Traffic event identification method and device and medium

Abstract

The invention discloses a traffic incident recognition method and device and a medium. The method comprises the following steps: acquiring a plurality of traffic perception data; under the condition that the traffic perception data meets a preset model triggering condition, a target event recognition model corresponding to the model triggering condition is called from a model library, the model library comprises a plurality of event recognition models, and each event recognition model is used for recognizing one traffic event; the target event recognition model is deployed on a target node of a server cluster, so that the target node operates the target event recognition model, the traffic perception data is input into the target event recognition model, a first recognition result is output, and the server cluster comprises a plurality of working nodes; each working node is used for operating different event recognition models in parallel; and obtaining a first identification result corresponding to each piece of traffic perception data. According to the invention, a plurality of traffic events within a period of time can be identified in parallel, so that the working efficiency of the vehicle-road cloud integrated system is improved.

Description

Technical Field

[0001] This application belongs to the field of traffic management technology, and in particular relates to a method, device and medium for traffic incident identification. Background Technology

[0002] With the rapid development of vehicle-road-cloud integrated technology, real-time identification and rapid response to traffic incidents have become core requirements for improving traffic safety and efficiency.

[0003] Existing technologies for traffic incident identification have the following limitations: they employ a single-algorithm serial processing mode, making it impossible to handle multiple traffic incident types simultaneously. For example, it is difficult to process illegal lane change incidents in parallel while identifying rear-end collision warnings, resulting in excessively long response times. Summary of the Invention

[0004] The embodiments of this application provide a traffic event identification method, apparatus, and medium, which can at least to some extent identify multiple traffic events in parallel over a period of time, thereby helping to reduce the response time of the vehicle-road-cloud integrated system, meet the low latency requirements of the vehicle-road-cloud integrated system, and improve work efficiency.

[0005] Other features and advantages of this application will become apparent from the following detailed description, or may be learned in part from practice of this application.

[0006] The first aspect of this application provides a traffic incident identification method, including: Acquire multiple traffic sensing data; When the traffic perception data meets the preset model triggering conditions, the target event recognition model corresponding to the model triggering conditions is called from the model library. The model library includes multiple event recognition models, and each event recognition model is used to identify a traffic event. The target event recognition model is deployed on the target node of the server cluster so that the target node runs the target event recognition model and the traffic perception data is input into the target event recognition model to output a first recognition result. The server cluster includes multiple working nodes, and each working node is used to run different event recognition models in parallel. Obtain the first identification result corresponding to each of the traffic perception data.

[0007] Optionally, calling the target event recognition model corresponding to the model triggering condition from the model library includes: Obtain the model identifier corresponding to the model triggering condition; Call the target event recognition model that matches the model identifier from the model library.

[0008] Optionally, deploying the target event recognition model on the target node of the server cluster includes: Obtain the hardware resource information required by the target event recognition model; Obtain a target node in the server cluster that matches the hardware resource information; The target event recognition model is deployed on the target node.

[0009] Optionally, the first recognition result includes a result image, and after obtaining the first recognition result corresponding to each of the traffic perception data, the method further includes: The resulting image is input into the first model for spatial feature extraction to obtain a spatial feature vector. The first model is obtained by training the first network to be trained using first historical samples. The first historical samples include historical result images and corresponding spatial feature labels. The traffic perception data of consecutive frames within a preset time window is input into the second model for temporal feature extraction to obtain a temporal feature vector. The second model is obtained by training the second network to be trained through a second historical sample. The second historical sample includes traffic perception data of consecutive historical frames and corresponding temporal feature labels. The spatial feature vector and the temporal feature vector are fused to obtain a fused feature vector; The fused feature vector is input into the classifier, and a second recognition result is output. The classifier is obtained by training a third network to be trained using a third historical sample. The third historical sample includes historical fused feature vectors and corresponding recognition result labels.

[0010] Optionally, the second identification result includes: identification confidence level and response time, wherein the response time is the time elapsed from acquiring the traffic perception data to outputting the second identification result. After outputting the second identification result, the method further includes: Obtain the second identification result for each of the traffic perception data; For each second identification result, the identification confidence is weighted based on a preset first weighting factor, and the response time is weighted based on a preset second weighting factor. The evaluation value of the second identification result is determined based on the weighted identification confidence and the weighted response time. The second identification result corresponding to the first n evaluation values ​​among the multiple evaluation values ​​is output, where n is an integer greater than or equal to 1.

[0011] Optionally, after obtaining the second identification result of each of the traffic perception data, the method further includes: Second identification results with a confidence level less than or equal to a preset confidence level threshold will be filtered out.

[0012] Optionally, before outputting the second recognition result corresponding to the first n evaluation values ​​among the plurality of evaluation values, the method further includes: Obtain target recognition results from multiple second recognition results whose evaluation values ​​are greater than or equal to preset evaluation values; If the number of target identification results is greater than n, and the target identification results include the identification results of target traffic events, then the identification results of the target traffic events will be used as the second identification result to be output.

[0013] Optionally, it also includes: Obtain the ambient temperature; The model triggering conditions are adjusted according to the ambient temperature, and / or the time window of the second model is adjusted according to the ambient temperature.

[0014] A second aspect of this application provides a traffic incident recognition device, comprising: The first acquisition unit is used to acquire multiple traffic sensing data; The calling unit is used to call a target event recognition model corresponding to the model triggering condition from the model library when the traffic perception data meets the preset model triggering condition. The model library includes multiple event recognition models, and each event recognition model is used to identify a traffic event. The deployment unit is used to deploy the target event recognition model on the target node of the server cluster, so that the target node runs the target event recognition model, inputs the traffic perception data into the target event recognition model, and outputs a first recognition result. The server cluster includes multiple working nodes, and each working node is used to run different event recognition models in parallel. The second acquisition unit is used to acquire the first identification result corresponding to each of the traffic perception data.

[0015] A third aspect of this application provides a computer-readable storage medium storing at least one computer program instruction, which is loaded and executed by a processor to perform the operations described in any of the methods described in the first aspect.

[0016] A fourth aspect of this application provides an electronic device including one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, and the at least one piece of program code is loaded and executed by the one or more processors to perform the operation as described in any of the methods in the first aspect.

[0017] The embodiments of the present invention provide one or more technical solutions that achieve at least the following technical effects or advantages: The traffic event recognition method of this application includes: acquiring multiple traffic perception data; when the traffic perception data meets preset model triggering conditions, calling a target event recognition model corresponding to the model triggering conditions from a model library, wherein the model library includes multiple event recognition models, each event recognition model is used to recognize a traffic event; deploying the target event recognition model on a target node of a server cluster, so that the target node runs the target event recognition model, inputting traffic perception data into the target event recognition model, and outputting a first recognition result, wherein the server cluster includes multiple working nodes, each working node is used to run different event recognition models in parallel; and acquiring the first recognition result corresponding to each traffic perception data. Therefore, this application embodiment can identify multiple traffic events in parallel over a period of time, thereby helping to reduce the response time of the vehicle-road-cloud integrated system, meeting the low latency requirements of the vehicle-road-cloud integrated system, and improving work efficiency.

[0018] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit this application. Attached Figure Description

[0019] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can derive other drawings from these drawings without creative effort. In the drawings: Figure 1 A flowchart of a traffic incident identification method according to an embodiment of this application is shown; Figure 2 The system architecture diagram on which the traffic incident identification method of this application is based is shown; Figure 3 A timing diagram of the traffic incident identification method according to an embodiment of this application is shown; Figure 4 A structural diagram of a traffic incident recognition device according to an embodiment of this application is shown; Figure 5 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown. Detailed Implementation

[0020] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of the embodiments. Based on the embodiments of this application, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of this application.

[0021] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, components, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0022] The block diagrams shown in the accompanying drawings are merely functional entities and do not necessarily correspond to physically independent entities. That is, these functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different models and / or processor devices and / or microcontroller devices.

[0023] The flowcharts shown in the accompanying drawings are merely illustrative and do not necessarily include all content and operations / steps, nor do they necessarily have to be performed in the described order. For example, some operations / steps can be broken down, while others can be combined or partially combined; therefore, the actual execution order may change depending on the specific circumstances.

[0024] It should also be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such uses of these terms can be interchanged where appropriate so that the embodiments of this application described herein can be implemented in orders other than those illustrated or described.

[0025] With the rapid development of vehicle-road-cloud integration technology (referring to intelligent transportation systems that deeply integrate vehicles, road infrastructure, and cloud platforms through next-generation information and communication technologies to achieve real-time data interaction, collaborative perception, and intelligent control), real-time identification and rapid response to traffic incidents have become core requirements for improving road safety and traffic efficiency. However, existing technological systems still have significant limitations in dealing with complex and ever-changing traffic scenarios, mainly manifested in the singleness and lag of identification modes.

[0026] Most current mainstream processing frameworks rely on a single-algorithm serial processing mode, meaning the system can only detect and analyze one specific type of event at a time. For example, when the system focuses on identifying the risk of rear-end collisions on the road ahead, it struggles to simultaneously process multiple events that may occur on the same road segment, such as illegal lane changes, pedestrian intrusions, or abnormal parking. This processing mechanism not only leads to underutilization of computing resources but also causes a delay in the time window for identifying critical events. In high-speed and high-density traffic flows, even a delay of just a few seconds can render the warning information worthless, thus missing the best opportunity to avoid accidents or alleviate congestion.

[0027] In view of this, embodiments of this application propose a traffic event identification method, which can at least to some extent identify multiple traffic events in parallel over a period of time, thereby helping to reduce the response time of the vehicle-road-cloud integrated system, meet the low latency requirements of the vehicle-road-cloud integrated system, and improve work efficiency.

[0028] The traffic incident identification method of this application embodiment will be described below with reference to the accompanying drawings.

[0029] Figure 1 A flowchart of a traffic incident identification method according to an embodiment of this application is shown.

[0030] The first aspect of this application provides a traffic incident identification method, which can be executed on a cloud server, and the cloud server can be deployed with a traffic incident identification system. The method includes, but is not limited to: Step S10. Acquire multiple traffic sensing data; For example, raw perception data collected by various sensors deployed on the roadside is continuously received through a pre-defined hardware interface. These raw perception data may differ in physical form, update frequency, and information dimensions. Therefore, after acquiring several sets of raw perception data, it is necessary to fuse and clean the raw perception data, and perform operations such as noise reduction, time synchronization, and format standardization to obtain multiple sets of traffic perception data.

[0031] In some embodiments, roadside sensors include, but are not limited to, cameras, millimeter-wave radar, lidar, and weather sensors. Cameras continuously capture video streams and output image sequences after decoding. Millimeter-wave radar stably outputs radial distance, velocity, and azimuth information of detected targets at a preset frequency. LiDAR provides high-precision three-dimensional spatial perception; its emitted laser beam forms a reflective point cloud in the environment, ensuring high point cloud density in key areas and accurately depicting the contours of detected targets such as roads and obstacles. Weather sensors monitor environmental conditions at a preset frequency, collecting key parameters affecting traffic behavior, such as visibility and rainfall intensity.

[0032] Understandably, heterogeneous data collected by different sensors can be accessed by traffic incident recognition systems through diverse interfaces. For example, video streams are typically retrieved via Ethernet (1000BASE-T) following the RTSP protocol; radar data (target lists from millimeter-wave radar, point cloud packets from lidar) are mostly transmitted using the high real-time UDP protocol; and control and status information from some vehicle-mounted or roadside equipment may be transmitted via the CANFD bus (with a speed of up to 500kbps).

[0033] In some embodiments, after acquiring the raw sensing data, the raw sensing data can be time-synchronized. For example, each frame of raw sensing data can be timestamped based on the High Precision Network Time Protocol (NTP) to ensure time alignment across sensor data. Then, the raw sensing data undergoes data cleaning, such as applying a mean square filtering algorithm to radar point clouds to effectively suppress isolated noise points and improve data quality. Finally, the cleaned and synchronized data is encapsulated into the efficient Protobuf serialization format. This format is not only structurally clear and highly scalable, but also has a small encoding size and fast parsing speed, improving data exchange efficiency.

[0034] In the embodiments of this application, each traffic perception data can refer to data corresponding to a traffic event; that is, each traffic perception data can have multiple sub-data. For example, for a potential rear-end collision event, the corresponding traffic perception data may include: event ID, geographical location, information of the vehicles involved (e.g., vehicle model, license plate number, real-time vehicle speed, etc.), relative information between vehicles (e.g., relative distance, relative speed, collision time, etc.), video stream data, radar data, environmental data, etc.

[0035] Step S20. When the traffic perception data meets the preset model triggering conditions, the target event recognition model corresponding to the model triggering conditions is called from the model library. The model library includes multiple event recognition models, and each event recognition model is used to identify a traffic event. In some embodiments, calling the target event recognition model corresponding to the model triggering condition from the model library includes: Step S201. Obtain the model identifier corresponding to the model triggering condition; Step S202. Call the target event recognition model that matches the model identifier from the model library.

[0036] In some embodiments, a rule base can be used to store the model triggering conditions for each traffic event recognition model. For example, for traffic events involving abnormal parking, the model triggering conditions could be: vehicle speed greater than or equal to 20 km / h, speed decreasing to 0 km / h within two seconds, continuous stopping time greater than 10 seconds, and not within a legal parking area. As another example, for traffic events involving illegal lane changes, the model triggering conditions could be: vehicle not using turn signals, lateral displacement crossing lane lines within 1.5 seconds, and distance from vehicles in front and behind the target lane less than a safety threshold.

[0037] To facilitate understanding, Table 1 below explains the triggering conditions for some traffic incident recognition models.

[0038] Table 1

[0039] Therefore, after acquiring certain traffic perception data, if it is detected that the traffic perception data meets a certain preset model trigger event, the traffic event recognition model corresponding to the model trigger event can be called to perform event recognition on the traffic perception data, thereby achieving accurate allocation and rapid response of computing resources, avoiding the running of irrelevant event recognition models, and improving resource allocation efficiency.

[0040] In some embodiments, multiple traffic event recognition models can be stored in a model library, which stores metadata corresponding to the SDK (Software Development Kit) of each traffic event recognition model.

[0041] In some embodiments, Kubernetes (K8s scheduler) can be used to register, invoke, and run multiple traffic event recognition models in parallel. The underlying principles of Kubernetes can be found in related technical documents and will not be elaborated upon here.

[0042] For example, developers upload the SDK package of a traffic event recognition model through the RESTful API (Representation Layer State Transition Application Programming Interface) of the registry center. The registry center parses its metadata (name, version, input / output interface protocols, and resource requirement configuration files, etc.) and stores it in the model library. Each traffic event recognition model can have a unique model identifier, such as a model ID. When traffic perception data meets preset model triggering conditions, the corresponding target traffic event recognition model can be called from the model library using the model identifier.

[0043] Step S30. Deploy the target event recognition model on the target node of the server cluster so that the target node runs the target event recognition model, inputs the traffic perception data into the target event recognition model, and outputs a first recognition result. The server cluster includes multiple working nodes, and each working node is used to run different event recognition models in parallel. It should be noted that the server cluster in this embodiment consists of multiple worker nodes, each of which can be an independent computing unit and a container that can host one or more traffic event recognition models. For example, through container orchestration technology, different worker nodes can run heterogeneous event recognition models in parallel without interfering with each other (such as one node handling illegal parking events and another node handling congestion events), ensuring the elastic scaling of computing resources and the high throughput of tasks.

[0044] In some embodiments, deploying the target event recognition model on the target node of the server cluster includes: Step S301. Obtain the hardware resource information required by the target event recognition model; Understandably, once it's determined that a specific target event recognition model needs to be invoked, the scheduler first queries the algorithm registry to obtain the predefined hardware resource information for that model. This information is a pre-declared list of resource specifications that guarantees optimal performance, typically existing in the form of configuration files. It mainly includes: computing resources, memory resources, and storage resources. Computing resources can include the required number of CPU cores and computing power (e.g., support for the AVX512 instruction set), whether a GPU is required, and the GPU's memory capacity and model requirements (e.g., requiring an NVIDIA T4 with at least 4GB of VRAM). Memory resources can include the minimum and recommended memory size required for operation (e.g., minimum 2GB, recommended 4GB). Storage resources can include the container image size and runtime temporary storage space requirements.

[0045] Step S302. Obtain the target node in the server cluster that matches the hardware resource information; Understandably, the scheduler continuously monitors the real-time status of all worker nodes in the server cluster, including: the total resource capacity of each node (total CPU, total memory, number of GPUs) and the currently allocated / remaining available resources. The demand list in step S301 is matched with the resource list of each node. A scheduling algorithm is used to evaluate all nodes that meet the criteria. Evaluation factors include: hard match, load balancing, affinity, etc. Finally, the scheduler selects the target node that matches the hardware resource information.

[0046] Step S303. Deploy the target event recognition model on the target node.

[0047] It should be noted that the process by which the target event recognition model obtains the first recognition result based on the traffic perception data can be referred to relevant traffic event recognition models, and will not be elaborated here.

[0048] In some embodiments, the first recognition result may be presented in the form of an image. In addition to an image, the first recognition result may also include a confidence level.

[0049] Step S40. Obtain the first identification result corresponding to each of the traffic perception data.

[0050] Understandably, in order to improve the accuracy of the traffic incident recognition model, after obtaining the first recognition result, the first recognition result is further processed by the following hybrid model recognition, which will be explained below.

[0051] In some embodiments, the first recognition result includes a result image, and after acquiring the first recognition result corresponding to each of the traffic perception data, the method further includes: Step S50. Input the result image into the first model for spatial feature extraction to obtain a spatial feature vector, wherein the first model is obtained by training the first network to be trained through the first historical samples, and the first historical samples include historical result images and corresponding spatial feature labels; Understandably, the resulting image from the first recognition output of S30 (e.g., a keyframe screenshot of a vehicle crossing a lane line in a lane-changing incident) is used as input to the first model. This image represents the region after preliminary detection and localization, focusing on the main subject of the event.

[0052] Understandably, the first model can be a deep convolutional neural network, such as ResNet, VisionTransformer, or a variant thereof. It is trained using first-historical samples, which contain a massive amount of labeled historical result images (historical result images) and spatial feature labels generated by expert systems or high-precision annotation tools. These labels not only contain event categories but can also include fine-grained attribute information (such as vehicle pose, lane line geometry, traffic sign positions, etc.), driving the network to learn to extract high-level semantic features from pixels (such as the relative positional relationship between the vehicle and lane lines, vehicle orientation). This results in a high-dimensional spatial feature vector. This vector is a mathematical representation of the image content, condensing the visual evidence of the event in the spatial dimension.

[0053] For example, the first model is a CNN network, which includes 3 convolutional layers (kernel size 3×3, stride 1) and 2 pooling layers (2×2) to extract spatial features of the resulting image (such as vehicle outlines and pedestrian poses).

[0054] Step S60. Input the traffic perception data of consecutive frames within a preset time window into the second model for temporal feature extraction to obtain a temporal feature vector. The second model is obtained by training the second network to be trained through the second historical samples. The second historical samples include traffic perception data of consecutive historical frames and corresponding temporal feature labels. Understandably, traffic perception data from multiple consecutive frames (e.g., a sequence of trajectories, speeds, or consecutive video frames containing the trajectory of a target vehicle before, during, and after an illegal lane change) is used as input to the second model. This provides process information that a single frame image cannot provide.

[0055] For example, the second model is typically a temporal neural network adept at processing sequential data, such as LSTM, GRU, or temporal convolutional networks, 3D CNNs, etc. It is trained using second historical samples, which consist of historical continuous frame data (traffic perception data of consecutive historical frames) and their corresponding temporal feature labels. The labels describe the dynamic characteristics of events, such as speed change curves, trajectory smoothness and abrupt change points, and the start and end times of actions. Through learning, the model acquires the ability to identify specific behavioral patterns from the timeline.

[0056] For example, the second model is an LSTM, which includes two layers of bidirectional LSTM (128 hidden units), processes time series data within a 3-second time window, and captures dynamic changes in events (such as vehicle acceleration trends and pedestrian movement trajectories).

[0057] Step S70. Perform feature fusion on the spatial feature vector and the temporal feature vector to obtain a fused feature vector; For example, feature fusion of the spatial feature vector and the temporal feature vector can be achieved by concatenating the spatial feature vector and the temporal feature vector, or by using an attention mechanism or a cross-modal fusion network to allow the model to automatically learn when, where, and with what weights to combine the two features. This results in a more comprehensive fused feature vector.

[0058] For example, a splicing operation is used to merge spatial features (dimension 512) and temporal features (dimension 256) to output a fused feature vector (dimension 768).

[0059] Step S80. Input the fused feature vector into the classifier and output the second recognition result, wherein the classifier is obtained by training the third network to be trained through the third historical samples, and the third historical samples include historical fused feature vectors and corresponding recognition result labels.

[0060] Understandably, the classifier itself is trained using third-party historical samples. These samples consist of historical fused feature vectors (historical fused feature vectors) and the final accurate identification result labels (such as confirming illegal lane changes). The training goal is to enable the classifier to learn to directly map the most accurate and reliable event categories and attributes from the fused features.

[0061] The classifier outputs a second identification result. Compared to the first identification result from S30, the second identification result typically has higher confidence and a lower false alarm rate, and may contain more granular information (such as event subtype, confidence score, and evolution stage). This result provides a more accurate basis for subsequent traffic warnings, signal control, or accident recording.

[0062] In some embodiments, the second identification result includes: identification confidence level and response time, wherein the response time is the time elapsed from acquiring the traffic perception data to outputting the second identification result; after outputting the second identification result, the method further includes: Step S901. Obtain the second identification result of each of the traffic perception data; Step S902. For each second identification result, the identification confidence is weighted based on a preset first weighting factor, and the response time is weighted based on a preset second weighting factor. Based on the weighted identification confidence and the weighted response time, the evaluation value of the second identification result is determined. Understandably, to further improve the efficiency of the traffic incident recognition system, the system does not simply rely on a single indicator (such as confidence level) for judgment, but comprehensively considers two key dimensions: recognition accuracy and processing timeliness (characterized by response time). Specifically, a first weighting factor and a second weighting factor are preset, and these two factors can be dynamically adjusted according to actual business needs (for example, in scenarios with extremely high security requirements, confidence level has a higher weight; in scenarios emphasizing real-time response, time has a higher weight). By weighting the confidence level and response time of each second recognition result separately, and calculating according to a preset formula (such as weighted summation or weighted scoring), a unified evaluation value is finally generated for each result.

[0063] For example: Score = 0.6×Accuracy + 0.4×(1 / ResponseTime); where Score represents the evaluation value, Accuracy represents the recognition confidence level, and ResponseTime represents the response time.

[0064] Step S903. Output the second identification result corresponding to the first n evaluation values ​​among the multiple evaluation values, where n is an integer greater than or equal to 1.

[0065] Understandably, by sorting the evaluation values ​​of all events generated in step S902 and based on preset thresholds or quantity requirements, only the top n (e.g., top 3) second identification results are output as the final result. On the one hand, this ensures the highest overall quality of the output results (e.g., avoiding outputting results with low confidence or severe delays); on the other hand, when resources are scarce or downstream processing capacity is limited, it can automatically filter out the most noteworthy and high-value events (such as serious violations that are both accurately and quickly identified), thereby optimizing resource allocation and improving the overall efficiency and targeting of traffic management response.

[0066] In some embodiments, after obtaining the second identification result of each of the traffic perception data, the method further includes: Second identification results with a confidence level less than or equal to a preset confidence level threshold will be filtered out.

[0067] For example, filtering out second recognition results with a confidence level of less than 60% can avoid outputting recognition results with low accuracy.

[0068] In some embodiments, before outputting the second recognition result corresponding to the first n evaluation values ​​among the plurality of evaluation values, the method further includes: Step S904. Obtain target recognition results from multiple second recognition results whose evaluation values ​​are greater than or equal to preset evaluation values; Step S905. If the number of target identification results is greater than n, and the target identification results include the identification results of target traffic events, then the identification results of the target traffic events are used as the second identification results to be output.

[0069] When the number of target identification results filtered in step S904 exceeds the planned output quota n, a more refined selection is required. In step S905, if this batch of target identification results includes specific types of target traffic events (such as hazardous chemical vehicles abnormally parking, pedestrians entering highways, multi-vehicle rear-end collisions, and other events with high safety levels), then regardless of whether these events' evaluation values ​​rank within the top n, the identification results for these events will be output. Therefore, for certain special event types concerning significant safety or requiring emergency response, the ranking is not entirely determined by the comprehensive score, avoiding the omission of key event identification results.

[0070] In some embodiments, the method further includes: acquiring ambient temperature; adjusting the model triggering conditions according to the ambient temperature; and / or adjusting the time window of the second model according to the ambient temperature.

[0071] Understandably, ambient temperature affects road conditions and vehicle performance, thus altering the threshold for safe driving. Based on real-time ambient temperature data, the system dynamically adjusts key parameters in the model's triggering conditions. For example, under low-temperature icy conditions, the system automatically lowers the speed change threshold for abnormal deceleration or skidding events, making the model more sensitive to potential skidding risks and enabling earlier warnings.

[0072] Understandably, ambient temperature affects the overall characteristics of traffic flow. In extreme heat or cold, driving behavior may be more cautious or more erratic, and the speed of event evolution may differ. Therefore, the time window length of the continuous frame data analyzed by the second model needs to be adjusted according to temperature. For example, during periods of high-temperature congestion, a longer analysis time window may be needed to accurately determine whether traffic is slow or completely stopped; while in low-temperature, icy road conditions, a shorter time window may be needed to quickly capture the instantaneous dynamics of emergency braking, thus improving response speed.

[0073] In the embodiments of this application, the hardware resources used in the traffic incident recognition method include, but are not limited to: a TITMS320C6678 processor (8 cores, 1.25GHz) that supports parallel computing; a SanDisk Extreme ProCFast storage array (capacity 128GB, read / write speed 525MB / s) that stores traffic perception data and the first recognition result; and a heat dissipation system that uses a heat spreader and a graphene thermal conductive film with a thermal resistance less than or equal to 0.1℃ / W.

[0074] For example, the ambient temperature is periodically collected by a temperature sensor, for example, every 100ms. When the temperature is greater than 85°C, the heat dissipation system is activated to dissipate heat. When the temperature is less than -20°C, a heating element (heating with a power of 5W) is used. When the temperature is high, the ΔV threshold of the collision warning can be widened from 5m / s² to 6m / s² to reduce the computational load. When the temperature is low, the time window of the second model is extended to 3 seconds to compensate for the sensor delay.

[0075] Figure 2 The system architecture diagram on which the traffic incident identification method of this application is based is shown.

[0076] In this embodiment of the application, the system architecture upon which the traffic incident identification method is based may include: The data acquisition layer 21 includes a data preprocessing module 211, which is used to preprocess the acquired raw sensing data to obtain multiple traffic sensing data. The scheduling layer 22 includes a threshold management module 221 and a scheduler 222. The threshold management module is used to acquire multiple traffic perception data, verify whether the traffic perception data meets the model triggering conditions, and after the verification is passed, initiate a scheduling request for the target event recognition model to the scheduler. The scheduler is used to call the target event type corresponding to the model triggering conditions from the model library according to the scheduling request, and deploy the target event recognition model on the target node of the server cluster so that the target node runs the target event recognition model and inputs the traffic perception data into the target event recognition model to output a first recognition result.

[0077] In some embodiments, the system architecture further includes: The hybrid model layer 23 includes a first model 231, a second model 232, and a classifier 233. The first model is used to output a spatial feature vector based on a first recognition result. The second model is used to output a temporal feature vector based on the traffic perception data of consecutive frames within a preset time window. The classifier is used to output a second recognition result based on the fused feature vector, wherein the fused feature vector is a vector obtained by fusing the spatial feature vector and the temporal feature vector.

[0078] Figure 3 A timing diagram of the traffic event identification method according to an embodiment of this application is shown.

[0079] Figure 3 The following example illustrates the timing of red-light violation identification. Figure 3 As shown, the traffic perception data sent by the data preprocessing module to the threshold management module includes red light status, vehicle position, and speed = 6km / h. The threshold management module verifies whether the model triggering conditions (red light + crossing the line + speed greater than 5km / h) are met. After the verification is successful, a call request (carrying algorithm (model) ID = 1001) is sent to the K8s scheduler. The K8s scheduler queries the algorithm metadata from the algorithm registry center. The algorithm registry center returns a resource request (including 2 CPU cores and 4GB of memory). The K8s scheduler schedules the algorithm container to the worker node (Node = 01). The worker node starts the SDK of the red light violation recognition model and returns the first recognition result (the intermediate result in the figure), which has a confidence level of 93%. K8s forwards the intermediate result to the hybrid model layer.

[0080] Figure 4 A structural diagram of a traffic incident recognition device according to an embodiment of this application is shown.

[0081] A second aspect of this application provides a traffic incident recognition device 300, comprising: The first acquisition unit 301 is used to acquire multiple traffic sensing data; The calling unit 302 is used to call a target event recognition model corresponding to the model triggering condition from the model library when the traffic perception data meets the preset model triggering condition. The model library includes multiple event recognition models, and each event recognition model is used to identify a traffic event. Deployment unit 303 is used to deploy the target event recognition model on the target node of the server cluster, so that the target node runs the target event recognition model, inputs the traffic perception data into the target event recognition model, and outputs a first recognition result. The server cluster includes multiple working nodes, and each working node is used to run different event recognition models in parallel. The second acquisition unit 304 is used to acquire the first identification result corresponding to each of the traffic perception data.

[0082] A third aspect of this application provides a computer-readable storage medium storing at least one computer program instruction, which is loaded and executed by a processor to perform the operations as described in any of the methods in the first aspect.

[0083] Computer-readable storage media may be portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the computer-readable storage medium of this application is not limited thereto. In this application, the readable storage medium may be any tangible medium that contains or stores a program that may be used by or in conjunction with an instruction execution system, apparatus, or device.

[0084] Readable storage media can be, for example, but not limited to, electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of readable storage media (a non-exhaustive list) include: electrical connections having one or more wires, portable disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM or flash memory), optical fibers, and portable compact disk read-only memory (CD). ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof.

[0085] Program code for performing the operations of this application can be written in any combination of one or more programming languages, including object-oriented programming languages ​​such as Java and C++, and conventional procedural programming languages ​​such as C or similar languages. The program code can execute entirely on the user's computing device, partially on the user's device, as a standalone software package, partially on the user's computing device and partially on a remote computing device, or entirely on a remote computing device or server. In cases involving remote computing devices, the remote computing device can be connected to the user's computing device via any type of network, including a local area network (LAN) or a wide area network (WAN), or it can be connected to an external computing device (e.g., via the Internet using an Internet service provider).

[0086] In the several embodiments provided in this application, it should be understood that the disclosed technical content can be implemented in other ways. The device embodiments described above are merely illustrative; for example, the division of units can be a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the displayed or discussed mutual couplings, direct couplings, or communication connections may be through some interfaces; indirect couplings or communication connections between units or modules may be electrical or other forms.

[0087] The units described as separate components may or may not be physically separate. Similarly, the components of the control device may or may not be physical units; that is, they may be located in one place or distributed across multiple units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0088] If the integrated unit is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, in essence, or the part that contributes to the prior art, or all or part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, read-only memory (ROM), random access memory (RAM), portable hard drives, magnetic disks, or optical disks.

[0089] Figure 5 A schematic diagram of the structure of a computer system suitable for implementing the electronic device of the present application is shown.

[0090] According to a fourth aspect of the present application, an electronic device is provided, including one or more processors and one or more memories, wherein at least one piece of program code is stored in the one or more memories, the at least one piece of program code being loaded and executed by the one or more processors to perform the operations performed as described in any of the methods in the first aspect.

[0091] like Figure 5 As shown, the electronic device 400 is manifested in the form of a general-purpose computing device. The components of the electronic device 400 may include, but are not limited to: at least one processing unit 410, at least one storage unit 420, and a bus 430 connecting different system components (including storage unit 420 and processing unit 410).

[0092] The storage unit stores program code, which can be executed by the processing unit 410, causing the processing unit 410 to perform the steps described in the "Embodiment Method" section above according to various exemplary embodiments of this application.

[0093] Storage unit 420 may include readable media in the form of volatile storage units, such as random access memory (RAM) 421 and / or cache 422, and may further include read-only memory (ROM) 423.

[0094] Storage unit 420 may also include a program / utility 424 having a set (at least one) of program modules 425, such program modules 425 including but not limited to: an operating system, one or more application programs, other program modules, and program data, each or some combination of these examples may include an implementation of a network environment.

[0095] Bus 430 can represent one or more of several bus structures, including a memory cell bus or memory cell controller, a peripheral bus, a graphics acceleration port, a processing unit, or a local bus using any of the multiple bus structures.

[0096] Electronic device 400 can also communicate with one or more external devices 500 (e.g., keyboard, pointing device, Bluetooth device, etc.), one or more devices that enable a user to interact with electronic device 400, and / or any device that enables electronic device 400 to communicate with one or more other computing devices (e.g., router, modem, etc.). This communication can be performed through I / O (input / output) interface 450, which can also be connected to display unit 440 to display the communication content. Furthermore, electronic device 400 can communicate with one or more networks (e.g., local area network (LAN), wide area network (WAN), and / or public networks, such as the Internet) via network adapter 460. As shown, network adapter 460 communicates with other modules of electronic device 400 via bus 430. It should be understood that, although not shown in the figures, other hardware and / or software modules can be used in conjunction with electronic device 400, including but not limited to: microcode, device drivers, redundant processing units, external disk drive arrays, RAID systems, tape drives, and data backup storage systems.

[0097] The functions described herein can be implemented in hardware, software executed by a processor, firmware, or any combination thereof. If implemented in software executed by a processor, the functions can be stored as one or more instructions or codes on or transmitted via a computer-readable medium. Other examples and embodiments are within the scope and spirit of this invention and the appended claims. For example, due to the nature of software, the functions described above can be implemented using software executed by a processor, hardware, firmware, hardwired, or any combination thereof. Furthermore, the functional units can be integrated into a single processing unit, or each unit can exist physically separately, or two or more units can be integrated into a single unit.

[0098] The above description is merely an embodiment of this application and is not intended to limit this application. Various modifications and variations can be made to this application by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the scope of the claims of this application.

Claims

1. A traffic incident detection method, characterized by, The method comprises: acquiring a plurality of traffic perception data; in a case where the traffic perception data meets a preset model triggering condition, calling a target event identification model corresponding to the model triggering condition from a model library, wherein the model library comprises a plurality of event identification models, and each event identification model is used for identifying a traffic event; deploying the target event identification model on a target node of a server cluster to enable the target node to run the target event identification model, inputting the traffic perception data into the target event identification model, and outputting a first identification result, wherein the server cluster comprises a plurality of working nodes, and each working node is used for running different event identification models in parallel; acquiring the first identification result corresponding to each traffic perception data.

2. The method of claim 1, wherein, The calling of the target event identification model corresponding to the model triggering condition from the model library comprises: acquiring a model identifier corresponding to the model triggering condition; calling a target event identification model matched with the model identifier from the model library.

3. The method of claim 1, wherein, The deploying of the target event identification model on the target node of the server cluster comprises: acquiring hardware resource information required by the target event identification model; acquiring a target node matched with the hardware resource information in the server cluster; deploying the target event identification model on the target node.

4. The method according to any of claims 1 to 3, characterized in that, The first identification result comprises a result image, and after the acquiring of the first identification result corresponding to each traffic perception data, the method further comprises: inputting the result image into a first model to perform spatial feature extraction to obtain a spatial feature vector, wherein the first model is obtained by training a first to-be-trained network through a first historical sample, and the first historical sample comprises a historical result image and a corresponding spatial feature label; inputting the traffic perception data of continuous frames within a preset time window into a second model to perform time sequence feature extraction to obtain a time sequence feature vector, wherein the second model is obtained by training a second to-be-trained network through a second historical sample, and the second historical sample comprises historical traffic perception data of continuous frames and a corresponding time sequence feature label; performing feature fusion on the spatial feature vector and the time sequence feature vector to obtain a fusion feature vector; inputting the fusion feature vector into a classifier to output a second identification result, wherein the classifier is obtained by training a third to-be-trained network through a third historical sample, and the third historical sample comprises a historical fusion feature vector and a corresponding identification result label.

5. The method of claim 4, wherein, The second identification result comprises an identification confidence and a response time, and the response time is a time experienced from the acquiring of the traffic perception data to the output of the second identification result, and after the output of the second identification result, the method further comprises: acquiring the second identification result of each traffic perception data; For each of the second recognition results, the recognition confidence is weighted based on a preset first weight factor, the response time is weighted based on a preset second weight factor, and an evaluation value of the second recognition result is determined according to the weighted recognition confidence and the weighted response time; The second recognition results corresponding to the first n evaluation values in the plurality of evaluation values are outputted, where n is an integer greater than or equal to 1.

6. The method of claim 5, wherein, After the second recognition result of each of the traffic perception data is obtained, the method further comprises: Second recognition results with a recognition confidence less than or equal to a preset confidence threshold are filtered out.

7. The method of claim 5, wherein, Before the second recognition results corresponding to the first n evaluation values in the plurality of evaluation values are outputted, the method further comprises: A target recognition result with an evaluation value greater than or equal to a preset evaluation value is obtained from the plurality of second recognition results. If the number of the target recognition results is greater than n and the target recognition results include a recognition result of a target traffic event, the recognition result of the target traffic event is taken as a second recognition result to be outputted.

8. The method of claim 4, wherein, Further comprising: An ambient temperature is obtained; The model triggering condition is adjusted according to the ambient temperature, and / or the time window of the second model is adjusted according to the ambient temperature.

9. A traffic incident detection apparatus characterized by comprising: Comprise: A first obtaining unit is configured to obtain a plurality of traffic perception data; A calling unit is configured to call a target event recognition model corresponding to a preset model triggering condition from a model library if the traffic perception data meets the model triggering condition, where the model library comprises a plurality of event recognition models, and each of the event recognition models is configured to recognize a traffic event; A deployment unit is configured to deploy the target event recognition model on a target node of a server cluster, so that the target node runs the target event recognition model, and input the traffic perception data into the target event recognition model to output a first recognition result, where the server cluster comprises a plurality of working nodes, and each of the working nodes is configured to run different event recognition models in parallel; A second obtaining unit is configured to obtain the first recognition result corresponding to each of the traffic perception data.

10. A computer-readable storage medium, characterized in that, The computer readable storage medium stores at least one computer program instruction, and the at least one computer program instruction is loaded and executed by the processor to implement the operations performed by the method according to any one of claims 1-8.

Images