Driving environment identification method and vehicle-mounted terminal

Abstract

The invention provides a driving environment identification method and a vehicle-mounted terminal. The driving environment identification method comprises the following steps: calling a target API based on position information of a target vehicle to obtain a corresponding target data set; wherein the target data set at least comprises a road scene data set and an administrative region data set; identifying a road scene type based on the road scene data set, and obtaining administrative region attributes based on the administrative region data set; and outputting a driving environment identification result in combination with the road scene type and the administrative region attribute. The method does not depend on a single sensor to sense the environment, and can stably obtain the data for identification under different environment conditions, thereby improving the adaptability to the environment, and reducing the possibility of misjudgment. By calling the target API to obtain the data set, the latest conditions of road scenes and administrative region attributes can be reflected in time, the problem of vehicle positioning matching deviation caused by untimely updating of the basic map is avoided, and then the accuracy of scene recognition is improved.

Description

Technical Field

[0001] This application relates to the field of driving environment perception technology, and in particular to a driving environment recognition method and vehicle terminal. Background Technology

[0002] Driving environment recognition is the core central link for autonomous driving systems to achieve environmental understanding and safety decisions; the current mainstream recognition schemes are mainly divided into two categories: one is sensor-driven, and the other is based on map assistance.

[0003] However, sensor-based solutions are less adaptable to the environment, are easily affected by weather and lighting conditions, and are prone to misjudgment; base maps, on the other hand, usually have long update cycles, making vehicle positioning and matching prone to errors, and scene recognition accuracy is also low. Summary of the Invention

[0004] This application provides a driving environment recognition method and an in-vehicle terminal to solve the above-mentioned technical problems.

[0005] Specifically, this application provides a driving environment recognition method, comprising the following steps: calling a target API based on the location information of a target vehicle to obtain a corresponding target dataset; wherein the target dataset includes at least a road scene dataset and an administrative region dataset; identifying the road scene type based on the road scene dataset and obtaining the administrative region attribute based on the administrative region dataset; and outputting a driving environment recognition result by combining the road scene type and the administrative region attribute.

[0006] In the above technical solution, the system does not rely on a single sensor to perceive the environment, thus avoiding interference from external environmental factors such as weather (e.g., heavy rain, fog, etc.) affecting sensor accuracy and light (e.g., strong light, darkness, etc., making it difficult for sensors to detect accurately) on the recognition results. It can stably acquire data for recognition under different environmental conditions, thereby improving the adaptability to the environment and reducing the possibility of misjudgment. By calling the target API to obtain the dataset, the road scene dataset and administrative region dataset can be updated in real time or near real time, which can reflect the latest situation of road scene and administrative region attributes in a timely manner. This avoids the vehicle positioning and matching deviation problem caused by the untimely update of the base map, thereby improving the accuracy of scene recognition.

[0007] Furthermore, the road scene dataset includes at least a terrain dataset, a navigation dataset, and a map dataset. The terrain dataset includes several terrain data sequences, each of which includes at least a terrain name, terrain type, location coordinates, and a first confidence level. Identifying the road scene type based on the road scene dataset includes: constructing a terrain keyword library based on the terrain name and terrain type, and calculating the semantic matching degree of each terrain data sequence based on the terrain keyword library and the confidence level; calculating the spatial circularity based on the location coordinates, and combining the semantic matching degree and the spatial circularity to calculate a comprehensive confidence level for terrain identification; and obtaining a terrain identification result based on the semantic matching degree and the spatial circularity, or based on the semantic matching degree and the comprehensive confidence level for terrain identification.

[0008] In the above technical solutions, traditional sensor solutions are affected by obstruction and bad weather, resulting in low accuracy in identifying special terrains (such as roundabouts). This application can accurately determine whether it is a roundabout from a semantic level through keyword matching, and spatial circularity verification further confirms it from spatial geometric features.

[0009] Furthermore, a scene keyword library is pre-built, and the navigation dataset includes at least road name, road type, real-time navigation command, and next command information; the identification of road scene type based on the road scene dataset also includes: obtaining a preliminary scene type based on the scene keyword library, road name, road type, and real-time navigation command; and obtaining a target scene type based on the target vehicle's location information, real-time speed, preliminary scene type, and next command information.

[0010] In the above technical solution, real-time scene recognition can be achieved based on vehicle status and navigation information, such as highway scenes, ramp scenes, and slope scenes, so as to provide more accurate information for subsequent driving decisions.

[0011] Furthermore, the map dataset includes at least predicted coordinates and map images; the identification of road scene types based on the road scene dataset further includes: extracting road edges and lane line edges based on the map images, and extracting straight line segments based on the road edges and lane line edges to obtain the number of intersecting roads and intersecting straight line segments; obtaining road angles based on the intersecting straight line segments and the predicted coordinates; obtaining the number of lane line pixels and the number of road area pixels based on the map images to obtain lane line distribution density; obtaining map feature vectors based on the number of intersecting roads, road angles, and lane line distribution density, and obtaining visual feature vectors based on the collected visual information, so as to obtain the intersection type probability distribution based on the map feature vectors and visual feature vectors; and obtaining the target road type based on the road type probability distribution and a second confidence level; wherein the second confidence level is calculated based on map feature confidence level, visual feature confidence level, map feature weight, and visual feature weight.

[0012] In the above-mentioned technical solutions, traditional vision solutions have a high misjudgment rate at intersections in complex environments. This application uses map features to extract key road information and combines it with visual language large model reasoning. Even in rainy days, nighttime backlight and other scenarios, the accuracy of intersection type recognition remains above 90%, with a relative robustness improvement of 50%, effectively avoiding the risk of turning conflicts caused by intersection misjudgment.

[0013] Furthermore, the step of identifying road scene types based on the road scene dataset also includes: determining whether there is a conflict based on the terrain recognition result, the target scene type, and the target road type; if there is, performing a verification process and correcting the road scene type based on the verification result; otherwise, combining the terrain recognition result, the target scene type, and the target road type to obtain the road scene type; wherein, the verification process includes at least road scene dataset validity verification, historical trend deviation verification, and cross-verification of recognition results.

[0014] The above technical solution solves the problem of single module identification deviation or conflict, so as to output a unique and reliable environment type based on the verification results.

[0015] Furthermore, obtaining administrative region attributes based on the administrative region dataset includes: obtaining administrative level information based on the administrative region dataset and counting the number of various regional features; obtaining administrative attribute coefficients based on the administrative level information and obtaining regional feature density based on the number of regional features; assigning regional weights to each regional feature based on the administrative attribute coefficients; and obtaining the administrative region affiliation probability based on the regional weight allocation result and the regional feature density, so as to obtain administrative region attributes based on the administrative region affiliation probability.

[0016] In the above technical solutions, the traditional division scheme based on administrative boundaries has a low accuracy rate in identifying urban-rural fringe areas and urban villages. This application takes into account the characteristics of different administrative regions by using dynamic weights of administrative attribute coefficients, and the density layer calculation is combined with the actual construction situation of the region. Combining the two, the comprehensive identification accuracy of cities, villages and urban villages reaches more than 95%, which is 35% higher than the traditional scheme, and can accurately adapt to driving strategies in different regions.

[0017] Furthermore, the process of calling the target API includes: executing a reconnection mechanism when the call fails; if the call fails for a preset number of consecutive times, obtaining the corresponding cached data as the target dataset.

[0018] In the above technical solution, the reconnection mechanism improves the success rate of API calls and ensures the normal acquisition of data; when a preset number of consecutive calls fail, cached data is obtained as the target dataset, ensuring that the system can continue to work in the event of API call anomalies, thereby improving the stability and reliability of the system.

[0019] Furthermore, it also includes: adjusting the map feature weights, visual feature weights, and / or region weights according to the real-time scene status.

[0020] In the above technical solutions, different real-time scenarios have different degrees of dependence on map features, visual features, and regional features. By dynamically adjusting the weights according to the real-time scenario status, the system can adapt to different scenarios more flexibly, further improving the accuracy and adaptability of recognition.

[0021] Based on the same concept, this application also provides an in-vehicle terminal, which includes at least a memory and a processor; the memory is used to store computer instructions for multiple functional layers, the functional layers including at least a first acquisition functional layer, a second acquisition functional layer and an output functional layer; the processor communicates with the memory via a bus and is used to execute the computer instructions for each of the multiple functional layers stored in the memory.

[0022] Furthermore, the first acquisition function layer is used to call the target API based on the location information of the target vehicle to obtain the corresponding target dataset; wherein, the target dataset includes at least a road scene dataset and an administrative region dataset; the second acquisition function layer is used to identify the road scene type based on the road scene dataset and obtain the administrative region attribute based on the administrative region dataset; and the output function layer is used to output the driving environment recognition result by combining the road scene type and the administrative region attribute.

[0023] The above technical solution breaks through the bottleneck of traditional sensors by fusing multi-source data, thereby achieving a qualitative improvement in recognition accuracy; it can also resist interference in extreme scenarios, thus comprehensively enhancing environmental adaptability.

[0024] Compared with the prior art, the beneficial effects of this application are as follows: This application does not rely on a single sensor to perceive the environment. It can stably acquire data for identification under different environmental conditions, thereby improving the adaptability to the environment and reducing the possibility of misjudgment. By calling the target API to obtain the dataset, it can reflect the latest situation of road scene and administrative region attributes in a timely manner, avoiding the vehicle positioning and matching deviation problem caused by the untimely update of the base map, and thus improving the accuracy of scene recognition. Attached Figure Description

[0025] Figure 1 This is a flowchart of the driving environment recognition method described in this application.

[0026] Figure 2 This is a framework diagram of the vehicle-mounted terminal described in this application.

[0027] Figure 3 This is a schematic diagram of the memory described in this application. Detailed Implementation

[0028] The following describes in further detail a driving environment recognition method and vehicle terminal of this application with reference to specific embodiments and accompanying drawings.

[0029] Please see Figure 1 This application provides a driving environment recognition method, including the following steps S100-S300.

[0030] In one feasible implementation, the system uses the vehicle's current location information as the center and a dynamic search radius to call the POI search API to obtain a terrain dataset. It then combines semantic feature extraction and spatial feature extraction results to obtain semantic matching degree and spatial circularity, and calculates the overall confidence score for terrain recognition to determine whether the current terrain is a roundabout. Simultaneously, it calls the path planning API to obtain a navigation dataset, and obtains the target scene type (e.g., ramp) through keyword quantization matching and temporal verification. Furthermore, it integrates prior features from the static map API and a large visual language model, obtaining the target road type (e.g., intersection) through feature quantization and weighted fusion. Additionally, it utilizes a hierarchical fusion of the administrative region query API and POI classification statistics (dynamic weight allocation + density score model) to construct a differentiated model for cities, rural areas, and urban villages, accurately capturing non-standard regional attributes such as the "half-city, half-rural" nature of urban-rural fringe areas and the "high-density mixed features" of urban villages. Finally, it combines the above recognition results to output a structured scene result, such as "urban area + roundabout + intersection, overall confidence score 0.91".

[0031] Step S100: Call the target API based on the location information of the target vehicle to obtain the corresponding target dataset; wherein, the target dataset includes at least a road scene dataset and an administrative region dataset.

[0032] In some embodiments, such as constructing a secure and controllable Gaode API vehicle access system, after applying for the API key, the key is stored in an in-vehicle security chip to avoid plaintext storage leading to leakage; in addition, a permission whitelist is configured on the map developer platform, only granting access permissions to the vehicle terminal device ID, and setting call frequency thresholds according to interface type to prevent interface abuse or abnormal calls. The road scene dataset includes at least a terrain dataset, a navigation dataset, and a map dataset; the target API is, for example, a POI search API, a route planning API, a static map API, and an administrative region query API; the target API is preferentially called using the HTTPS protocol to ensure encrypted data transmission, and a cross-platform interface adaptation layer is also configured, such as reducing TCP handshake time through long socket connections, and controlling the network latency of a single interface call to within 200ms in a 4G environment.

[0033] The process of calling the target API includes: when the call fails, executing a reconnection mechanism; if the call fails for a preset number of consecutive times, obtaining the corresponding cached data as the target dataset.

[0034] In some embodiments, for scenarios with fluctuating vehicle networks, an interface timeout retry strategy is set (e.g., ≤3 retries, incremented by 100ms interval); if an interface fails to be called 3 times consecutively (e.g., no network in the tunnel), a degradation scheme is automatically triggered, such as switching the terrain dataset to local cached data (e.g., surrounding POIs within the last hour) and using preloaded offline tiles for the map dataset to ensure that the core scene recognition is not interrupted.

[0035] In the above technical solution, the reconnection mechanism improves the success rate of API calls and ensures the normal acquisition of data; when a preset number of consecutive calls fail, cached data is obtained as the target dataset, ensuring that the system can continue to work in the event of API call anomalies, thereby improving the stability and reliability of the system.

[0036] Step S200: Identify the road scene type based on the road scene dataset, and obtain the administrative region attribute based on the administrative region dataset.

[0037] The terrain dataset includes several terrain data sequences, each of which includes at least a terrain name, terrain type, location coordinates, and a first confidence level. Identifying road scene types based on the road scene dataset includes: constructing a terrain keyword library based on the terrain name and terrain type, and calculating the semantic matching degree of each terrain data sequence based on the terrain keyword library and the confidence level; calculating spatial circularity based on the location coordinates, and combining the semantic matching degree and the spatial circularity to calculate a comprehensive confidence level for terrain identification; and obtaining a terrain identification result based on the semantic matching degree and the spatial circularity, or based on the semantic matching degree and the comprehensive confidence level for terrain identification.

[0038] In some embodiments, the vehicle's current location information ( Latitude Set the dynamic search radius centered on longitude. (Based on the common roundabout diameter of 15-40 meters, take...) The POI search API is invoked (parameters: location=Lat_0, Lon_0, radius=r, types=190300 (road ancillary facilities class)) to obtain the surrounding POI dataset (i.e., the terrain dataset). ,in For the number of POIs, each (That is, the terrain data sequence) includes terrain names Terrain type Location coordinates First confidence level Field.

[0039] Further analysis and Construct a terrain keyword library, such as K = {"roundabout", "circular island", "roundabout", "circular intersection"}, and calculate each... semantic matching degree : ; Fuzzy matching is determined by the string edit distance; a match is considered to occur when the edit distance is ≤2.

[0040] Furthermore, to verify the ring-shaped distribution characteristics of the terrain, all data were first fitted using the least squares method. The center of the circular trajectory This can be achieved by fitting only 6 key POIs, reducing processing time per frame and computational cost; then each... Euclidean distance to the center of the circle (Calculated after converting latitude and longitude to plane coordinates): ; in, for Mercator projection plane coordinates, To fit the planar coordinates of the circle's center; define the spatial cyclicity. This is the ratio of the standard deviation to the mean distance: ; in, Standard deviation, ; The closer the value is to 1, the closer the spatial distribution of POIs is to a ring (threshold set to 1). ).

[0041] Construct a semantic-spatial dual-dimensional comprehensive judgment model to calculate the comprehensive confidence score for terrain recognition. : ; in, For semantic feature weights, Spatial feature weights (which can be dynamically adjusted according to the environment, such as increasing when the sensor is blocked). ).

[0042] Set threshold judgment rules: If it exists satisfy and ,or ( For an indicator function, if the condition is met, then it is 1) and If the terrain is roundabout, it is determined to be roundabout terrain; otherwise, it is determined to be non-roundabout terrain (such as straight roads or ordinary intersections).

[0043] In the above technical solutions, traditional sensor solutions are affected by occlusion and bad weather, resulting in low accuracy in identifying special terrains (such as roundabouts). This application can accurately determine whether it is a roundabout from a semantic level through keyword matching, and spatial circularity verification further confirms it from spatial geometric features. Subsequently, by combining the road topology features of static maps, the accuracy of roundabout identification is steadily improved to over 90%, with a relative improvement of 40%, completely solving the problem of misjudgment in complex terrain.

[0044] Furthermore, a scene keyword library is pre-built, and the navigation dataset includes at least road name, road type, real-time navigation command, and next command information; the identification of road scene type based on the road scene dataset also includes: obtaining a preliminary scene type based on the scene keyword library, road name, road type, and real-time navigation command; and obtaining a target scene type based on the target vehicle's location information, real-time speed, preliminary scene type, and next command information.

[0045] In some embodiments, the route planning API is invoked to obtain the navigation dataset for the current driving segment. ,in, For road names (e.g., "G60 Shanghai-Kunming Expressway"), This refers to the road type (e.g., "110200" represents a highway). For real-time navigation instructions (such as "Enter XX ramp 500 meters ahead"), The remaining distance to the next instruction point (i.e., the next instruction information, in meters).

[0046] This includes a pre-built scenario keyword library. For example, for scenarios such as highways, ramps, elevated roads, and slopes, the keyword sets are defined as follows: High-speed scenarios {"Highway", "Expressway", "G", "National Expressway", "Toll Station", "Service Area"}; Ramp Scene {"Ramp", "Entrance", "Exit", "Turn", "Decelerate to", "Merge in"}; Elevated road scene {"Elevated road", "Expressway", "Bridge", "Elevated road exit", "Continuous elevated road"}; Ramp scene {"Uphill", "Downhill", "Slope", "Continuous Slope", "Gentle Slope", "Steep Slope"}.

[0047] Furthermore, a dynamic sliding window with a length of, for example, L=500m is used (and... (Matching), calculate the matching frequency of keywords for each scene within the window. ( (Representing four types of scenarios): ; in, The navigation text within the sliding window (i.e., the road name, road type, and real-time navigation instructions). This represents the total number of words in the text. For indicator functions (keywords) Existence Then it is 1); set the frequency threshold. ,when At that time, the initial scene type is k.

[0048] Furthermore, calculate the remaining distance for navigation instructions. Actual distance from vehicle deviation : ; in, Calculated using vehicle odometer or location information (unit: m); when If the navigation command matches the vehicle's location, the verification is successful.

[0049] Then combine the real-time vehicle speed from the vehicle's CAN bus. (Unit: km / h) Calculate the deviation from the standard vehicle speed in the scenario. .

[0050] For example, in high-speed scenarios, the standard vehicle speed (Minimum speed limit on highways) , must meet .

[0051] Furthermore, in a certain scenario satisfy" "and" "and" When this is determined, the target scene type is ultimately identified as a scene. The response time is controlled as follows: ;in, The time spent on keyword matching To reduce the time consumption for timing verification, a millisecond-level response time is achieved.

[0052] In the above technical solution, real-time scene recognition can be achieved based on vehicle status and navigation information, such as highway scenes, ramp scenes, and slope scenes, so as to provide more accurate information for subsequent driving decisions.

[0053] Furthermore, the map dataset includes at least predicted coordinates and map images; the identification of road scene types based on the road scene dataset further includes: extracting road edges and lane line edges based on the map images, and extracting straight line segments based on the road edges and lane line edges to obtain the number of intersecting roads and intersecting straight line segments; obtaining road angles based on the intersecting straight line segments and the predicted coordinates; obtaining the number of lane line pixels and the number of road area pixels based on the map images to obtain lane line distribution density; obtaining map feature vectors based on the number of intersecting roads, road angles, and lane line distribution density, and obtaining visual feature vectors based on the collected visual information, so as to obtain the intersection type probability distribution based on the map feature vectors and visual feature vectors; and obtaining the target road type based on the road type probability distribution and a second confidence level; wherein the second confidence level is calculated based on map feature confidence level, visual feature confidence level, map feature weight, and visual feature weight.

[0054] In some embodiments, the static map API is invoked based on pre-set parameters (such as center coordinates, zoom level, image size, and layer control) to obtain predicted coordinates. and map images .

[0055] It should be noted that the center coordinates, i.e. the predicted coordinates of the vehicle in front of it at d=50~100m, are calculated from the current positioning and heading angle. The zoom level is such as zoom=18, the image size is such as 600×600 pixels, and the layer control is such as keeping only the road layer (layers=road|on) and the lane line layer (layers=lane|on).

[0056] Furthermore, regarding Perform grayscale processing, for example, using the Sobel operator to calculate the horizontal and vertical gradients, and extract the edges of the road and lane lines: ; ; in, For convolution operations, This is the grayscale map image. The total gradient (gradient value) (Pixels are identified as edges).

[0057] Furthermore, for example, by extracting line segments from the edge using Hough Line Transform, the number of intersection points of the line segments is calculated. When there are ≥3 intersection points and the angle difference between the line segments is ≥30°, it is determined to be... When four roads intersect (e.g., four straight road segments intersect with an angle approximately 90°), then... For intersecting straight segments, the included angle of the roads is calculated using the law of cosines: ; in, The length of the edges of two adjacent roads. The length of the line segment opposite the included angle (in pixels), based on the predicted coordinates. Obtained through calculation.

[0058] Furthermore, the number of lane line pixels in the statistical image was analyzed. Number of pixels in the road area The ratio of the two values ​​is used to obtain the lane line distribution density. .

[0059] Furthermore, map features Convert to a text description vector (i.e., the map feature vector). (e.g., "4 roads intersect at a 90° angle, with a lane density of 0.3"), while simultaneously displaying real-time intersection images (i.e., the aforementioned visual information) captured by the vehicle-mounted camera. Convert to visual feature vector Then and Input a pre-trained large visual language model (such as qwen), and normalize the output intersection type probability distribution using a function such as Softmax: ; ; Where T = {"crossroads", "T-junction", "no intersection"} is the set of target types. Calculate the feature similarity score for the model.

[0060] Furthermore, combining map feature confidence levels confidence level of visual features Calculate the second confidence level : ; in, (Map feature weights, unaffected by environmental interference) (Visual feature weights) The confidence level for intersection recognition based on camera images (dynamically adjusted by image sharpness, such as in rainy weather) (decreased to 0.2); when When the intersection type is determined, the target road type is obtained.

[0061] In the above-mentioned technical solutions, traditional vision solutions have a high misjudgment rate at intersections in complex environments. This application uses map features to extract key road information and combines it with visual language large model reasoning. Even in rainy days, nighttime backlight and other scenarios, the accuracy of intersection type recognition remains above 90%, with a relative robustness improvement of 50%, effectively avoiding the risk of turning conflicts caused by intersection misjudgment.

[0062] Furthermore, the step of identifying road scene types based on the road scene dataset also includes: determining whether there is a conflict based on the terrain recognition result, the target scene type, and the target road type; if there is, performing a verification process and correcting the road scene type based on the verification result; otherwise, combining the terrain recognition result, the target scene type, and the target road type to obtain the road scene type; wherein, the verification process includes at least road scene dataset validity verification, historical trend deviation verification, and cross-verification of recognition results.

[0063] In some embodiments, when the identification results of different modules conflict (e.g., the obtained target scene type is "high-speed (0.85)" but the terrain identification result does not contain any "high-speed" related keywords), a three-level verification mechanism is triggered: Road scene dataset validity verification: Check the validity of the module input data (e.g., for highway scenarios, it is necessary to verify that the vehicle speed is ≥60km / h; if the vehicle speed is only 40km / h, the confidence of the module should be reduced to 0.5).

[0064] Historical trend deviation verification: For example, if the current result deviates from the historical trend by more than 50% from the recognition results of the last 5 frames (e.g., the previous 4 frames were all "urban roads", and the current result suddenly becomes "highway"), it is marked as a suspicious result.

[0065] Cross-validation of recognition results: Validate the results using third-party modules (e.g., if the target scene type is "ramp", the target road type must have the feature of "circular lane line", otherwise reduce the confidence level).

[0066] The above technical solution solves the problem of single module identification deviation or conflict, so as to output a unique and reliable environment type based on the verification results.

[0067] Furthermore, obtaining administrative region attributes based on the administrative region dataset includes: obtaining administrative level information based on the administrative region dataset and counting the number of various regional features; obtaining administrative attribute coefficients based on the administrative level information and obtaining regional feature density based on the number of regional features; assigning regional weights to each regional feature based on the administrative attribute coefficients; and obtaining the administrative region affiliation probability based on the regional weight allocation result and the regional feature density, so as to obtain administrative region attributes based on the administrative region affiliation probability.

[0068] In some embodiments, the administrative region query API is invoked to obtain the vehicle's current location information. Information on the administrative level to which it belongs ( It is a prefecture-level city. For districts and counties, (For townships), define administrative attribute coefficients. (Weights used to dynamically adjust the feature density of subsequent regions): ; Furthermore, with Statistical radius centered on Within the range (area) The number of three types of regional features: Number of urban features (e.g., type codes: 090100-office building, 080100-shopping mall, 150500-subway station); number of rural features (e.g., type codes: 110100-farmland, 130200-village committee, 190100-rural road); number of urban village features (e.g., type codes: 070200-self-built houses, 080900-small shops, 070100-dense residential areas).

[0069] Define three types of regional feature densities (unit: , ): ; Furthermore, the weights of the three types of regional feature densities are assigned based on administrative attribute coefficients. : ; ; .

[0070] in, Ensure weight normalization; The larger (the closer to the city) The larger, the smaller The larger, Maximum performance in urban-rural fringe areas (suitable for urban village scenarios).

[0071] Furthermore, construct the probability of administrative region affiliation. , , For example, when a certain type of probability When, it is determined to be the corresponding administrative region attribute (such as...) (This refers to urban areas).

[0072] In the above technical solutions, the traditional division scheme based on administrative boundaries has a low accuracy rate in identifying urban-rural fringe areas and urban villages. This application takes into account the characteristics of different administrative regions by using dynamic weights of administrative attribute coefficients, and the density layer calculation is combined with the actual construction situation of the region. Combining the two, the comprehensive identification accuracy of cities, villages and urban villages reaches more than 95%, which is 35% higher than the traditional scheme, and can accurately adapt to driving strategies in different regions.

[0073] Furthermore, it also includes: adjusting the map feature weights, visual feature weights, and / or region weights according to the real-time scene status.

[0074] In some embodiments, the module weights are dynamically adjusted according to the real-time scenario. For example, in a rainy scenario, the map feature weight is increased from 0.6 to 0.8 (visual features are affected by rain); in a suburban scenario, the regional weight is increased from 0.25 to 0.35 (the attributes of urban villages need to be distinguished first), to ensure that the weight allocation matches the scenario requirements.

[0075] In the above technical solutions, different real-time scenarios have different degrees of dependence on map features, visual features, and regional features. By dynamically adjusting the weights according to the real-time scenario status, the system can adapt to different scenarios more flexibly, further improving the accuracy and adaptability of recognition.

[0076] Step S300: Output the driving environment recognition result by combining the road scene type and administrative region attribute.

[0077] In some embodiments, the fused output structured scene results (including "road scene type, administrative region attribute + comprehensive confidence + contribution of each module", such as "urban area + roundabout + intersection, comprehensive confidence 0.91") are transmitted to the intelligent driving decision system via CAN bus (response latency ≤ 50ms); at the same time, the raw data of each frame of the fusion process (module output, weight, conflict handling result) are recorded to the vehicle SD card for subsequent testing, optimization and fault tracing.

[0078] Furthermore, it should be noted that when location coordinates are missing in the topographic data sequence, interpolation of adjacent locations is used to complete them; when administrative region level information is empty, the name of the superior administrative region is filled in by default (e.g., the name of the district / county is used instead when the township field is missing); and when acquiring the target dataset, data exceeding a reasonable range is removed based on actual application needs, and deduplication is performed using a sliding window; data standardization transformation is also involved, standardizing the naming and type of each field, i.e., unifying the map coordinate system to the coordinate system adapted to the vehicle system; in addition, normalization is required when extracting features to ensure a uniform input scale for subsequent algorithms. The above preprocessing processes are performed in parallel, and the preprocessed data is stored in the vehicle's shared memory for retrieval, reducing data copying time.

[0079] In summary, this application does not rely on a single sensor to perceive the environment, avoiding interference from external environmental factors such as weather (e.g., heavy rain, fog, etc.) affecting sensor accuracy and light (e.g., strong light, darkness, etc., making accurate sensor detection difficult). It can stably acquire data for recognition under different environmental conditions, thereby improving environmental adaptability and reducing the possibility of misjudgment. By calling the target API to obtain the dataset, the road scene dataset and administrative region dataset can be updated in real time or near real time, which can promptly reflect the latest status of road scene and administrative region attributes, avoiding vehicle positioning and matching deviations caused by untimely updates of the base map, thereby improving the accuracy of scene recognition.

[0080] Based on the same concept, please refer to Figure 2 This application also provides an in-vehicle terminal, comprising at least a memory and a processor; the memory is used to store computer instructions for multiple functional layers (such as...). Figure 3 The processor (as shown) includes at least a first acquisition function layer, a second acquisition function layer, and an output function layer; the processor communicates with the memory via a bus and is used to execute computer instructions for each of the multiple function layers stored in the memory.

[0081] Furthermore, the first acquisition function layer is used to call the target API based on the location information of the target vehicle to obtain the corresponding target dataset; wherein, the target dataset includes at least a road scene dataset and an administrative region dataset; the second acquisition function layer is used to identify the road scene type based on the road scene dataset and obtain the administrative region attribute based on the administrative region dataset; and the output function layer is used to output the driving environment recognition result by combining the road scene type and the administrative region attribute.

[0082] It should be noted that each of the above functional layers implements its corresponding function using the specific implementation steps of the above driving environment recognition method, which will not be elaborated here. Among them, the vehicle real-time dispatcher allocates CPU resources according to module priority, with road scene type recognition having the highest priority and administrative region attribute recognition having the second highest priority, to ensure that the recognition of key scenes such as highway ramps and intersections is not delayed.

[0083] The above technical solution breaks through the bottleneck of traditional sensors by fusing multi-source data, thereby achieving a qualitative improvement in recognition accuracy; it can also resist interference in extreme scenarios, thus comprehensively enhancing environmental adaptability.

[0084] Although exemplary embodiments have been described herein with reference to the accompanying drawings, it should be understood that the above exemplary embodiments are merely illustrative and are not intended to limit the scope of this application. Various changes and modifications can be made therein by those skilled in the art without departing from the scope and spirit of this application. All such changes and modifications are intended to be included within the scope of this application as claimed in the appended claims.

[0085] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0086] In the several embodiments provided in this application, it should be understood that the disclosed devices and methods can be implemented in other ways. For example, the device embodiments described above are merely illustrative. For instance, the division of units is only 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 device, or some features may be ignored or not executed.

[0087] The various component embodiments of this application can be implemented in hardware, or as software modules running on one or more processors, or a combination thereof. Those skilled in the art will understand that microprocessors or digital signal processors (DSPs) can be used in practice to implement some or all of the functions of some modules according to the embodiments of this application. This application can also be implemented as an apparatus program (e.g., a computer program and computer program product) for performing part or all of the methods described herein. Such an implementation of this application can be stored on a computer-readable medium, or can be in the form of one or more signals. Such signals can be downloaded from an Internet website, provided on a carrier signal, or provided in any other form.

[0088] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0089] Although the description of this application has been made in conjunction with the specific embodiments described above, it will be apparent to those skilled in the art that many substitutions, modifications, and variations can be made based on the foregoing. Therefore, all such substitutions, modifications, and variations are included within the spirit and scope of the appended claims.

Claims

1. A driving environment recognition method characterized by comprising: The method comprises the following steps: calling a target API based on position information of the target vehicle to obtain a corresponding target data set; wherein the target data set at least comprises a road scene data set and an administrative region data set; identifying a road scene type based on the road scene data set, and obtaining an administrative region attribute based on the administrative region data set; and outputting a driving environment recognition result in combination with the road scene type and the administrative region attribute.

2. The driving environment recognition method according to claim 1, characterized by, The road scene data set at least comprises a terrain data set, a navigation data set and a map data set, the terrain data set comprises a plurality of terrain data sequences, each terrain data sequence at least comprises a terrain name, a terrain type, a position coordinate and a first confidence degree; the identification of the road scene type based on the road scene data set comprises: constructing a terrain keyword library based on the terrain name and the terrain type, and calculating a semantic matching degree of each terrain data sequence according to the terrain keyword library and the confidence degree; calculating a spatial ring degree according to the position coordinate, and calculating a terrain recognition comprehensive confidence degree in combination with the semantic matching degree and the spatial ring degree; and obtaining a terrain recognition result based on the semantic matching degree and the spatial ring degree, or based on the semantic matching degree and the terrain recognition comprehensive confidence degree.

3. The driving environment recognition method according to claim 2, characterized by, The scene keyword library is constructed in advance, the navigation data set at least comprises a road name, a road type, a real-time navigation instruction and next instruction information; the identification of the road scene type based on the road scene data set further comprises: obtaining a preliminary scene type based on the scene keyword library, the road name, the road type and the real-time navigation instruction; and obtaining a target scene type based on the position information of the target vehicle, a real-time vehicle speed, the preliminary scene type and the next instruction information.

4. The driving environment recognition method according to claim 2, characterized by, The map data set at least comprises a prediction coordinate and a map image; the identification of the road scene type based on the road scene data set further comprises: extracting a road edge and a lane line edge based on the map image, and extracting a straight line segment according to the road edge and the lane line edge to obtain a number of intersecting roads and intersecting straight line segments; obtaining a road angle according to the prediction coordinate based on the intersecting straight line segments; obtaining a lane line distribution density according to a number of lane line pixels and a number of road region pixels obtained from the map image; obtaining a map feature vector based on the number of intersecting roads, the road angle and the lane line distribution density, and obtaining a visual feature vector based on collected visual information, to obtain a road type probability distribution according to the map feature vector and the visual feature vector; and obtaining a target road type according to the road type probability distribution and a second confidence degree; wherein the second confidence degree is obtained based on a map feature confidence degree, a visual feature confidence degree, a map feature weight and a visual feature weight.

5. The driving environment recognition method according to any one of claims 3-4, characterized by, The identification of the road scene type based on the road scene data set further comprises: If there is a conflict, verification is performed, and the road scene type is corrected based on the verification result; otherwise, the road scene type is obtained in combination with the terrain recognition result, target scene type, and target road type. The verification work includes at least road scene dataset validity verification, historical trend deviation verification, and recognition result cross verification.

6. The driving environment recognition method according to claim 1, characterized by, The administrative region attribute is obtained based on the administrative region dataset, including: The administrative region attribute is obtained based on the administrative region dataset, including: The administrative region attribute is obtained based on the administrative region dataset, including: The administrative region attribute is obtained based on the administrative region dataset, including:

7. The driving environment recognition method according to claim 1, characterized by, The administrative region attribute is obtained based on the administrative region dataset, including: In the process of calling the target API, including:

8. The driving environment recognition method according to any one of claims 4 or 6, characterized by, When the call fails, a reconnection mechanism is executed; if the call fails for a preset number of consecutive times, the corresponding cache data is obtained as the target dataset. Also including:

9. A vehicle terminal, characterized by Adjusting the map feature weight, visual feature weight, and / or region weight according to the real-time scene state. At least including a memory and a processor; The memory is used to store computer instructions of multiple function layers, including at least a first acquisition function layer, a second acquisition function layer, and an output function layer; 10. The vehicle terminal according to claim 9, characterized by The processor communicates with the memory through a bus and is used to execute computer instructions of each function layer in the multiple function layers stored in the memory. The first acquisition function layer is used to call a target API based on the position information of the target vehicle to obtain a corresponding target dataset; the target dataset includes at least a road scene dataset and an administrative region dataset; The second acquisition function layer is used to identify a road scene type based on the road scene dataset and obtain an administrative region attribute based on the administrative region dataset; And the output function layer is used to output a driving environment recognition result in combination with the road scene type and the administrative region attribute.

Images