical behavior data of the to-be-controlled device is generated according to the previous frame of input data, taking the data frame number of the previous frame of input data as an index, generating the behavior prediction structured data of the previous frame of input data according to the historical behavior data, the time stamp of the previous frame of input data and the device identification information of the to-be-controlled device; taking the time stamp of the previous frame of input data as a key and the behavior prediction structured data as a value, writing the behavior prediction structured data as a cache entry in the historical behavior information area of the shared cache in a key-value pair data structure manner, so as to read each frame of historical behavior data from the shared cache through a read request command containing the key.

[0107] As an exemplary embodiment of the above embodiment, the cache module can be further used for: when a historical data read instruction containing a target time stamp is received, determining a data read starting time stamp according to the maximum window length and the target time stamp; reading each target historical behavior data corresponding to the data read starting time stamp to the target time stamp from the shared cache.

[0108] As another exemplary embodiment of the above embodiment, the cache module can be further used for: obtaining the historical intention encoding data corresponding to the previous frame of input data; generating the behavior prediction structured data of the previous frame of input data according to the historical trajectory data and the corresponding historical intention encoding data, the time stamp of the previous frame of input data and the device identification information of the to-be-controlled device.

[0109] As another exemplary embodiment of the above embodiment, the cache module can be further used for: obtaining the time stamp of the behavior prediction structured data, the position data of each trajectory point of the historical trajectory data, the data type of the historical intention encoding data and the device identification information as field type constraint description information; writing the behavior prediction structured data and each data type as a cache entry in the historical behavior information area of the shared cache in a key-value pair data structure manner and a data cache format of the field type constraint description information.

[0110] Exemplarily, in some other embodiments of the present embodiment, the device can further comprise an intention encoding module, which can: read the historical intention encoding data of the previous frame which is semantically associated with the behavior prediction result from the shared cache; encode the target data element different from the data element of the historical behavior data in the behavior prediction data of the current data frame to generate local intention encoding data; and splice the local intention encoding data to the historical intention encoding data to generate the intention encoding result corresponding to the current data frame.

[0111] In some other embodiments of the present embodiment, the local prediction update module 503 can be further configured to: take the historical behavior data and the data elements of the behavior prediction data of the current data frame as data points, generate point-to-point error information according to the point-to-point difference between the historical behavior data and the behavior prediction data of the current data frame; compare each point-to-point difference in the point-to-point error information with the preset difference threshold value in sequence, and assign the point-to-point difference that is greater than or equal to the preset difference threshold value as a first mask value, and assign the point-to-point difference that is less than the preset difference threshold value as a second mask value, so as to convert the point-to-point error information into a behavior difference mask; and when the total number of the first mask values in the behavior difference mask is less than or equal to a preset local offset threshold value, replace the first mask values of the behavior difference mask with the corresponding data elements of the behavior prediction data of the current data frame in sequence, and replace the second mask values with the corresponding data elements of the historical behavior data, so as to generate the behavior prediction result corresponding to the current data frame.

[0112] As an exemplary embodiment of the above embodiment, the intention encoding module can be further configured to: read the historical intention encoding data of the previous frame that is semantically associated with the behavior prediction result from the shared cache; sequentially traverse each mask value of the behavior difference mask, if the mask value is the second mask value, take the data element of the behavior prediction result corresponding to the second mask value as a data prediction point, encode the data prediction point, and replace the second mask value with a new point intention encoding value; if the mask value is the first mask value, obtain the historical point intention encoding value of the historical intention encoding data corresponding to the first mask value, and replace the first mask value with the historical point intention encoding value; output the replaced behavior difference mask as the intention encoding result corresponding to the current data frame, and store the intention encoding result in the shared cache.

[0113] As another exemplary embodiment of the above embodiment, the local prediction update module 503 can be further configured to: when the total number of the first mask values in the behavior difference mask is greater than the preset local offset threshold value, output the behavior prediction data of the current data frame as the behavior prediction result corresponding to the current data frame, and store the behavior prediction result in the shared cache; encode the behavior prediction data of the current data frame as the intention encoding result corresponding to the current data frame, and output and store the intention encoding result in the shared cache.

[0114] The dynamic environment adaptive control device mentioned above is described from the perspective of functional modules. Further, the present application also provides an electronic device, which is described from the perspective of hardware. Figure 6 The electronic device provided by the present embodiment has the structure shown in the structural schematic diagram of an embodiment. The electronic device includes a memory 60 and a processor 61, the memory 60 stores a computer program, and the processor 61 is configured to run the computer program to execute the steps in any of the above dynamic environment adaptive control method embodiments.

[0115] The embodiment of the present application further provides a computer readable storage medium, which stores a computer program, and the computer program is configured to execute the steps in any of the above dynamic environment adaptive control method embodiments when running.

[0116] In an example embodiment, the above computer readable storage medium can include, but is not limited to, a U disk, a read-only memory (ROM), a random access memory (RAM), a mobile hard disk, a magnetic disk or an optical disk, and various media that can store computer programs.

[0117] The embodiment of the present application further provides a computer program product, which includes a computer program, and the computer program is executed by a processor to implement the steps in any of the above dynamic environment adaptive control method embodiments.

[0118] The embodiment of the present application further provides another computer program product, which includes a non-volatile computer readable storage medium or a non-volatile storage medium, and the non-volatile computer readable storage medium stores a computer program, and the computer program is executed by a processor to implement the steps in any of the above dynamic environment adaptive control method embodiments.

[0119] The above describes in detail a dynamic environment adaptive control method, device, electronic equipment, non-volatile storage medium and computer program product provided by the present application. In the present specification, each embodiment is described in a progressive manner, and each embodiment mainly describes the difference from other embodiments. The same or similar parts of each embodiment can be referred to each other. The units and algorithm steps of each example described in each disclosed embodiment are executed in the form of electronic hardware or computer software, which depends on the specific application and design constraints of the technical solution. The skilled in the art can use different methods to implement the described functions for each specific application, and such implementation should not be considered beyond the scope of the present application. Without departing from the principles of the present application, the present application can be improved and modified in several ways, and these improvements and modifications also fall within the protection scope of the present application.

Claims

1. A dynamic environment adaptive control method, characterized in that, include: Based on environmental sensing data and the operating status data of the device to be controlled within a preset time period, the prediction result inference method of the device to be controlled is determined; wherein, the environmental sensing data is cached in a pre-built shared cache after being sensed; If the prediction result inference method is the behavior prediction stop update method, then the behavior prediction result stop update flag parameter is cached as the behavior prediction result in the shared cache; If the prediction result inference method is behavior prediction method, then read the historical behavior data of the previous frame of the device to be controlled from the shared cache. If some data elements of the historical behavior data and the behavior prediction data of the current data frame are different, then determine the behavior prediction result corresponding to the current data frame based on the behavior difference data and the historical behavior data. The process for determining the prediction result inference method of the device to be controlled is as follows: The system acquires environmental indication signal data from the environmental perception data at the target time; it acquires the behavior stop area and the target spatial position of the device to be controlled at the target time from the environmental perception data at the target time, and calculates the forward travel distance between the device to be controlled and the behavior stop area along the forward movement direction; if the environmental indication signal data remains unchanged within a preset future time period and the forward travel distance is less than a preset distance threshold, the behavior prediction stop update method is the prediction data cache freeze method; if the environmental indication signal data changes within a preset future time period, or the forward travel distance is greater than or equal to the preset distance threshold, and there are no consecutive frames before this data frame that use the behavior prediction stop update method, the behavior prediction stop update method is the frame skip update method. Accordingly, the process of stopping the update of the behavior prediction result identifier parameter as a behavior prediction result cache is as follows: If the behavior prediction stop update method is the prediction data cache freeze method, then the generation of behavior prediction data for the next frame of the device to be controlled is suppressed. At the same time, the behavior prediction status of the current data frame is set to the first identifier value that identifies the prediction data cache freeze method, and the first identifier value and the corresponding timestamp are used as the behavior prediction result stop update identifier parameter. If the behavior prediction stop update method is the frame skip update method, then the historical behavior data of the previous frame is used as the behavior prediction data for the next frame of the device to be controlled. At the same time, the behavior prediction status of the current data frame is set to the second identifier value that identifies the frame skip update method, and the second identifier value and the corresponding timestamp are used as the behavior prediction result stop update identifier parameter.

2. The dynamic environment adaptive control method according to claim 1, characterized in that, Based on environmental perception data and the operating status data of the device to be controlled within a preset time period, the prediction result inference method for the device to be controlled is determined, including: Acquire target operating status data at a target time, historical operating status data within a preset time period, and historical environmental perception data; the target time is the perception time corresponding to the current data frame. If, based on the target operating status data, it is determined that the action amplitude of the device to be controlled is less than a preset action amplitude threshold, based on the historical operating status data, it is determined that the change in the action of the device to be controlled within a preset time period meets a preset quasi-zero acceleration state condition, and based on the historical environmental perception data, it is determined that there are multiple target objects in the forward space of the device to be controlled that meet a preset action maintenance condition, then the prediction result inference method is the behavior prediction stop update method. If the prediction result inference method is not the behavior prediction stop update method, then the prediction result inference method is the behavior prediction method.

3. The dynamic environment adaptive control method according to claim 2, characterized in that, Also includes: If the new environmental perception data or the new operating status data does not meet the conditions corresponding to the behavior prediction stop update method, then the behavior prediction status of this data frame is set to the third identifier value, which indicates the resumption of normal behavior prediction operation and cache update operation.

4. The dynamic environment adaptive control method according to claim 1, characterized in that, The behavior prediction result corresponding to the current data frame is determined based on the behavior difference data and the historical behavior data, including: Calculate the difference between each data element of the historical behavior data and the corresponding data element of the behavior prediction data of the current data frame; If the total number of target data elements greater than or equal to the preset difference threshold is less than or equal to the preset local offset threshold, then the behavior prediction method is the local behavior data update method. The behavior prediction result corresponding to the current data frame is determined based on the value of the behavior prediction data of each target data element in the current data frame and the historical behavior data, and stored in the shared cache. If the total number of target data elements greater than or equal to the preset difference threshold is greater than the preset local offset threshold, then the behavior prediction method is the full behavior data update method, and the behavior prediction data of the current data frame is used as the behavior prediction result corresponding to the current data frame and stored in the shared cache.

5. The dynamic environment adaptive control method according to claim 1, characterized in that, Before basing the data on environmental sensing data and the operating status data of the controlled equipment within a preset time period, the following is also included: Whenever an environmental perception data frame is generated, the spatial location data and / or attitude data and / or state data of each environmental target are extracted from the environmental perception data frame to serve as the environmental target perception data of each environmental target. Using the data frame number of the environmental perception data frame as an index, environmental cache structured data corresponding to the environmental perception data frame is generated based on the timestamp of the environmental perception data frame and the perception data of each environmental target. Using the timestamp of the environmental perception data frame as the key and the environmental cache structured data as the value, the environmental cache structured data is written as a cache entry into the environmental state information area of ​​the shared cache in a key-value pair data structure manner, so as to read the environmental target perception data from the shared cache through a read request command containing the key.

6. The dynamic environment adaptive control method according to claim 5, characterized in that, Following a key-value pair data structure, the structured environmental cache data is written as a cache entry to the environment state information area of ​​the shared cache, including: Obtain the data type corresponding to the state data of the structured data in the environment cache, and use it as the first type of description information; The data types of the timestamp, spatial location data, and attitude data of the environmental perception data frame are obtained as the second type of descriptive information; The first type of description information and the second type of description information are used as field type constraint description information; According to the key-value pair data structure and the data caching format of the field type constraint description information, the environment cache structured data, the first type description information and the second type description information are written as a cache entry to the environment state information area of ​​the shared cache.

7. The dynamic environment adaptive control method according to claim 1, characterized in that, Read the historical behavior data of the device to be controlled from the shared cache for the previous frame, including: When the historical behavior data of the device to be controlled is generated based on the input data of the previous frame, the data frame number of the input data of the previous frame is used as an index, and the behavior prediction structured data of the input data of the previous frame is generated based on the historical behavior data, the timestamp of the input data of the previous frame, and the device identification information of the device to be controlled. Using the timestamp of the previous frame's input data as the key and the behavior prediction structured data as the value, the behavior prediction structured data is written as a cache entry into the historical behavior information area of ​​the shared cache in a key-value pair data structure manner, so that historical behavior data of each frame can be read from the shared cache through a read request command containing the key.

8. The dynamic environment adaptive control method according to claim 7, characterized in that, The shared cache maintains the historical behavior information region in a circular array. Historical behavior data for each frame is read from the shared cache, including: When a historical data read instruction containing a target timestamp is received, the data read start timestamp is determined based on the maximum window length and the target timestamp. Read the historical behavior data of each target from the data reading start timestamp to the target timestamp from the shared cache.

9. The dynamic environment adaptive control method according to claim 7, characterized in that, The behavior prediction data of the device to be controlled is trajectory prediction data, and the historical behavior data is historical trajectory data. Based on the historical behavior data, the timestamp of the previous frame input data, and the device identification information of the device to be controlled, structured behavior prediction data for the previous frame input data is generated, including: Obtain the historical intent encoding data corresponding to the input data of the previous frame; Based on historical trajectory data and corresponding historical intent encoding data, the timestamp of the previous frame input data, and the device identification information of the device to be controlled, behavioral prediction structured data of the previous frame input data is generated.

10. The dynamic environment adaptive control method according to claim 9, characterized in that, Following a key-value pair data structure, the structured data for behavior prediction is written as a cache entry to the historical behavior information area of ​​the shared cache, including: The data types of the behavior prediction structured data, the location data of each trajectory point in the historical trajectory data, the historical intent encoding data, and the device identification information are obtained as field type constraint description information. According to the key-value pair data structure and the data caching format of the field type constraint description information, the behavior prediction structured data and each data type are written as a cache entry to the historical behavior information area of ​​the shared cache.

11. The dynamic environment adaptive control method according to any one of claims 1 to 10, characterized in that, After determining the behavior prediction result corresponding to the current data frame based on the behavior difference data and the historical behavior data, the method further includes: Read historical intent-encoded data from the previous frame that is semantically related to the behavior prediction result from the shared cache; For the target data elements in the behavior prediction data of the current data frame that are different from the data elements in the historical behavior data, intent encoding is performed to generate local intent encoding data; The local intent-encoded data is concatenated with the historical intent-encoded data to generate the intent-encoded result corresponding to the current data frame.

12. The dynamic environment adaptive control method according to any one of claims 1 to 10, characterized in that, The behavior prediction result corresponding to the current data frame is determined based on the behavior difference data and the historical behavior data, including: Using the data elements of the historical behavior data and the behavior prediction data of the current data frame as data points, point-to-point error information is generated based on the point-to-point difference between the historical behavior data and the behavior prediction data of the current data frame. The point-to-point difference values ​​in the point-to-point error information are compared with a preset difference threshold value in turn. Point-to-point difference values ​​that are greater than or equal to the preset difference threshold value are assigned as the first mask value, and point-to-point difference values ​​that are less than the preset difference threshold value are assigned as the second mask value, so as to convert the point-to-point error information into a line difference mask. When the total number of first mask values ​​in the behavior difference mask is less than or equal to a preset local offset threshold, the behavior prediction result corresponding to the current data frame is generated by sequentially replacing the first mask value of the behavior difference mask with the corresponding data element of the behavior prediction data of the current data frame and replacing the second mask value with the corresponding data element of the historical behavior data.

13. The dynamic environment adaptive control method according to claim 12, characterized in that, After generating the behavior prediction result corresponding to this data frame, the following is also included: Read historical intent-encoded data from the previous frame that is semantically related to the behavior prediction result from the shared cache; The mask values ​​of the behavior difference mask are traversed sequentially. If it is the second mask value, the data element corresponding to the second mask value in the behavior prediction result is used as the data prediction point. The data prediction point is then used for intent encoding, and the second mask value is replaced with the new point intent encoding value. If it is the first mask value, the historical point intent encoding value corresponding to the first mask value in the historical intent encoding data is obtained, and the first mask value is replaced with the historical point intent encoding value. The replaced behavior difference mask is output as the intent encoding result corresponding to this data frame and stored in the shared buffer.

14. The dynamic environment adaptive control method according to claim 12, characterized in that, Also includes: When the total number of the first mask values ​​in the behavior difference mask is greater than the preset local offset threshold, the behavior prediction data of the current data frame is output as the behavior prediction result corresponding to the current data frame and stored in the shared cache. Intent encoding is performed based on the behavior prediction data of the current data frame, and the result is output as the intent encoding result corresponding to the current data frame and stored in the shared cache.

15. A dynamic environment adaptive control device, characterized in that, include: The prediction update method determination module is used to determine the prediction result inference method of the device to be controlled based on the environmental perception data and the operating status data of the device to be controlled within a preset time period; wherein, the environmental perception data is cached in a pre-built shared cache after being perceived; The stop prediction module is used to cache the behavior prediction result stop update flag parameter as the behavior prediction result in the shared cache if the prediction result inference method is behavior prediction stop update method; The local prediction update module is used to read the historical behavior data of the device to be controlled from the previous frame in the shared cache if the prediction result inference method is behavior prediction method; if some data elements of the historical behavior data and the behavior prediction data of the current data frame are different, the behavior prediction result corresponding to the current data frame is determined according to the behavior difference data and the historical behavior data. The prediction update method determination module is further configured to: acquire environmental indication signal data from the environmental perception data at the target time; acquire the behavior stop area and the target spatial position of the device to be controlled at the target time from the environmental perception data at the target time, and calculate the forward travel distance between the device to be controlled and the behavior stop area along the forward movement direction; if the environmental indication signal data remains unchanged within a preset future time period and the forward travel distance is less than a preset distance threshold, then the behavior prediction stop update method is the prediction data cache freezing method; if the environmental indication signal data changes within a preset future time period, or the forward travel distance is greater than or equal to the preset distance threshold, and there are no consecutive frames using the behavior prediction stop update method before this data frame, then the behavior prediction stop update method is the frame skipping update method. The prediction update method determination module is further configured to: if the behavior prediction stop update method is the prediction data cache freeze method, then suppress the generation of behavior prediction data for the next frame of the device to be controlled, and simultaneously set the behavior prediction status of the current data frame to a first identifier value that identifies the prediction data cache freeze method, and use the first identifier value and the corresponding timestamp as behavior prediction result stop update identifier parameters; if the behavior prediction stop update method is the frame skip update method, then use the historical behavior data of the previous frame as the behavior prediction data for the next frame of the device to be controlled, and simultaneously set the behavior prediction status of the current data frame to a second identifier value that identifies the frame skip update method, and use the second identifier value and the corresponding timestamp as behavior prediction result stop update identifier parameters.

16. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor for executing the computer program to implement the steps of the dynamic environment adaptive control method as described in any one of claims 1 to 14.

17. A non-volatile storage medium, characterized in that, The non-volatile storage medium stores a computer program, which, when executed by a processor, implements the steps of the dynamic environment adaptive control method as described in any one of claims 1 to 14.

18. A computer program product comprising a computer program / instructions, characterized in that, When the computer program / instructions are executed by the processor, they implement the steps of the dynamic environment adaptive control method according to any one of claims 1 to 14.

Images