Automatic driving method and device based on thunder-vision fusion and electric snowmobile

Abstract

The embodiment of the invention discloses an automatic driving method and device based on thunder-vision fusion and an electric snowmobile, and belongs to the technical field of automatic driving. According to the embodiment of the invention, a plurality of initial automatic driving strategies can be generated according to the driving user request information, and the task elements representing each strategy are mined through the Leiyu fusion state vector, so that the automatic driving system can comprehensively and meticulously analyze the execution requirements of each strategy under the Leiyu fusion technology. The intelligent driving state evaluation record is obtained, the road network environment quality characteristics of the initial strategies are calculated based on the intelligent driving state evaluation record, the previous empirical data are fully utilized in the process, and the strategy evaluation is more reliable and scientific. By screening out the target automatic driving strategy and issuing the target automatic driving strategy to the target electric snowmobile, it can be ensured that the driving strategy most suitable for the current road network environment is selected. And the adaptability and the safety of automatic driving of the electric snowmobile are integrally improved.

Description

Technical Field

[0001] The embodiments of the present invention belong to the field of autonomous driving technology, and specifically relate to an autonomous driving method, device and electric snowmobile based on radar and vision fusion. Background Art

[0002] There are many technical challenges in the field of autonomous driving of electric snowmobiles. Traditional autonomous driving strategy generation often lacks comprehensive consideration of multiple factors. For example, different driving users have different needs, but existing technologies may not be able to comprehensively and meticulously generate diverse initial strategies based on these needs. Moreover, for the application of radar-visual fusion technology in autonomous driving, there is a lack of effective ways to deeply analyze the specific needs of radar-visual fusion when each strategy is executed, that is, the strategy execution task elements. At the same time, when evaluating whether the autonomous driving strategy is suitable for the current road network environment, past experience data is not fully utilized. This makes it difficult to accurately select the optimal autonomous driving strategy under different road conditions (such as snow-covered roads, where the depth of snow, road flatness and width all affect driving), traffic flow (the number of snowmobiles and skiers changes frequently), and weather conditions (bad weather such as snowfall, low visibility and strong winds), thereby increasing driving risks. Summary of the invention

[0003] The embodiments of the present invention provide an automatic driving method, device and electric snowmobile based on radar and vision fusion, which can solve or partially solve the technical problems involved in the above-mentioned background technology.

[0004] An embodiment of the present invention provides an automatic driving method based on radar and vision fusion, which is applied to a radar and vision fusion automatic driving device. The method comprises: processing the acquired driving user request information to generate X initial automatic driving strategies for the driving user request information; X is a positive integer; performing radar and vision fusion state vector mining on each of the initial automatic driving strategies to obtain the radar and vision fusion state vectors corresponding to each of the initial automatic driving strategies; a radar and vision fusion state vector corresponding to the initial automatic driving strategy is used to characterize the strategy execution task element of the initial automatic driving strategy; obtaining an intelligent driving state evaluation record; the intelligent driving state evaluation record contains a description of the association between a past radar and vision fusion state vector set and a past road network environment quality feature set, and one of the past radar and vision fusion state vector sets There is an association description between a past radar-vision fusion state vector and a past road network environment quality feature in the past road network environment quality feature set, and a past radar-vision fusion state vector refers to a radar-vision fusion state vector corresponding to a past autonomous driving strategy of which a driving user requested information in the past; the road network environment quality feature of each initial autonomous driving strategy is calculated through the association description between the radar-vision fusion state vector corresponding to each initial autonomous driving strategy, the past radar-vision fusion state vector set in the intelligent driving state evaluation record, and the past road network environment quality feature set; a target autonomous driving strategy is selected from the X initial autonomous driving strategies according to the road network environment quality features of each initial autonomous driving strategy, and the target autonomous driving strategy is sent to a target electric snowmobile.

[0005] An embodiment of the present invention provides a radar-vision fusion automatic driving device, comprising at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the above method.

[0006] An embodiment of the present invention provides a readable storage medium, on which a program or instruction is stored. When the program or instruction is executed by a processor, the steps of the above method are implemented.

[0007] An embodiment of the present invention provides an electric snowmobile, which is communicatively connected to a radar-vision fusion automatic driving device. The electric snowmobile is used to receive a target automatic driving strategy issued by the radar-vision fusion automatic driving device, and the target automatic driving strategy is determined by the above method.

[0008] In an embodiment of the present invention, multiple initial autonomous driving strategies can be generated according to the driving user's request information, and the task elements of each strategy can be characterized by mining the state vector of the radar-visual fusion, which enables the autonomous driving system to comprehensively and meticulously analyze the execution requirements of each strategy under the radar-visual fusion technology. The intelligent driving state evaluation record is obtained and the road network environment quality characteristics of each initial strategy are calculated based on this. This process makes full use of past experience data to make the strategy evaluation more reliable and scientific. By screening out the target autonomous driving strategy and sending it to the target electric snowmobile, it can ensure that the driving strategy that best suits the current road network environment is selected. The overall adaptability and safety of the electric snowmobile's autonomous driving are improved. Regardless of different road conditions (such as snow thickness, road width), traffic flow (vehicle or skier density) or weather conditions (snowfall, visibility, wind speed), the vehicle can make the best decision, effectively avoiding driving risks caused by improper strategy selection, while improving the user's travel experience. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Figure 1 A flowchart of an autonomous driving method based on radar and vision fusion provided in an embodiment of the present invention.

[0010] Figure 2 A schematic diagram of the structure of a radar-vision fusion automatic driving device provided in an embodiment of the present invention. DETAILED DESCRIPTION

[0011] The following will be combined with the drawings in the embodiments of the present invention to clearly and completely describe the technical solutions in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of them. Based on the embodiments in the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the embodiments of the present invention.

[0012] The terms "first", "second", etc. in the embodiments of the present invention are used to distinguish similar objects, and are not used to describe a specific order or sequence. It should be understood that the data used in this way can be interchangeable under appropriate circumstances, so that the embodiments of the embodiments of the present invention can be implemented in an order other than those illustrated or described here, and the objects distinguished by "first", "second", etc. are generally of the same type, and the number of objects is not limited. For example, the first object can be one or more. In addition, in the embodiments of the present invention, "and / or" means at least one of the connected objects, and the character " / " generally indicates that the objects associated with each other are in an "or" relationship.

[0013] Figure 1 An automatic driving method based on radar and vision fusion is shown, which is applied to a radar and vision fusion automatic driving device. The method includes the following steps 110 to 150.

[0014] Step 110: The radar-vision fusion autonomous driving device processes the acquired driving user request information and generates X initial autonomous driving strategies for the driving user request information; X is a positive integer.

[0015] In an embodiment of the present invention, step 110 involves the step of generating an initial autonomous driving strategy, in which the focus is on obtaining and processing the driver's request information. The radar-vision fusion autonomous driving device must first obtain the driver's request information. The driver's request information may contain a variety of content, such as destination information (such as coordinates or specific place names, such as the coordinates of a specific area of ​​a ski resort (45.321, -73.567)), driving speed requirements (such as a maximum speed limit of 50 kilometers per hour), driving mode requirements (such as comfort mode, energy-saving mode, etc.), etc.

[0016] The Leisight Fusion Autonomous Driving Device will process the acquired driving user request information. This processing may involve operations such as data parsing and format conversion. For example, if the user request information is input in the form of natural language, such as "I want to go to the rest area of ​​the ski resort in comfort mode", the Leisight Fusion Autonomous Driving Device needs to convert it into a machine-understandable format and parse out key information such as the destination is the rest area of ​​the ski resort and the driving mode is comfort mode.

[0017] Based on the processed user request information, the radar-based fusion autonomous driving device generates X initial autonomous driving strategies. Here, X is a positive integer, and the specific value may depend on many factors, such as different driving route selections, different speed adjustment strategies, etc. For example, for a request to go to the rest area of ​​a ski resort, the following three initial autonomous driving strategies may be generated (X=3).

[0018] Initial autonomous driving strategy 1: Choose the shortest route and drive at the upper limit allowed by traffic regulations, but gradually slow down to 10 km / h when approaching the rest area and maintain a larger safety distance (such as 5 meters).

[0019] Initial autonomous driving strategy 2: Choose a route with better scenery but a slightly longer distance, drive at a moderate speed (such as 30 kilometers per hour) in energy-saving mode, and set the safety distance to 3 meters.

[0020] Initial autonomous driving strategy three: choose the route with the least traffic, dynamically adjust the speed according to road conditions (up to 40 km / h), and keep a safe distance of 4 meters.

[0021] Step 120: The radar-vision fusion autonomous driving device mines the radar-vision fusion state vector of each of the initial autonomous driving strategies to obtain the radar-vision fusion state vector corresponding to each of the initial autonomous driving strategies; a radar-vision fusion state vector corresponding to an initial autonomous driving strategy is used to characterize the strategy execution task element of the initial autonomous driving strategy.

[0022] Further, step 120 turns to the mining process of the radar-vision fusion state vector. The radar-vision fusion state vector is a vector used to characterize the strategy execution task elements of the initial autonomous driving strategy. It contains information of multiple dimensions related to the radar-vision fusion technology, which is of great significance for evaluating the feasibility and effectiveness of the autonomous driving strategy in actual execution. For each initial autonomous driving strategy, the radar-vision fusion autonomous driving device will mine the radar-vision fusion state vector. For example, taking strategy one as an example, the mined radar-vision fusion state vector may contain the following elements: LiDAR-related factors: the expected detection range requirements of the LiDAR on the shortest path (e.g., a sector area of ​​100 meters in front and 10 meters on each side), the resolution requirements of the LiDAR (e.g., 0.1 degrees horizontal resolution and 0.5 degrees vertical resolution to ensure accurate detection of small obstacles such as skis, etc.); Visual sensor-related factors: the types of targets that the visual sensor needs to identify (such as road signs, other vehicles on snow, pedestrians, etc.), the clarity requirements for visual images (such as at least 1080p resolution to accurately identify traffic signs in the distance), and the sensitivity requirements for different colors and textures (for example, high sensitivity to identifying black objects against a white snow background).

[0023] Fusion-related factors: the time interval for data fusion in the radar-vision fusion algorithm (such as data fusion every 0.1 seconds to ensure timely acquisition of the latest environmental information), the processing priority of the fused data (for example, giving priority to the radar-vision fusion data of objects that are closer to deal with emergencies), etc.

[0024] For other initial autonomous driving strategies (Strategy 2 and Strategy 3), their corresponding radar-vision fusion state vectors will also be mined in a similar manner. These vectors reflect the requirements and dependencies of different strategies on various aspects of radar-vision fusion technology during execution.

[0025] Step 130: The radar-vision fusion autonomous driving device obtains the intelligent driving status assessment record.

[0026] The intelligent driving status assessment record includes a description of the association between a set of past radar-vision fusion state vectors and a set of past road network environment quality characteristics. There is a description of the association between a past radar-vision fusion state vector in the set of past radar-vision fusion state vectors and a past road network environment quality feature in the set of past road network environment quality features. A past radar-vision fusion state vector refers to a radar-vision fusion state vector corresponding to a past autonomous driving strategy for which a past driving user requested information.

[0027] Next, step 130 is used to obtain the intelligent driving state evaluation record. The intelligent driving state evaluation record is a record containing the association description between the past radar-visual fusion state vector set and the past road network environment quality feature set. This record has important reference value for evaluating the newly generated initial autonomous driving strategy.

[0028] Each past radar-visual fusion state vector in the past radar-visual fusion state vector set is a vector corresponding to a past autonomous driving strategy for a past driver request. For example, in the past, a driver requested to go from the ski resort entrance to the ski lift area. The radar-visual fusion state vector corresponding to the autonomous driving strategy generated at that time may include the detection frequency of the laser radar on the route (detection every 0.2 seconds), the recognition accuracy requirement of the visual sensor for the lift logo in the snow (at least 90%), and other factors.

[0029] The past road network environmental quality characteristics in the past road network environmental quality characteristics set reflect the road network conditions at that time, such as the snow depth of the road (such as the average snow depth is 20 cm), the flatness of the road (such as there are some small ups and downs, the slope is within 5%), the traffic flow (5 other snow vehicles per kilometer), the weather conditions (light snow, visibility is 500 meters), etc.

[0030] The association description establishes the connection between the two. For example, when the lidar detection frequency in the past radar-vision fusion state vector is high and the visual sensor recognition accuracy requirement is high, the corresponding past road network environmental quality characteristics may be large snow depth and moderate traffic flow, which means that more accurate environmental perception is required to ensure safe driving under such road conditions.

[0031] Step 140: The radar and vision fusion autonomous driving device calculates the road network environment quality characteristics of each of the initial autonomous driving strategies through the radar and vision fusion state vectors corresponding to each of the initial autonomous driving strategies, the association description between the past radar and vision fusion state vector set in the intelligent driving state evaluation record and the past road network environment quality characteristic set.

[0032] Then, step 140 involves calculating the road network environment quality characteristics of the initial autonomous driving strategy. The radar-visual fusion autonomous driving device calculates the road network environment quality characteristics of each initial autonomous driving strategy through the radar-visual fusion state vector corresponding to each initial autonomous driving strategy, the association description between the past radar-visual fusion state vector set in the intelligent driving state evaluation record and the past road network environment quality characteristic set.

[0033] Take the initial autonomous driving strategy 1 as an example. For example, the lidar detection range in its radar-vision fusion state vector is similar to some records in the past intelligent driving state evaluation records, such as past records with similar detection range requirements. If in those past situations, the corresponding road network environmental quality characteristics are shallow snow depth (less than 10 cm), large traffic volume (more than 10 snow vehicles per kilometer) and good road flatness (slope within 3%), then it can be preliminarily inferred that strategy 1 may have a better performance under such road network environmental quality characteristics.

[0034] In order to calculate more accurately, some algorithm models can be used. For example, a regression model based on a neural network can be used. The radar-vision fusion state vector of the initial autonomous driving strategy is used as input, and the association description between the past radar-vision fusion state vector set and the past road network environment quality feature set is used as training data. The output of the model is the quantitative value of the road network environment quality feature of the initial autonomous driving strategy. For example, in a trained neural network model, the number of input layer nodes is determined according to the dimension of the radar-vision fusion state vector (such as 20 nodes), the hidden layer adopts a 3-layer structure (the number of nodes in each layer is 30, 20, and 10 respectively), and the output layer is the quantitative value of the road network environment quality feature (such as a value between 0 and 1, 0 represents extremely poor road network environment adaptability, and 1 represents excellent road network environment adaptability).

[0035] In the same way, the other initial autonomous driving strategies (Strategy 2 and Strategy 3) are calculated to obtain their respective road network environment quality characteristic quantization values. For example, the road network environment quality characteristic quantization value obtained by Strategy 2 is 0.6, and the quantization value obtained by Strategy 3 is 0.7.

[0036] Step 150: The radar-vision fusion autonomous driving device selects a target autonomous driving strategy from the X initial autonomous driving strategies according to the road network environment quality characteristics of each of the initial autonomous driving strategies, and sends the target autonomous driving strategy to the target electric snowmobile.

[0037] Finally, step 150 is used to implement the screening and distribution of the target autonomous driving strategy. The radar-vision fusion autonomous driving device screens the target autonomous driving strategy from the X initial autonomous driving strategies based on the road network environment quality characteristics of each initial autonomous driving strategy. This screening process is based on the quantitative value of the road network environment quality characteristics calculated previously.

[0038] For example, when the quantitative values ​​of the road network environment quality characteristics of strategies one, two and three calculated above are 0.5, 0.6 and 0.7 respectively, since the quantitative value of the road network environment quality characteristics of strategy three is the highest, strategy three is selected as the target autonomous driving strategy.

[0039] Once the target autonomous driving strategy (here is strategy 3) is screened, the Leishi fusion autonomous driving device will send the target autonomous driving strategy to the target electric snowmobile. After receiving this target autonomous driving strategy, the target electric snowmobile can perform autonomous driving operations according to the requirements of strategy 3, including driving along the selected route with the least traffic, dynamically adjusting the speed according to the road conditions (up to 40 kilometers per hour), maintaining a safe distance of 4 meters, etc.

[0040] In summary, through a series of steps such as processing the driving user's request information by the radar-vision fusion autonomous driving device, mining the radar-vision fusion state vector, utilizing the intelligent driving status evaluation records, and calculating the road network environment quality characteristics, the target autonomous driving strategy that best suits the current situation is finally screened out and sent to the target electric snowmobile, thereby improving the autonomous driving performance and safety of the electric snowmobile with the support of radar-vision fusion technology.

[0041] In the above technical solution, the radar-visual fusion state vector is a multi-dimensional vector that integrates various state information of the lidar and visual sensors during the execution of the autonomous driving strategy. Each dimension of this vector represents a specific attribute related to the radar-visual fusion technology.

[0042] From the perspective of LiDAR, the vector may contain dimensions related to the detection range of the LiDAR. For example, in the scenario of an electric snowmobile, it may include dimensions such as the front detection distance (e.g., in meters, with a range of values ​​from tens of meters to hundreds of meters, such as 150 meters in front), side detection width (e.g., 15 meters on each side). It may also include dimensions related to the resolution of the LiDAR, such as horizontal resolution (which can be a value between 0.1 degrees and 1 degree, such as 0.3 degrees) and vertical resolution (e.g., 0.5 degrees), which directly affect the detection accuracy of obstacles. In addition, the detection frequency of the LiDAR (e.g., once every 0.1 seconds to 1 second, for example, once every 0.2 seconds) is also a dimension of the vector, which determines the speed at which environmental information is updated.

[0043] For visual sensors, the radar-visual fusion state vector contains dimensions related to visual recognition targets. In the electric snowmobile driving scene, this may include the type of target to be recognized (such as classifying the target type into categories such as traffic signs, other vehicles, pedestrians, special snow signs, etc., where all types may need to be recognized), sensitivity to different colors and textures (for example, black objects against a white snow background are more sensitive, represented by a value between 0 and 1, with 0.8 indicating higher sensitivity), etc. The clarity requirement of the visual image is also a dimension, for example, expressed in pixel resolution (such as at least 720p or higher 1080p).

[0044] In terms of radar-vision fusion, the vector includes the dimension of the time interval for data fusion (such as data fusion every 0.05 to 0.5 seconds, or every 0.1 seconds), the dimension of the processing priority of the fused data (for example, targets can be divided into three levels: near, medium, and far according to their distance, and the fused data of targets that are close are processed first, with 1 representing the highest priority), etc.

[0045] The main function of the radar-vision fusion state vector is to comprehensively and accurately describe the requirements and characteristics of an initial autonomous driving strategy at the radar-vision fusion technology level. It provides a detailed basis for the subsequent evaluation of the adaptability of autonomous driving strategies in different road network environments. Through this vector, it is possible to clearly understand how the lidar and visual sensors need to work together when executing a specific autonomous driving strategy, and what performance status each of them needs to achieve. For example, in an autonomous driving strategy that requires frequent avoidance of skiers, the values ​​of dimensions such as the lidar detection frequency and the visual sensor's recognition sensitivity to pedestrians (skiers) in the radar-vision fusion state vector will be adjusted accordingly to ensure that the skier's position and action intentions can be perceived in a timely and accurate manner, thereby ensuring the safe driving of the electric snowmobile.

[0046] In addition, the strategy execution task elements are the key components involved in the execution of the initial autonomous driving strategy. These elements can be roughly divided into several categories, such as driving route related elements, speed control related elements and safety assurance related elements.

[0047] The driving route-related elements include the determination of the destination (e.g., expressed in coordinate form, with the specific coordinates of the ski resort (42.123, -75.456) as the destination), the choice of the driving route (e.g., different selection strategies such as choosing the shortest path, the path with better scenery, or the path with the least traffic).

[0048] Speed ​​control-related factors include speed setting modes (such as fixed speed mode, set at 30 km / h; or dynamic speed mode, dynamically adjusting the speed according to road conditions and traffic rules, for example, the maximum speed on an open snow road does not exceed 50 km / h, and the speed is reduced to 10 km / h when approaching a crowded area), the basis for speed adjustment (such as based on the distance to the vehicle or pedestrian ahead, the curvature of the road, or traffic signs, etc.).

[0049] Safety assurance-related factors include the setting of safe distances (such as maintaining a safe distance of 3 meters from the vehicle in front during normal driving, and increasing it to 5 meters under special road or weather conditions), and emergency response strategies (for example, when a skier is suddenly detected sliding quickly across the snowmobile's route, should emergency braking be performed or evasive measures be taken, and how to determine the direction and magnitude of the evasion, etc.).

[0050] Furthermore, the radar-visual fusion state vector is a mapping and support for the strategic execution task elements from the perspective of the radar-visual fusion technology. For example, the route selection in the strategic execution task element will affect the lidar detection range in the radar-visual fusion state vector and the target type identified by the visual sensor. If a route with heavy traffic is selected, the detection range of the lidar in the radar-visual fusion state vector may need to be wider to detect more vehicles in advance; the visual sensor needs to pay more attention to the identification of vehicle types, including distinguishing snow vehicles of different sizes and uses. Similarly, speed control-related elements are also related to the radar-visual fusion state vector. When driving at high speeds, the lidar detection frequency in the radar-visual fusion state vector may need to be higher, and the image clarity requirements of the visual sensor are also higher to ensure that environmental changes can be perceived in a timely and accurate manner at high speeds to ensure safety. The safety distance setting in the safety assurance-related element is related to the detection accuracy in the radar-visual fusion state vector. More accurate radar-visual fusion detection can more reasonably determine the safety distance.

[0051] In the embodiment of the present invention, the road network environmental quality characteristics are a comprehensive description of the road network environmental conditions on which the electric snowmobile travels, and are mainly composed of road physical characteristics, traffic flow characteristics, and weather condition characteristics.

[0052] The physical characteristics of the road include snow depth (in centimeters, ranging from a few centimeters to tens of centimeters, such as 15 centimeters), road flatness (expressed as a percentage of slope, such as a 2% slope means the road is relatively flat), road width (for example, a single lane width of 3 meters or a double lane width of 6 meters, etc.), etc.

[0053] Traffic flow characteristics refer to the number of vehicles and pedestrians (including skiers) on the road. For example, the number of snowmobiles per kilometer of road (e.g., 8 snowmobiles per kilometer), the density of pedestrians (skiers) (20 skiers per square kilometer), etc.

[0054] Weather condition characteristics include snowfall (such as light snow, moderate snow or heavy snow, which can be quantified as snowfall per hour, such as light snow is 0.1 mm snowfall per hour), visibility (in meters, such as visibility is 300 meters), wind speed (in meters / second, such as wind speed is 5 meters / second), etc.

[0055] These features can be expressed in a quantitative way, for example, by normalizing values ​​such as snow depth, traffic flow, and snowfall to a value range between 0 and 1. For snow depth, 0 means no snow, and 1 means the snow depth reaches a preset maximum value (such as 50 cm); a traffic flow of 0 means no vehicles and pedestrians, and 1 means the traffic flow reaches saturation, etc.

[0056] The road network environment quality characteristics play a key role in evaluating the applicability of the initial autonomous driving strategy. In the radar-visual fusion autonomous driving device, the adaptability quantitative value of the strategy under the current road network environment quality characteristics is calculated by comparing the radar-visual fusion state vector corresponding to the initial autonomous driving strategy with the associated description in the past intelligent driving state evaluation record. For example, the radar-visual fusion state vector corresponding to an initial autonomous driving strategy indicates that its lidar detection range is narrow and the visual sensor has weak target recognition ability under low visibility. If the current road network environment quality characteristics are heavy snow weather (extremely low visibility) and narrow roads (more snow causes the road to become narrower), then the adaptability quantitative value of this strategy in this road network environment will be low and may not be suitable for adoption. On the contrary, if a strategy has a wide detection range and strong target recognition ability in bad weather in the radar-visual fusion state vector, the adaptability quantitative value will be higher under the same road network environment, and it is more likely to be selected as the target autonomous driving strategy.

[0057] It can be seen that the embodiment of the present invention can generate multiple initial autonomous driving strategies according to the driving user's request information, and mine the task elements that characterize each strategy through the radar-visual fusion state vector, which enables the autonomous driving system to comprehensively and meticulously analyze the execution requirements of each strategy under the radar-visual fusion technology. The intelligent driving state evaluation record is obtained and the road network environment quality characteristics of each initial strategy are calculated based on this. This process makes full use of past experience data to make the strategy evaluation more reliable and scientific. By screening out the target autonomous driving strategy and sending it to the target electric snowmobile, it can ensure that the driving strategy that best suits the current road network environment is selected. The overall adaptability and safety of the electric snowmobile's autonomous driving are improved. Regardless of different road conditions (such as snow thickness, road width), traffic flow (vehicle or skier density) or weather conditions (snowfall, visibility, wind speed), the vehicle can make the best decision, effectively avoiding driving risks caused by improper strategy selection, and improving the user's travel experience.

[0058] In some examples, each of the initial autonomous driving strategies is composed of one or more strategy task nodes; the performing radar-vision fusion state vector mining on each of the initial autonomous driving strategies to obtain the radar-vision fusion state vector corresponding to each of the initial autonomous driving strategies includes: determining any initial autonomous driving strategy among the X initial autonomous driving strategies as the target initial autonomous driving strategy; obtaining the node depth of each of the strategy task nodes in the target initial autonomous driving strategy, and prioritizing one or more strategy task nodes in the target initial autonomous driving strategy based on the node depth of each strategy task node to obtain a first strategy task node chain; performing node radar-vision fusion state vector mining on each of the strategy task nodes in the first strategy task node chain to obtain the node radar-vision fusion state vector corresponding to each of the strategy task nodes; prioritizing the node radar-vision fusion state vectors corresponding to the one or more strategy task nodes based on the priority order of the one or more strategy task nodes in the first strategy task node chain to obtain a node radar-vision fusion state vector chain; and determining the node radar-vision fusion state vector chain as the radar-vision fusion state vector corresponding to the target initial autonomous driving strategy.

[0059] Furthermore, the node lightning-vision fusion state vector mining is performed on each of the policy task nodes in the first policy task node chain to obtain the node lightning-vision fusion state vector corresponding to each of the policy task nodes, including: determining any policy task node in the first policy task node chain as a target policy task node; performing multi-mode node element identification on the target policy task node to obtain the multi-mode node element vector of the target policy task node; performing mutual attention operation on the multi-mode node element vector of the target policy task node to obtain the node lightning-vision fusion state vector of the target policy task node.

[0060] Furthermore, the mutual attention operation is performed on the multi-mode node element vector of the target strategy task node to obtain the node lightning vision fusion state vector of the target strategy task node, including: based on the task node conduction characteristics of the target strategy task node, a feature connection operation is performed on the multi-mode node element vector of the target strategy task node to obtain the task connection characteristics of the target strategy task node; according to the lightning vision weight enhancement rule, the task connection characteristics of the target strategy task node are subjected to lightning vision weight enhancement to obtain the task connection weighted characteristics of the target strategy task node; and the task connection weighted characteristics of the target strategy task node are determined as the node lightning vision fusion state vector of the target strategy task node.

[0061] In the actual application process, the radar-visual fusion state vector of each of the initial autonomous driving strategies is mined, and the detailed expansion description of the radar-visual fusion state vector corresponding to each of the initial autonomous driving strategies is obtained as follows.

[0062] 1. The composition of the initial autonomous driving strategy and the overall process of mining the radar-vision fusion state vector 1) Structure of the initial autonomous driving strategy In the embodiment of the present invention, each initial autonomous driving strategy is composed of one or more strategy task nodes. These strategy task nodes are the basic units that constitute the initial autonomous driving strategy. Each node carries specific task information, and these task information together constitute the entire autonomous driving strategy. For example, an initial autonomous driving strategy may include strategy task nodes such as "detecting the road conditions ahead", "adjusting the speed", and "maintaining a safe distance".

[0063] 2) The process of mining the state vector of the radar and vision fusion 2.1) Determine the target initial autonomous driving strategy: First, any initial autonomous driving strategy among the X initial autonomous driving strategies is determined as the target initial autonomous driving strategy. This step is to mine the radar-vision fusion state vector for each initial autonomous driving strategy.

[0064] 2.2) Priority sorting of strategy task nodes Get node depth: To get the node depth of each strategy task node in the target initial autonomous driving strategy. The node depth can reflect the hierarchical relationship of the strategy task nodes in the entire strategy. For example, "Detect the road ahead" may be a more basic task node, and its node depth is shallow, set to 1; while the node "Adjust the speed according to the road ahead" may depend on the result of the "Detect the road ahead" node, and its node depth can be 2.

[0065] Priority sorting: Based on the node depth of each strategy task node, one or more strategy task nodes in the target initial autonomous driving strategy are prioritized to obtain the first strategy task node chain. In this node chain, nodes are arranged from high to low according to priority. For example, according to the node depth set previously, the "Detect the road ahead" node will be ranked before the "Adjust the speed according to the road ahead" node.

[0066] 2.3) Node radar and vision fusion state vector mining Operation of a single node: Perform node radar and vision fusion state vector mining on each strategy task node in the first strategy task node chain to obtain the node radar and vision fusion state vector corresponding to each strategy task node. This process is performed separately for each strategy task node.

[0067] Constructing a node lightning and vision fusion state vector chain: Based on the priority of one or more policy task nodes in the first policy task node chain, the node lightning and vision fusion state vectors corresponding to one or more policy task nodes are prioritized to obtain a node lightning and vision fusion state vector chain. For example, if there are three policy task nodes A, B, and C, and their corresponding node lightning and vision fusion state vectors are VA, VB, and VC, respectively, the node lightning and vision fusion state vector chain can be [VA, VB, VC] in order of priority.

[0068] Determine the radar-visual fusion state vector corresponding to the target initial autonomous driving strategy: Finally, the node radar-visual fusion state vector chain is determined as the radar-visual fusion state vector corresponding to the target initial autonomous driving strategy. This radar-visual fusion state vector comprehensively reflects the relevant states of each task node of the target initial autonomous driving strategy under the radar-visual fusion technology.

[0069] 2. Detailed steps of mining node radar-vision fusion state vector 1) Determination of target strategy task nodes When mining the node radar-vision fusion state vector of each policy task node in the first policy task node chain, any policy task node in the first policy task node chain is determined as the target policy task node. This is to process each node step by step.

[0070] 2) Identification of multi-mode node elements and acquisition of multi-mode node element vectors Perform multi-mode node element identification on the target strategy task node. For example, for the target strategy task node "detect the road conditions ahead", the multi-mode node elements may include the detection range of the lidar (for example, within 100 meters ahead), the detection frequency (once every 0.2 seconds), the type of target identified by the visual sensor (such as vehicles, pedestrians, traffic signs, etc.), and the image clarity requirement (720p). These elements together constitute the multi-mode node element vector of the target strategy task node. For example, a vector representing this multi-mode node element vector may be [100, 0.2, 111, 720], where 111 is a code indicating the identification of three target types: vehicles, pedestrians, and traffic signs.

[0071] 3) Mutual attention operation to obtain the node radar fusion state vector 3.1) Feature connection operation: Based on the task node conduction feature of the target strategy task node, the multi-mode node element vector of the target strategy task node is subjected to a feature connection operation to obtain the task connection feature of the target strategy task node. The task node conduction feature can be a feature related to the information conduction of the task node in the entire strategy. For example, the task node conduction feature is a weight vector [0.5, 0.3, 0.2], which is subjected to a feature connection operation (here, it can be a weighted summation operation or the like) with the multi-mode node element vector [100, 0.2, 111, 720] to obtain the task connection feature.

[0072] 3.2) Radar weight enhancement: According to the radar weight enhancement rule, the task connection features of the target strategy task node are enhanced by radar weight to obtain the task connection weighted features of the target strategy task node. The radar weight enhancement rule may be to assign different weights according to the importance of lidar and visual sensors in the radar fusion technology. For example, the weight of lidar-related factors is increased by 0.1, and the weight of visual sensor-related factors is adjusted according to their importance. For example, after the radar weight enhancement, the task connection weighted features become [110, 0.22, 120, 750] (this is just an example value, which is actually calculated according to the specific weight enhancement rule).

[0073] 3.3) Determine the node lightning and vision fusion state vector: The task connection weighted features of the target strategy task node are determined as the node lightning and vision fusion state vector of the target strategy task node. This node lightning and vision fusion state vector accurately reflects the state of the target strategy task node under the lightning and vision fusion technology, taking into account multiple factors such as multi-mode node elements, task node conduction characteristics, and lightning and vision weight enhancement.

[0074] With this design, first, the initial autonomous driving strategy is decomposed into strategy task nodes, and the priorities are sorted based on the node depth, which can analyze the internal structure and execution order of each strategy in more detail and improve the rationality of strategy execution. For example, when dealing with complex road conditions, the high-priority "detect road conditions ahead" node is executed first, which can obtain environmental information in time so that subsequent nodes can make correct decisions. Secondly, the node radar fusion state vector is obtained through multi-mode node element recognition, mutual attention operation and other steps, which can comprehensively and accurately reflect the state of each strategy task node under the radar fusion technology. Taking a specific numerical example, under different multi-mode node elements such as detection range and detection frequency, the node radar fusion state vector obtained through precise calculation and weight enhancement provides a more reliable basis for evaluating the adaptability of the strategy in the radar fusion environment. Finally, the overall process helps to improve the accuracy and adaptability of the autonomous driving strategy, thereby improving the safety and efficiency of electric snowmobile autonomous driving.

[0075] In the next step, based on the task node conduction characteristics of the target policy task node, a feature connection operation is performed on the multi-mode node element vector of the target policy task node to obtain the task connection characteristics of the target policy task node, including: performing a node clustering operation on the target policy task node based on the task node conduction characteristics of the target policy task node to obtain the target node cluster to which the target policy task node belongs; updating the first intelligent driving control event feature in the multi-mode node element vector of the target policy task node into the second intelligent driving control event feature used to characterize the vehicle-mounted human-computer interaction event indicated by the target node cluster to obtain the task connection characteristics of the target policy task node.

[0076] Based on this, the task connection features of the target strategy task node are subjected to thunder vision weight enhancement according to the thunder vision weight enhancement rule to obtain the task connection weighted features of the target strategy task node, including: obtaining the connection features to be processed whose feature attention distribution does not match the thunder vision weight enhancement rule from the task connection features of the target strategy task node; adjusting the attention weight of the connection features to be processed based on the thunder vision weight enhancement rule to obtain the target connection features; the feature attention distribution of the target connection features matches the thunder vision weight enhancement rule; determining the feature set composed of the target connection features and the remaining connection features as the task connection weighted features of the target strategy task node; the remaining connection features refer to the connection features in the task connection features of the target strategy task node except the connection features to be processed.

[0077] In an embodiment of the present invention, the specific implementation methods of "based on the task node conduction characteristics of the target strategy task node, performing a feature connection operation on the multi-mode node element vector of the target strategy task node to obtain the task connection characteristics of the target strategy task node" and "according to the thunder vision weight enhancement rule, performing thunder vision weight enhancement on the task connection characteristics of the target strategy task node to obtain the task connection weighted characteristics of the target strategy task node" are as follows.

[0078] 1. Feature connection operation based on task node conduction features 1) Node clustering operation and determination of target node cluster In the process of performing feature connection operation on the multi-mode node element vector of the target policy task node based on the task node conduction feature of the target policy task node to obtain the task connection feature, the target policy task node is firstly subjected to node clustering operation based on the task node conduction feature of the target policy task node. The task node conduction feature contains characteristics related to information transmission and interaction of the task node, which can be used to divide the target policy task node into specific groups, namely, the target node cluster.

[0079] For example, the task node conduction feature contains some identification information, such as the task type identification, information flow identification, etc. If clustering is performed according to the task type identification, for a target strategy task node with the "road condition detection and processing" type identification, it may be divided into a specific node cluster. This node cluster may contain other task nodes related to road conditions, such as "recognize road signs" and "detect road flatness". For example, this node cluster is marked as cluster 1.

[0080] 2) Feature update of multi-mode node feature vector After obtaining the target node cluster to which the target strategy task node belongs, the first intelligent driving control event feature in the multi-mode node element vector of the target strategy task node is updated to the second intelligent driving control event feature used to characterize the vehicle-mounted human-computer interaction event indicated by the target node cluster, thereby obtaining the task connection feature of the target strategy task node.

[0081] Taking a specific numerical example, for example, the multi-mode node element vector of the target strategy task node is [100, 0.2, 111, 720, 5], where the last value 5 represents the first intelligent driving control event feature, which may represent a specific driving control mode, such as manual driving assistance mode. If the vehicle-mounted human-computer interaction event indicated by the target node cluster (cluster 1) corresponds to a fully autonomous driving mode, then the second intelligent driving control event feature can be 1, indicating a fully autonomous driving mode. The updated multi-mode node element vector (i.e., the task connection feature) becomes [100, 0.2, 111, 720, 1]. This update makes the multi-mode node element vector of the target strategy task node match the vehicle-mounted human-computer interaction event of the node cluster to which it belongs, thereby better reflecting the task connection characteristics in a specific cluster environment.

[0082] 2. Weight enhancement operation based on the thunder vision weight enhancement rules 1) Obtaining the characteristics of the connection to be processed In the process of performing thunder-vision weight enhancement on the task connection features of the target strategy task node according to the thunder-vision weight enhancement rule to obtain the task connection weighted features, the connection features to be processed whose feature attention distribution does not match the thunder-vision weight enhancement rule are first obtained from the task connection features of the target strategy task node.

[0083] For example, the radar weight enhancement rule stipulates that the total attention weight of lidar-related factors (such as detection range, detection frequency, etc.) in a specific task connection feature should account for 40%, the total attention weight of visual sensor-related factors (such as target type recognition, image clarity requirements, etc.) should account for 30%, and the total attention weight of other related factors (such as intelligent driving control event features, etc.) should account for 30%. If the task connection feature is [100, 0.2, 111, 720, 1], after analysis, it is found that the total attention weight of lidar-related factors is only 30%, then these lidar-related factors (such as detection range 100 and detection frequency 0.2) may be determined as connection features to be processed.

[0084] 2) Attention weight adjustment and acquisition of target connection features Based on the radar weight enhancement rule, the attention weight of the connection feature to be processed is adjusted to obtain the target connection feature. Continuing with the above example, for the lidar related elements to be processed, according to the radar weight enhancement rule, the total attention weight needs to be increased to 40%. For example, through a certain weight adjustment algorithm, such as increasing the weight value of each element according to a certain proportion. If the original weight of the detection range 100 is 0.15, and the original weight of the detection frequency 0.2 is 0.15, after adjustment, the weight of the detection range 100 may become 0.2, and the weight of the detection frequency 0.2 may become 0.2 (this is just an example value, and the actual calculation is based on the specific adjustment algorithm), and the adjusted connection feature (i.e., the target connection feature) becomes [120, 0.25, 111, 720, 1] (the numerical changes in the detection range and detection frequency here are to reflect the effect of the weight adjustment, not the actual simple numerical correspondence), at this time The feature attention distribution of the target connection feature matches the radar weight enhancement rule.

[0085] 3) Determination of task connection weighted features The feature set composed of the target connection feature and the remaining connection feature is determined as the task connection weighted feature of the target strategy task node. In the above example, the remaining connection feature is [111, 720, 1]. Then the task connection weighted feature is the feature set composed of [120, 0.25] in the target connection feature [120, 0.25, 111, 720, 1] and the remaining connection feature [111, 720, 1], that is, [120, 0.25, 111, 720, 1]. This task connection weighted feature not only meets the radar weight enhancement rule, but also comprehensively considers the weight adjustment of different factors, and more accurately reflects the characteristics of the target strategy task node under the radar fusion technology.

[0086] With this design, firstly, through node clustering operations and feature updates of multi-mode node element vectors, the target strategy task nodes can be better matched with the vehicle-mounted human-computer interaction events of the node cluster to which they belong, which helps to improve the adaptability of the autonomous driving strategy in different interaction scenarios. For example, accurately adjusting the intelligent driving control event characteristics according to the requirements of the node cluster can ensure the correct operation of the vehicle in different driving modes. Secondly, the weight enhancement operation performed according to the radar and vision weight enhancement rules obtains and adjusts the connection features to be processed so that the attention distribution of the task connection features meets the requirements of the rules. Taking the specific attention weight ratio requirements as an example, this operation can reasonably allocate the weights of elements such as lidar and visual sensors, improve the accuracy and effectiveness of radar and vision fusion technology in autonomous driving strategies, and thus improve the performance and safety of the entire autonomous driving system.

[0087] Under some preferred design ideas, the processing of the acquired driving user request information to generate X initial autonomous driving strategies for the driving user request information includes: receiving the driving user request information uploaded by the vehicle-mounted server; performing intelligent driving demand mining on the driving user request information to obtain an intelligent driving demand feature map of the driving user request information; mapping and decoding a strategy vector based on the intelligent driving demand feature map to generate X initial autonomous driving strategies for the driving user request information.

[0088] In the embodiment of the present invention, the detailed expansion description of the technical solution recorded in the preferred design idea is as follows.

[0089] 1. Overall process of generating initial autonomous driving strategy 1) Receiving driving user request information Under this preferred design idea, the first thing to do is to receive the driving user request information uploaded by the vehicle server. As the core hub for information interaction within the vehicle, the driving user request information uploaded by the vehicle server contains the user's various expectations and requirements for autonomous driving. For example, the driving user request information may be transmitted in a specific data format, which includes the destination information (in the form of geographic coordinates, such as longitude 120.123 and latitude 30.456), the expected driving speed range (such as the maximum speed not exceeding 60 kilometers per hour), and the requirements for driving comfort (such as smooth driving, avoiding sudden acceleration and braking).

[0090] 2) Intelligent driving demand mining and acquisition of intelligent driving demand feature map Intelligent driving demand mining: The received driving user request information is mined for intelligent driving demand. This process is to deeply analyze the various implicit demands related to intelligent driving in the user request information. For example, if the user request information contains the destination of a mountain ski resort and the current weather is snowing, in addition to the clear destination coordinates, it also implies that the vehicle needs to have the ability to cope with snowy road conditions (such as anti-skid control, recognition of snow-covered roads, etc.) and low visibility (such as visual sensor adjustment in rime weather, etc.).

[0091] Construction of intelligent driving demand feature map: The intelligent driving demand feature map of the driving user's request information is obtained through intelligent driving demand mining. The intelligent driving demand feature map is a comprehensive description of intelligent driving demand expressed in graphical or vector form. For example, when constructing an intelligent driving demand feature map in vector form, it may contain values ​​of multiple dimensions. For example, if the vector dimension is [road condition adaptability, speed control requirements, safety requirements, environmental perception requirements], for a request to go to a mountain ski resort, the road condition adaptability dimension may correspond to a higher value, such as 0.8 (between 0-1, 0.8 indicates a higher requirement for adapting to snowy road conditions); the speed control requirement dimension may be a value corresponding to a speed range, such as the coded value corresponding to [0, 60] kilometers per hour; the safety requirement dimension may be set to a higher value according to the user's requirements for comfort, such as 0.7 (indicating a higher requirement for safe distance and stable driving); the environmental perception requirement dimension may be set to 0.9 due to snowy weather (indicating a high requirement for identifying targets in snow and low visibility).

[0092] 3) Mapping and decoding of strategy vector and generation of initial autonomous driving strategy Mapping and decoding of strategy vectors: Mapping and decoding of strategy vectors are performed based on the intelligent driving demand feature map. This process involves converting each characteristic value in the intelligent driving demand feature map into specific autonomous driving strategy elements. For example, based on the value of 0.8 in the road condition adaptability dimension, it may be determined during the mapping and decoding process that the vehicle needs to have a specific snow driving mode, such as adjusting tire pressure, optimizing the power distribution of the four-wheel drive system, etc. For the speed control requirement dimension of [0, 60] km / h, it will be converted into a specific speed control strategy, such as dynamically adjusting the vehicle speed within this speed range according to road conditions and traffic rules on different sections of the road (such as flat snowy roads, sloping snow slopes, etc.).

[0093] Generation of initial autonomous driving strategies: Through the mapping and decoding of the strategy vector, X initial autonomous driving strategies are finally generated based on the driving user's request information. Here X is a positive integer, and its specific value depends on many factors. For example, based on different route selections (such as the shortest path, the safest path, the path with the best scenery), different vehicle performance adjustment strategies (such as adjustments in energy-saving mode, adjustments in high-performance mode), and other factors, multiple different initial autonomous driving strategies may be generated. For example, three initial autonomous driving strategies (X=3) are generated based on the above user request information: Strategy 1: Choose the shortest path to the destination and use the energy-saving mode. The speed is dynamically adjusted within [0, 60] km / h according to the road conditions. The vehicle is adjusted to the snow energy-saving driving mode (such as reducing engine power output, optimizing the energy consumption of the anti-skid system, etc.). The safety distance is set to 3 meters. The visual sensor and lidar work in a mode that adapts to the low visibility of snow (such as reducing the frame rate of the visual sensor and adjusting the detection frequency of the lidar, etc.).

[0094] Strategy 2: Choose the safest route to the destination, use the standard performance mode, maintain a speed of about 40 km / h, adjust the vehicle to the snow safety driving mode (such as enhancing anti-skid control, increasing the safety distance warning sensitivity, etc.), set the safety distance to 5 meters, and the visual sensor and lidar work in high-precision mode (such as improving the resolution of the visual sensor, increasing the detection accuracy of the lidar, etc.).

[0095] Strategy 3: Choose the most scenic route to the destination, use high-performance mode, dynamically adjust the speed according to the road conditions within [0, 60] km / h, adjust the vehicle to high-performance snow driving mode (such as maximizing engine power output to cope with possible climbing situations, optimizing the handling performance of the four-wheel drive system, etc.), set the safety distance to 4 meters, and use the visual sensor and lidar to work in a comprehensive balance mode (while ensuring a certain visual effect and detection accuracy, taking into account energy consumption and system resource usage).

[0096] Therefore, firstly, by mining the intelligent driving demand of driving user request information to obtain the intelligent driving demand characteristic map, we can fully and deeply understand the user's needs, and no longer be limited to the surface request content. Judging from the numerical examples in the intelligent driving demand characteristic map, the values ​​of different dimensions accurately reflect the degree of user demand in terms of road conditions, speed, safety and environmental perception. Secondly, the initial autonomous driving strategy is generated by mapping and decoding the strategy vector based on the intelligent driving demand characteristic map, making the generated strategy more targeted and reasonable. For example, the differences in different initial autonomous driving strategies in terms of path selection, vehicle mode, speed control and sensor working mode can meet the diverse needs of users and adapt to different driving conditions, improving the flexibility and adaptability of the autonomous driving system, thereby improving the user experience and the safety of autonomous driving.

[0097] Under some other optional design ideas, the road network environment quality characteristics of each initial autonomous driving strategy are calculated through the association description between the radar and vision fusion state vectors corresponding to each initial autonomous driving strategy, the past radar and vision fusion state vector set in the intelligent driving state evaluation record, and the past road network environment quality feature set, including: determining any initial autonomous driving strategy among the X initial autonomous driving strategies as the target initial autonomous driving strategy; roaming the past radar and vision fusion state vector set in the intelligent driving state evaluation record based on the radar and vision fusion state vector corresponding to the target initial autonomous driving strategy; if there is a radar and vision fusion state vector corresponding to the target initial autonomous driving strategy in the past radar and vision fusion state vector set, If there is a target past radar and vision fusion state vector with the same value, the past road network environment quality characteristics corresponding to the target past radar and vision fusion state vector are determined as the road network environment quality characteristics of the target initial autonomous driving strategy; if there is no target past radar and vision fusion state vector that is the same as the radar and vision fusion state vector corresponding to the target initial autonomous driving strategy in the past radar and vision fusion state vector set, the target initial autonomous driving strategy is converted to obtain Y autonomous driving sub-strategies, and based on the Y autonomous driving sub-strategies, the association description between the past radar and vision fusion state vector set in the intelligent driving status evaluation record and the past road network environment quality characteristics set, the road network environment quality characteristics of the target initial autonomous driving strategy are determined; Y is a positive integer.

[0098] In the next step, the road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the Y autonomous driving secondary strategies, the association description between the past radar vision fusion state vector set in the intelligent driving status evaluation record and the past road network environment quality characteristic set, including: obtaining the local radar vision fusion state vector corresponding to each of the autonomous driving secondary strategies; wandering the past radar vision fusion state vector set according to the local radar vision fusion state vector corresponding to each of the autonomous driving secondary strategies; and determining the road network environment quality characteristics of the target initial autonomous driving strategy based on the wandering results of the past radar vision fusion state vector set.

[0099] In a further embodiment, the target initial autonomous driving strategy is composed of one or more strategy task nodes; the wandering result of the past radar vision fusion state vector set is the result of including the target local radar vision fusion state vector in the past radar vision fusion state vector set, and the target local radar vision fusion state vector refers to any local radar vision fusion state vector among Y local radar vision fusion state vectors; the road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the wandering results of the past radar vision fusion state vector set, including: determining the autonomous driving secondary strategy corresponding to the target local radar vision fusion state vector as the target autonomous driving secondary strategy, and setting the target autonomous driving secondary strategy The strategy task nodes slightly included are determined as strategy task quality labeling nodes; the strategy task nodes other than the strategy task quality labeling nodes in the target initial autonomous driving strategy are determined as strategy task nodes to be processed, and the radar vision sensor acquisition data corresponding to the strategy task nodes to be processed are obtained; the radar vision sensor acquisition data includes the past road network environment quality characteristics corresponding to the target local radar vision fusion state vector; the node environment quality sub-characteristics corresponding to the strategy task node to be processed are determined based on the radar vision sensor acquisition data corresponding to the strategy task node to be processed; the road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the node environment quality sub-characteristics corresponding to the strategy task node to be processed.

[0100] It can be understood that the above optional design ideas and the subsequent two embodiments are described in detail as follows.

[0101] 1. Overall process of calculating the road network environment quality characteristics for the initial autonomous driving strategy 1. Determination of the initial autonomous driving strategy and the wandering of the radar-vision fusion state vector 1) Determination of the target initial autonomous driving strategy In the process of calculating the road network environment quality characteristics of each initial autonomous driving strategy, we first need to select one of the X initial autonomous driving strategies as the target initial autonomous driving strategy. This selection is to calculate the road network environment quality characteristics for each initial autonomous driving strategy separately. X is a positive integer, which represents the number of initial autonomous driving strategies. For example, X=5 means that there are 5 different initial autonomous driving strategies to choose from. Each strategy has its own unique components, which may involve different driving route planning, speed control strategies, safety distance settings, etc.

[0102] 2) Walk through the set of past lightning-vision fusion state vectors based on the lightning-vision fusion state vector Once the target initial autonomous driving strategy is determined, it is necessary to perform a roaming operation in the set of past radar-visual fusion state vectors in the intelligent driving state evaluation record based on the radar-visual fusion state vector corresponding to the strategy. The radar-visual fusion state vector is a multi-dimensional vector that contains various state information related to the lidar and visual sensors. For example, for the radar-visual fusion state vector of the target initial autonomous driving strategy, the lidar-related dimensions may include detection range (for example, 120 meters), detection frequency (for example, once every 0.15 seconds), horizontal resolution (such as 0.3 degrees), etc.; the visual sensor-related dimensions may include image clarity requirements (such as 1080p), encoding for identifying target types (for example, represented by binary encoding, identifying vehicles, pedestrians, and traffic signs as 111 respectively), etc.

[0103] When walking in the past radar-vision fusion state vector set, each dimension of the radar-vision fusion state vector of the target initial autonomous driving strategy is compared with each vector in the past radar-vision fusion state vector set. For example, a vector in the past radar-vision fusion state vector set may be [110 meters, once every 0.18 seconds, 0.25 degrees, 720p, 101]. It is necessary to compare whether the values ​​of these two vectors in each dimension are the same or match within a certain error range.

[0104] (II) Processing when there is the same radar-vision fusion state vector 1) Determine the road network environmental quality characteristics If a target past radar-vision fusion state vector identical to the radar-vision fusion state vector corresponding to the target initial autonomous driving strategy is found in the past radar-vision fusion state vector set, then the past road network environment quality characteristics corresponding to the target past radar-vision fusion state vector are directly determined as the road network environment quality characteristics of the target initial autonomous driving strategy. The past road network environment quality characteristics are also a set containing many aspects of information. For example, it may include the snow depth of the road (for example, 20 cm), the road flatness (expressed in slope, such as a 3% slope), the traffic flow (8 vehicles per kilometer), the weather conditions (such as light snow, visibility of 400 meters), etc. These past road network environment quality characteristics are directly mapped to the target initial autonomous driving strategy as its road network environment quality characteristics, thereby providing a direct basis for evaluating the applicability of the strategy in the current similar road network environment.

[0105] (III) Processing when there is no identical radar-vision fusion state vector 1) Conversion of target initial autonomous driving strategy If there is no target past radar-vision fusion state vector in the set of past radar-vision fusion state vectors that is identical to the radar-vision fusion state vector corresponding to the target initial autonomous driving strategy, it is necessary to convert the target initial autonomous driving strategy to obtain Y autonomous driving secondary strategies. The conversion here is a rule-based adjustment, the purpose of which is to generate a strategy that is more comparable to past experience. For example, the target initial autonomous driving strategy originally set a fixed speed of 50 km / h, and the converted autonomous driving secondary strategy may be a dynamic adjustment of the speed between [45, 55] km / h (an example case when Y=1). If Y=3, there may be three different autonomous driving secondary strategies, each adjusting the target initial autonomous driving strategy from different angles, such as changes in driving route selection, speed adjustment basis, or safety distance setting.

[0106] 2) Determine road network environment quality characteristics based on autonomous driving secondary strategies Based on the Y autonomous driving secondary strategies, the association description between the past radar-visual fusion state vector set and the past road network environment quality feature set in the intelligent driving state evaluation record, the road network environment quality features of the target initial autonomous driving strategy are determined. This step involves the following sub-steps: 2. Steps to determine road network environmental quality characteristics based on autonomous driving secondary strategies (I) Obtain the local radar fusion state vector and walk through the set of past radar fusion state vectors 1) Acquisition of local radar-vision fusion state vector First, we need to obtain the local radar-visual fusion state vectors corresponding to each autonomous driving secondary strategy. These local radar-visual fusion state vectors are specially constructed for autonomous driving secondary strategies. Their structure and meaning are similar to the radar-visual fusion state vectors, but they reflect the characteristics of the autonomous driving secondary strategy. For example, for an autonomous driving secondary strategy, the lidar detection range in its local radar-visual fusion state vector may become 110 meters (adjusted from 120 meters of the target initial autonomous driving strategy), the detection frequency becomes once every 0.12 seconds, and the image clarity requirement of the visual sensor becomes 720p (adjusted from 1080p), etc.

[0107] 2) Walk through the set of past radar fusion state vectors based on the local radar fusion state vector According to the local radar-vision fusion state vector corresponding to each autonomous driving secondary strategy, the wandering operation is performed again in the set of past radar-vision fusion state vectors. Similar to the wandering of the radar-vision fusion state vector based on the target initial autonomous driving strategy, each dimension of the local radar-vision fusion state vector is compared with the vectors in the set of past radar-vision fusion state vectors to find matching vectors.

[0108] 2. Road network environment quality characteristics for determining the target initial autonomous driving strategy based on the wandering results 1) Determination of target autonomous driving secondary strategy and strategy task quality labeling nodes When the target initial autonomous driving strategy is composed of one or more strategy task nodes, the walking result of the past radar vision fusion state vector set is the result of including the target local radar vision fusion state vector in the past radar vision fusion state vector set, and the target local radar vision fusion state vector refers to any local radar vision fusion state vector among the Y local radar vision fusion state vectors. Once such a target local radar vision fusion state vector is found, its corresponding autonomous driving sub-strategy is determined as the target autonomous driving sub-strategy. For example, if in the case of Y=3, the autonomous driving sub-strategy corresponding to one of the local radar vision fusion state vectors is found to satisfy the walking result, then this autonomous driving sub-strategy is determined as the target autonomous driving sub-strategy.

[0109] Next, the policy task nodes included in the target autonomous driving sub-strategy are determined as policy task quality labeling nodes. For example, if the target autonomous driving sub-strategy includes the three policy task nodes of "detecting the road conditions ahead", "adjusting the vehicle speed according to the road conditions", and "maintaining a safe distance", then these three nodes are determined as policy task quality labeling nodes.

[0110] 2) Obtaining the strategy task nodes to be processed and the data collected by the radar sensor The strategy task nodes other than the strategy task quality annotation nodes in the target initial autonomous driving strategy are determined as pending strategy task nodes. For example, in addition to the three annotation nodes mentioned above, the target initial autonomous driving strategy also has a node called "Identify Road Signs", so "Identify Road Signs" is a pending strategy task node.

[0111] Get the radar sensor data corresponding to the pending strategy task node. The radar sensor data here includes the past road network environment quality characteristics corresponding to the target local radar fusion state vector. For example, for the pending strategy task node "Identify Road Signs", the past road network environment quality characteristics in the radar sensor data may include the distribution density of traffic signs on the road (5 traffic signs per kilometer), the degree to which the clarity of road signs is affected by weather (such as light snow weather reduces the clarity of signs by 30%), etc.

[0112] 3) Determination of node environmental quality sub-characteristics and road network environmental quality characteristics The node environmental quality sub-feature corresponding to the pending strategy task node is determined based on the radar sensor data corresponding to the pending strategy task node. For example, for the pending strategy task node "identifying road signs", its node environmental quality sub-feature can be determined based on the distribution density of traffic signs on the road and the degree to which the clarity of the signs is affected by the weather. If the distribution density of traffic signs is high and the clarity of the signs is greatly affected by the weather, then this node environmental quality sub-feature may be expressed as a high difficulty in identifying road signs in this road network environment, which may be expressed as 0.7 (0 means easy, 1 means very difficult).

[0113] Finally, the road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the node environment quality sub-characteristics corresponding to the pending strategy task nodes. This may involve a comprehensive calculation algorithm. For example, the target initial autonomous driving strategy has multiple pending strategy task nodes, and their node environment quality sub-characteristics are calculated to be 0.7, 0.6, and 0.8 respectively. Through a weighted average algorithm (for example, the weights are 0.3, 0.3, and 0.4 respectively), the road network environment quality characteristics of the target initial autonomous driving strategy are calculated to be (0.7×0.3+0.6×0.3+0.8×0.4)=0.71.

[0114] Applying the above technical solution, first, the matching vector is found through the set of past radar-vision fusion state vectors in the wandering intelligent driving state evaluation record, making full use of past experience data. When the same vector exists, the road network environment quality characteristics are directly determined, which is efficient and accurate. Secondly, when the same vector does not exist, it is converted to an autonomous driving secondary strategy and subsequent operations are performed, which increases the flexibility of strategy evaluation. The process of determining the strategy task quality annotation nodes and the strategy task nodes to be processed, obtaining the radar-vision sensor acquisition data to determine the node environment quality sub-features and then determining the road network environment quality characteristics can carefully analyze the situation of each strategy task node in different road network environments, thereby improving the adaptability of the entire autonomous driving strategy to the road network environment and ensuring the safety and effectiveness of autonomous driving.

[0115] In some independent technical solutions, the target initial autonomous driving strategy is composed of one or more policy task nodes; the walking result of the past radar vision fusion state vector set is the result that the past radar vision fusion state vector set does not contain any local radar vision fusion state vector; the road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the walking results of the past radar vision fusion state vector set, including: obtaining the node depth of each of the policy task nodes in the target initial autonomous driving strategy, and prioritizing one or more policy task nodes in the target initial autonomous driving strategy based on the node depth of each of the policy task nodes to obtain a second policy task node chain corresponding to the target initial autonomous driving strategy; based on the priority order of the one or more policy task nodes in the second policy task node chain, sequentially mining the node environment quality sub-features corresponding to each of the policy task nodes; determining the node environment quality sub-features corresponding to the policy task terminal node as the road network environment quality characteristics of the target initial autonomous driving strategy; the policy task terminal node refers to the policy task node at the end of the node chain in the second policy task node chain.

[0116] On the basis of the above content, based on the wandering results of the past radar-vision fusion state vector set, the specific implementation of determining the road network environment quality characteristics of the target initial autonomous driving strategy is as follows.

[0117] 1. Overall process of determining the road network environment quality characteristics of the target initial autonomous driving strategy based on specific walk results 1. Structure of the target initial autonomous driving strategy and strategy task nodes 1) Composition of the target initial autonomous driving strategy In this independent technical solution, the target initial autonomous driving strategy is composed of one or more policy task nodes. These policy task nodes are the basic components of the target initial autonomous driving strategy. Each node undertakes a specific task or function and works together to implement the entire autonomous driving strategy. For example, a target initial autonomous driving strategy may include policy task nodes such as "start vehicle self-check", "plan initial driving route", "detect surrounding traffic participants", and "adjust vehicle speed according to road conditions". These nodes are interconnected in a certain logical order and dependency relationship, and jointly determine the execution process and effect of the autonomous driving strategy.

[0118] 2. Priority Arrangement Based on Node Depth and Formation of the Second Strategy Task Node Chain 1) Obtaining and Significance of Node Depth First, we need to obtain the node depth of each strategy task node in the target initial autonomous driving strategy. The node depth reflects the hierarchical relationship and execution order of the strategy task nodes in the entire strategy. For example, "starting vehicle self-check" may be the most basic operation, with a node depth of 1; "planning the initial driving route" may depend on the result of the vehicle self-check, with a node depth of 2; "detecting surrounding traffic participants" depends on the planning of the driving route, with a node depth of 3; "adjusting the speed according to road conditions" depends on the detection of surrounding traffic participants, with a node depth of 4.

[0119] 2) Priority sorting and construction of the second strategy task node chain Based on the node depth of each policy task node, one or more policy task nodes in the target initial autonomous driving strategy are prioritized to obtain a second policy task node chain corresponding to the target initial autonomous driving strategy. In this node chain, the policy task nodes are arranged from high to low priority. For example, according to the node depth mentioned above, the second policy task node chain may be ["start vehicle self-check", "plan initial driving route", "detect surrounding traffic participants", "adjust vehicle speed according to road conditions"]. This priority order determined by the node depth reflects the logical order in the execution process of the autonomous driving strategy, that is, the basic and pre-tasks are executed first, and then the subsequent tasks that depend on the previous results are gradually carried out.

[0120] 3. Sequentially mining node environmental quality sub-features and determining road network environmental quality features 1) Mining node environmental quality sub-features in order Based on the priority of one or more policy task nodes in the second policy task node chain, the node environmental quality sub-features corresponding to each policy task node are mined in order. For each policy task node, its node environmental quality sub-features reflect the difficulty of the node to perform tasks in a specific road network environment or related environmental quality influencing factors. For example, for the "start vehicle self-inspection" node, its node environmental quality sub-features may be related to environmental factors such as temperature and humidity at the location of the vehicle. For example, if the temperature at the location of the vehicle is -10°C, according to empirical data, there may be certain risks in vehicle self-inspection in such a low temperature environment. Its node environmental quality sub-features can be represented by a numerical value, such as 0.6 (0 means no risk, 1 means high risk).

[0121] For the "Planning Initial Driving Route" node, its node environmental quality sub-feature may be related to the congestion of the road network, road construction conditions, etc. If the congestion index of the current road network is 0.4 (0 means no congestion, 1 means severe congestion), and the proportion of road construction sections is 0.1 (i.e., 10% of the sections are under construction), through a specific algorithm (e.g., weighted average algorithm, such as congestion index weight is 0.6, road construction section weight is 0.4), the node environmental quality sub-feature of the node is calculated to be (0.4×0.6+0.1×0.4)=0.28.

[0122] In this way, the node environmental quality sub-features of nodes such as "detect surrounding traffic participants" and "adjust vehicle speed according to road conditions" are calculated in sequence. For example, the node environmental quality sub-features of the "detect surrounding traffic participants" node may be affected by factors such as weather conditions (such as visibility) and the density of traffic participants. For example, if the visibility is 500 meters (the lower the visibility, the greater the detection difficulty), and the density of traffic participants is 30 per square kilometer (the higher the density, the greater the detection difficulty), through a specific function (such as based on the inverse proportional function relationship between visibility and traffic participant density, and normalized), the node environmental quality sub-feature of the node is obtained as 0.7.

[0123] 2) Determine the road network environment quality characteristics of the target initial autonomous driving strategy The node environment quality sub-characteristic corresponding to the end node of the strategy task is determined as the road network environment quality characteristic of the target initial autonomous driving strategy. In the second strategy task node chain constructed previously ["Start vehicle self-check", "Plan initial driving route", "Detect surrounding traffic participants", "Adjust vehicle speed according to road conditions"], "Adjust vehicle speed according to road conditions" is the end node of the strategy task. For example, if the node environment quality sub-characteristic of this node is 0.7, then this value 0.7 is determined as the road network environment quality characteristic of the target initial autonomous driving strategy. This determination method reflects that the final execution effect of the entire autonomous driving strategy depends to a large extent on the execution of the last key task node in the road network environment. Taking it as a representative of the road network environment quality characteristics of the entire strategy can concisely and effectively reflect the adaptability and feasibility of the entire strategy in a specific road network environment.

[0124] With this design, by obtaining the node depth of the strategy task node and sorting out the priorities to obtain the second strategy task node chain, the execution logic order of the target initial autonomous driving strategy can be clearly reflected. This helps to deeply understand the dependencies between the various task nodes, so as to better evaluate the rationality of the entire strategy. Secondly, the node environmental quality sub-features are sequentially mined and the sub-features of the strategy task terminal node are determined as the road network environmental quality features, taking into account the impact of each link on the final result during the entire strategy execution process. Taking a specific numerical example, the sub-feature values ​​calculated by different nodes based on their respective relevant environmental factors can accurately reflect the status of the node in the road network environment. The final determined road network environmental quality features can effectively measure the adaptability of the target initial autonomous driving strategy in a specific road network environment, help to screen out autonomous driving strategies that are more suitable for the current environment, and improve the safety and effectiveness of autonomous driving.

[0125] In some other independent technical solutions, a target autonomous driving strategy is selected from the X initial autonomous driving strategies based on the road network environment quality characteristics of each of the initial autonomous driving strategies, including: obtaining a road network environment quality characteristic with the highest feature normalization value from the road network environment quality characteristics of the X initial autonomous driving strategies; and determining the initial autonomous driving strategy indicated by the road network environment quality characteristic with the highest feature normalization value as the target autonomous driving strategy for which the driving user requests information.

[0126] In this technical solution, the screening process of the target autonomous driving strategy is as follows.

[0127] 1. The overall process of selecting the target autonomous driving strategy based on the road network environment quality characteristics (I) Obtaining feature normalization values 1) Review of road network environmental quality characteristics In the previous technical solution, the road network environment quality characteristics have been calculated for each initial autonomous driving strategy. These road network environment quality characteristics are obtained after comprehensive consideration of multiple factors, such as road conditions (such as snow depth, flatness, etc.), traffic flow (number of vehicles and pedestrians, etc.), weather conditions (visibility, snowfall, etc.) and other factors that affect the execution of autonomous driving strategies. Each initial autonomous driving strategy has its corresponding road network environment quality characteristics, which are expressed in a specific form, which may be a vector containing multiple values ​​or a value processed by a specific algorithm.

[0128] 2) Necessity and operation of feature normalization In order to facilitate the comparison of the road network environment quality characteristics of different initial autonomous driving strategies, feature normalization operations are required. Feature normalization is the process of converting values ​​of different ranges and dimensions into a unified range. For example, for the road network environment quality characteristics of an initial autonomous driving strategy, if it contains the influencing values ​​of three factors: snow depth, traffic flow, and visibility, the snow depth may be an actual value in centimeters (such as 15 centimeters), the traffic flow is expressed in the number of vehicles per kilometer (such as 8 vehicles / kilometer), and the visibility is in meters (such as 400 meters). Through feature normalization, these values ​​of different dimensions can be converted into values ​​between 0 and 1.

[0129] The specific normalization algorithm can adopt different methods according to the actual situation. Taking linear normalization as an example, if the maximum value of the snow depth is set to 50 cm (this is set according to the maximum snow depth that may be encountered in reality), then the normalized value of the snow depth of 15 cm is 15 / 50=0.3. For traffic flow, for example, the maximum traffic flow is set to 20 vehicles per kilometer, and the traffic flow of 8 vehicles / kilometer is normalized to 8 / 20=0.4. For visibility, the minimum visibility is set to 100 meters and the maximum visibility is 1000 meters. The visibility of 400 meters is normalized to (400-100) / (1000-100)=0.33. These normalized values ​​are recombined or further processed to obtain the characteristic normalized values ​​of the road network environment quality characteristics of the initial autonomous driving strategy.

[0130] 2. Determination of target autonomous driving strategy 1) Obtain the road network environment quality feature with the highest feature normalization value After obtaining the normalized values ​​of the road network environment quality characteristics of X initial autonomous driving strategies, the road network environment quality characteristics with the highest normalized value are obtained from these values. For example, if there are 3 initial autonomous driving strategies (X=3), and their road network environment quality characteristics have normalized values ​​of 0.55, 0.6, and 0.45, respectively, then the highest normalized value is 0.6.

[0131] 2) Determine the target autonomous driving strategy The initial autonomous driving strategy indicated by the road network environment quality feature with the highest feature normalization value is determined as the target autonomous driving strategy for the driving user's request for information. Continuing with the above example, the initial autonomous driving strategy corresponding to the feature normalization value of 0.6 is determined as the target autonomous driving strategy. The basis for this determination method is that the higher the feature normalization value, the better adaptability of the initial autonomous driving strategy in the current road network environment. Because this feature normalization value is obtained by combining various factors that affect autonomous driving (such as the road conditions, traffic flow, weather conditions, etc. mentioned above), a higher value means that the strategy has a relative advantage in dealing with various conditions in the current road network environment.

[0132] With this design, first, the road network environment quality characteristics of different initial autonomous driving strategies are converted into a unified measurement standard through feature normalization, which solves the problem that numerical values ​​of different dimensions and ranges are difficult to compare directly. Taking specific numerical examples, different factors such as snow depth, traffic flow and visibility can be compared at the same scale after normalization. Secondly, the target autonomous driving strategy is determined according to the highest feature normalization value, and the most adaptable strategy in the current road network environment can be selected. This helps to improve the efficiency and safety of autonomous driving, because the selected target autonomous driving strategy has a relatively optimal performance when considering various environmental factors, so that it can better cope with actual driving conditions and reduce risks that may be caused by improper strategy selection.

[0133] Further, Figure 2 FIG. 2 is a schematic diagram of a structure of a radar-vision fusion automatic driving device 200 provided in an embodiment of the present invention. Figure 2 The radar-vision fusion automatic driving device 200 shown includes a processor 210, which can call and run a computer program from a memory to implement the method in the embodiment of the present invention.

[0134] Alternatively, if Figure 2 As shown, the radar-vision fusion automatic driving device 200 may further include a memory 230. The processor 210 may call and run a computer program from the memory 230 to implement the method in the embodiment of the present invention.

[0135] The memory 230 may be a separate device independent of the processor 210 , or may be integrated into the processor 210 .

[0136] Alternatively, if Figure 2 As shown, the radar-vision fusion automatic driving device 200 may also include a transceiver 220, and the processor 210 may control the transceiver 220 to interact with other devices. Specifically, it may send information or data to other devices, or receive information or data sent by other devices.

[0137] Optionally, the radar-vision fusion autonomous driving device 200 can implement the corresponding processes corresponding to the storage engine or components in the storage engine (such as a processing module) or a device deployed with a storage engine in each method of the embodiments of the present invention. For the sake of brevity, they will not be repeated here.

[0138] It should be understood that the processor of the embodiment of the present invention may be an integrated circuit chip with signal processing capability.

[0139] It is understood that the memory in the embodiments of the present invention may be a volatile memory or a non-volatile memory, or may include both volatile and non-volatile memory. It should be noted that the memory of the systems and methods described herein is intended to include but is not limited to suitable types of memory.

[0140] Based on the above, a readable storage medium is provided, on which a program or instruction is stored, and when the program or instruction is executed by a processor, the steps of the above method are implemented.

[0141] Based on the same or similar inventive concept as above, an electric snowmobile is also provided. The snowmobile can adopt a wheeled structure and can adopt an EMB brake as a braking system. The electric snowmobile is communicatively connected to a radar-vision fusion automatic driving device. The electric snowmobile is used to receive a target automatic driving strategy issued by the radar-vision fusion automatic driving device, and the target automatic driving strategy is determined by the above method.

[0142] It should be noted that, in this article, the term "comprises", "includes" or any other variant thereof is intended to cover non-exclusive inclusion, so that the process, method, article or device including a series of elements includes not only those elements, but also includes other elements not explicitly listed, or also includes elements inherent to such process, method, article or device. In the absence of further restrictions, the elements defined by the sentence "including one..." do not exclude the presence of other identical elements in the process, method, article or device including the element. In addition, it should be pointed out that the scope of the method and device in the embodiment of the present invention is not limited to performing functions in the order shown or discussed, and may also include performing functions in a substantially simultaneous manner or in the opposite order according to the functions involved, for example, the described method may be performed in an order different from that described, and various steps may also be added, omitted, or combined. In addition, the features described with reference to certain examples may be combined in other examples.

[0143] Through the description of the above implementation methods, those skilled in the art can clearly understand that the above-mentioned implementation methods can be implemented by means of software plus a necessary general hardware platform, and of course, by hardware, but in many cases the former is a better implementation method. Based on such an understanding, the technical solution of the embodiment of the present invention is essentially or the part that contributes to the prior art can be embodied in the form of a computer software product, which is stored in a storage medium (such as ROM / RAM, a magnetic disk, or an optical disk), and includes a number of instructions for enabling a terminal (which can be a mobile phone, a computer, a server, or a network device, etc.) to execute the methods described in each embodiment of the embodiment of the present invention.

[0144] The embodiments of the embodiments of the present invention are described above in conjunction with the accompanying drawings, but the embodiments of the present invention are not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the embodiments of the present invention, ordinary technicians in this field can also make many forms without departing from the purpose of the embodiments of the present invention and the scope of protection of the embodiments of the present invention, all of which are within the protection of the embodiments of the present invention.

Claims

1. An automatic driving method based on radar and vision fusion, characterized in that: The method comprises: Processing the acquired driving user request information to generate X initial automatic driving strategies for the driving user request information; X is a positive integer; Mining the radar-visual fusion state vector of each of the initial autonomous driving strategies to obtain the radar-visual fusion state vector corresponding to each of the initial autonomous driving strategies; a radar-visual fusion state vector corresponding to the initial autonomous driving strategy is used to characterize the strategy execution task element of the initial autonomous driving strategy; Obtaining an intelligent driving status evaluation record; the intelligent driving status evaluation record includes a description of an association between a past set of radar-visual fusion state vectors and a past set of road network environment quality characteristics, wherein a past radar-visual fusion state vector in the past set of radar-visual fusion state vectors and a past road network environment quality characteristic in the past set of road network environment quality characteristics have an association description, and a past radar-visual fusion state vector refers to a radar-visual fusion state vector corresponding to a past autonomous driving strategy for which a past driving user requested information; Calculate the road network environment quality characteristics of each of the initial autonomous driving strategies through the association description between the radar-visual fusion state vectors corresponding to each of the initial autonomous driving strategies, the past radar-visual fusion state vector set in the intelligent driving state evaluation record, and the past road network environment quality characteristic set; A target autonomous driving strategy is selected from the X initial autonomous driving strategies according to the road network environment quality characteristics of each of the initial autonomous driving strategies, and the target autonomous driving strategy is sent to the target electric snowmobile.

2. The method according to claim 1, characterized in that Each of the initial autonomous driving strategies is composed of one or more strategy task nodes; The performing of radar-visual fusion state vector mining on each of the initial autonomous driving strategies to obtain the radar-visual fusion state vector corresponding to each of the initial autonomous driving strategies includes: Determine any initial autonomous driving strategy among the X initial autonomous driving strategies as a target initial autonomous driving strategy; Obtaining a node depth of each of the policy task nodes in the target initial autonomous driving strategy, and prioritizing one or more policy task nodes in the target initial autonomous driving strategy based on the node depth of each of the policy task nodes to obtain a first policy task node chain; Performing node radar-vision fusion state vector mining on each of the strategy task nodes in the first strategy task node chain to obtain a node radar-vision fusion state vector corresponding to each of the strategy task nodes; Based on the priority order of the one or more policy task nodes in the first policy task node chain, the node radar-vision fusion state vectors corresponding to the one or more policy task nodes are prioritized to obtain a node radar-vision fusion state vector chain; The node radar-visual fusion state vector chain is determined as the radar-visual fusion state vector corresponding to the target initial autonomous driving strategy.

3. The method according to claim 2, characterized in that The performing of node radar-vision fusion state vector mining on each of the strategy task nodes in the first strategy task node chain to obtain the node radar-vision fusion state vector corresponding to each of the strategy task nodes includes: Determine any policy task node in the first policy task node chain as a target policy task node; Performing multi-mode node element identification on the target strategy task node to obtain a multi-mode node element vector of the target strategy task node; A mutual attention operation is performed on the multi-mode node element vector of the target strategy task node to obtain a node radar-vision fusion state vector of the target strategy task node.

4. The method according to claim 3, characterized in that The performing a mutual attention operation on the multi-mode node element vector of the target strategy task node to obtain a node radar-vision fusion state vector of the target strategy task node includes: Based on the task node conduction characteristics of the target strategy task node, a feature connection operation is performed on the multi-mode node element vector of the target strategy task node to obtain the task connection characteristics of the target strategy task node; According to the thunder and lightning weight strengthening rule, thunder and lightning weight strengthening is performed on the task connection characteristics of the target strategy task node to obtain the task connection weighted characteristics of the target strategy task node; The task connection weighted feature of the target strategy task node is determined as the node radar-vision fusion state vector of the target strategy task node.

5. The method according to claim 4, characterized in that The task node conduction feature of the target strategy task node is based on the task node conduction feature of the target strategy task node, and a feature connection operation is performed on the multi-mode node element vector of the target strategy task node to obtain the task connection feature of the target strategy task node, including: Performing a node clustering operation on the target policy task node based on the task node conduction characteristics of the target policy task node to obtain a target node cluster to which the target policy task node belongs; The first intelligent driving control event feature in the multi-mode node element vector of the target strategy task node is updated to a second intelligent driving control event feature used to characterize the in-vehicle human-computer interaction event indicated by the target node cluster, so as to obtain the task connection feature of the target strategy task node.

6. The method according to claim 4, characterized in that According to the thunder vision weight strengthening rule, thunder vision weight strengthening is performed on the task connection feature of the target strategy task node to obtain the task connection weighted feature of the target strategy task node, including: Obtaining the connection features to be processed whose feature attention distribution does not match the thunder vision weight reinforcement rule from the task connection features of the target strategy task node; The attention weight of the connection feature to be processed is adjusted based on the thunder vision weight enhancement rule to obtain a target connection feature; the feature attention distribution of the target connection feature matches the thunder vision weight enhancement rule; The feature set composed of the target connection feature and the remaining connection features is determined as the task connection weighted feature of the target strategy task node; the remaining connection feature refers to the connection features in the task connection features of the target strategy task node except the connection features to be processed.

7. The method according to claim 1, characterized in that The step of processing the acquired driving user request information to generate X initial automatic driving strategies for the driving user request information includes: Receiving driving user request information uploaded by the vehicle server; Performing intelligent driving demand mining on the driving user request information to obtain an intelligent driving demand feature graph of the driving user request information; Mapping and decoding of the strategy vector are performed based on the intelligent driving demand characteristic graph to generate X initial autonomous driving strategies for the driving user request information.

8. The method according to claim 1, characterized in that The calculating of the road network environment quality characteristics of each of the initial autonomous driving strategies by using the radar-visual fusion state vectors corresponding to each of the initial autonomous driving strategies, the association description between the past radar-visual fusion state vector set in the intelligent driving state evaluation record and the past road network environment quality characteristic set includes: Determine any initial autonomous driving strategy among the X initial autonomous driving strategies as a target initial autonomous driving strategy; According to the radar-visual fusion state vector corresponding to the target initial autonomous driving strategy, the set of past radar-visual fusion state vectors in the intelligent driving state evaluation record is roamed; If there is a target past radar-vision fusion state vector in the past radar-vision fusion state vector set that is the same as the radar-vision fusion state vector corresponding to the target initial autonomous driving strategy, the past road network environment quality feature corresponding to the target past radar-vision fusion state vector is determined as the road network environment quality feature of the target initial autonomous driving strategy; If there is no target past radar-vision fusion state vector identical to the radar-vision fusion state vector corresponding to the target initial autonomous driving strategy in the past radar-vision fusion state vector set, the target initial autonomous driving strategy is converted to obtain Y autonomous driving secondary strategies, and based on the Y autonomous driving secondary strategies, the association description between the past radar-vision fusion state vector set in the intelligent driving state evaluation record and the past road network environment quality feature set, the road network environment quality feature of the target initial autonomous driving strategy is determined; Y is a positive integer; The determining of the road network environment quality characteristics of the target initial autonomous driving strategy based on the Y autonomous driving secondary strategies, the association description between the past radar-visual fusion state vector set in the intelligent driving state evaluation record and the past road network environment quality characteristic set includes: Obtaining a local radar-vision fusion state vector corresponding to each of the autonomous driving secondary strategies; Walking the set of past radar-visual fusion state vectors according to the local radar-visual fusion state vectors corresponding to each of the autonomous driving secondary strategies; Determining the road network environment quality characteristics of the target initial autonomous driving strategy based on the wandering results of the past radar-vision fusion state vector set; The target initial autonomous driving strategy is composed of one or more strategy task nodes; the walking result of the past radar vision fusion state vector set is the result of including the target local radar vision fusion state vector in the past radar vision fusion state vector set, and the target local radar vision fusion state vector refers to any local radar vision fusion state vector among Y local radar vision fusion state vectors; The determining of the road network environment quality characteristics of the target initial autonomous driving strategy based on the wandering results of the past radar-vision fusion state vector set includes: Determine the autonomous driving secondary strategy corresponding to the target local radar-vision fusion state vector as the target autonomous driving secondary strategy, and determine the strategy task node included in the target autonomous driving secondary strategy as the strategy task quality labeling node; Determine the strategy task nodes other than the strategy task quality labeling nodes in the target initial autonomous driving strategy as the strategy task nodes to be processed, and obtain the radar vision sensor acquisition data corresponding to the strategy task nodes to be processed; the radar vision sensor acquisition data includes the past road network environment quality characteristics corresponding to the target local radar vision fusion state vector; Determine the node environment quality sub-characteristic corresponding to the strategy task node to be processed according to the radar sensor collection data corresponding to the strategy task node to be processed; The road network environment quality characteristics of the target initial autonomous driving strategy are determined based on the node environment quality sub-characteristics corresponding to the strategy task node to be processed.

9. A radar-vision fusion automatic driving device, characterized in that: The method comprises at least one processor and a memory; the memory stores computer-executable instructions; the at least one processor executes the computer-executable instructions stored in the memory, so that the at least one processor executes the method according to any one of claims 1 to 8.

10. An electric snowmobile, characterized in that: The electric snowmobile is communicatively connected to a radar-vision fusion automatic driving device, and the electric snowmobile is used to receive a target automatic driving strategy issued by the radar-vision fusion automatic driving device, wherein the target automatic driving strategy is determined by the method described in any one of claims 1-8.

Images