Vehicle control method, model training method, and vehicle

Abstract

A vehicle control method, a model training method and a vehicle are disclosed, and relate to the technical field of intelligent driving. In the method, a road surface image of a vehicle during driving is obtained. Based on the similarity between the features of a target obstacle in the road surface image and the features of a reference obstacle, the type of the target obstacle is identified by an obstacle identification model. The features of the reference obstacle are obtained by contrast learning of the reference obstacle in the training data by the obstacle identification model. Subsequently, the vehicle is controlled based on the type of the target obstacle. The obstacle identification model used by the application to identify the target obstacle not only does not rely on fixed training or manual annotation, but also supports unfamiliar road environments and can accurately identify road obstacles, so that the vehicle can make reasonable obstacle avoidance and control decisions.

Description

Technical Field

[0001] This application relates to the field of intelligent driving technology, and in particular to a vehicle control method, a model training method, and a vehicle. Background Technology

[0002] When a vehicle travels over obstacles or damaged road surfaces, it will experience bumps, affecting the passenger experience. Related technologies typically use traditional obstacle recognition algorithms to identify obstacles and employ sensors such as cameras and radar to detect the distance between the vehicle and the vehicle in front. Then, they use corresponding control algorithms to control the vehicle (such as adjusting the suspension stiffness).

[0003] However, traditional obstacle recognition algorithms rely on fixed training or manual annotation, which do not support unfamiliar road environments, resulting in the inability to accurately identify road obstacles and causing vehicles to be unable to make reasonable obstacle avoidance and control decisions. Summary of the Invention

[0004] In view of this, embodiments of this application propose a vehicle control method, a model training method, and a vehicle to improve the above-mentioned problems.

[0005] According to a first aspect of the embodiments of this application, a vehicle control method is provided, the method comprising: acquiring a road surface image in which the vehicle is driving; identifying the type of the target obstacle by means of an obstacle recognition model based on the similarity between the features of the target obstacle in the road surface image and the features of the reference obstacle; obtaining the features of the reference obstacle by means of the obstacle recognition model through comparative learning of the reference obstacle in the training data; and controlling the vehicle based on the type of the target obstacle.

[0006] In one possible implementation, controlling the vehicle based on the type of the target obstacle includes: determining the control execution level of the vehicle based on the type of the target obstacle; and controlling the vehicle to execute control commands corresponding to the control execution level.

[0007] In one possible implementation, the vehicle control method further includes: determining the physical distance between the vehicle and the target obstacle based on a distance prediction model using a road surface image; determining the control duration of the vehicle based on the physical distance and the vehicle speed; the above-mentioned control of the vehicle based on the type of the target obstacle includes: controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle.

[0008] In one possible implementation, the above-mentioned determination of the physical distance between a vehicle and a target obstacle based on a road surface image and a distance prediction model includes: determining the position of the target obstacle in the road surface image and the reference distance from the vehicle to the horizontal centerline of the road surface image using the distance prediction model; and obtaining the physical distance based on the reference distance and the position of the target obstacle in the road surface image.

[0009] In one possible implementation, the vehicle includes a front suspension and a rear suspension, and the control duration includes a first control duration. The above-mentioned control of the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: if the first control duration is greater than a control duration threshold, controlling the front suspension and the rear suspension to execute control commands within the first control duration based on the type of the target obstacle.

[0010] In one possible implementation, the vehicle includes a rear suspension, and the control time includes a first control duration. The aforementioned control of the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: if the first control duration is less than or equal to a control duration threshold, determining a second control duration based on physical distance, the vehicle's wheelbase, and the vehicle's speed; and if the second control duration is greater than the control duration threshold, controlling the rear suspension to execute control commands within the second control duration based on the type of the target obstacle.

[0011] According to a second aspect of the embodiments of this application, a model training method is provided. The method includes: acquiring multiple sets of first images and multiple sets of second images, wherein the first images include a first obstacle and the second images do not include the first obstacle, and the first images and second images are used to characterize road surface information where a vehicle is driving; narrowing the similarity distance between the first obstacle in the multiple sets of first images to obtain a first feature, and widening the similarity distance between the multiple sets of second images and the first obstacle to obtain a second feature; and training a first model based on the first feature and the second feature to obtain an obstacle recognition model.

[0012] In one possible implementation, the above-mentioned acquisition of multiple sets of first images includes: acquiring multiple sets of third images, performing data augmentation on each set of third images in the multiple sets of third images respectively, to obtain multiple sets of first images.

[0013] In one possible implementation, the model training method further includes: dividing each of the multiple sets of first images into at least two parts, and marking the position of the first obstacle in the lower part of the first image; training the second model based on the marked multiple sets of first images to obtain a distance prediction model, which is used to predict the distance between the vehicle and the target obstacle based on the position of the target obstacle in the road image.

[0014] According to a third aspect of the embodiments of this application, a vehicle is provided, the vehicle including one or more processors; the processors are coupled to a memory; the memory stores one or more program instructions, which are executed by the one or more processors to implement the method as described in either the first or second aspect.

[0015] The beneficial effects of the technical solution provided in this application include at least the following: In this application, an obstacle recognition model obtained through comparative learning is used to identify the types of target obstacles in road images. On one hand, this improves the accuracy of obstacle recognition, enabling vehicles to make reasonable obstacle avoidance and control decisions. On the other hand, the learning process of the obstacle recognition model does not rely on manual annotation; by comparing the similarity between samples, it can learn obstacle features with high recognition accuracy, saving data and labor costs. Furthermore, the obstacle recognition model can utilize unlabeled data generated during vehicle movement for comparative learning, covering infinitely long-tail scenarios that are difficult to learn with fixed training (such as manhole covers with rare materials, speed bumps under extreme lighting, etc.), improving the generalization ability of the obstacle recognition model in complex environments.

[0016] It should be understood that the above general description and the following detailed description are exemplary and explanatory only, and do not limit the embodiments of this application. Attached Figure Description

[0017] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application. It is obvious that the drawings described below are merely some embodiments of this application, and those skilled in the art can obtain other drawings based on these drawings without any inventive effort.

[0018] Figure 1 This is a schematic diagram of a vehicle driving scenario provided in an embodiment of this application; Figure 2 This is a schematic flowchart of a vehicle control method provided in an embodiment of this application; Figure 3 This is a schematic diagram of an obstacle feature extraction method provided in an embodiment of this application; Figure 4 This is a schematic flowchart of a model training method provided in an embodiment of this application; Figure 5 This is a schematic diagram of a vehicle control device structure provided in an embodiment of this application; Figure 6 This is a schematic diagram of a model training device structure provided in an embodiment of this application; Figure 7 This is a hardware structure diagram of a vehicle provided in an embodiment of this application. Detailed Implementation

[0019] Exemplary embodiments will now be described more fully with reference to the accompanying drawings. However, these exemplary embodiments can be implemented in many forms and should not be construed as limited to the examples set forth herein; rather, these embodiments are provided to make this application more comprehensive and complete, and to fully convey the concept of the exemplary embodiments to those skilled in the art.

[0020] Furthermore, the described features, structures, or characteristics can be combined in any suitable manner in one or more embodiments. Numerous specific details are provided in the following description to give a thorough understanding of embodiments of this application. However, those skilled in the art will recognize that the technical solutions of this application can be practiced without one or more of the specific details, or other methods, apparatuses, steps, etc., can be employed. In other instances, well-known methods, apparatuses, implementations, or operations are not shown or described in detail to avoid obscuring various aspects of this application.

[0021] Please see Figure 1 , Figure 1 This application provides a schematic diagram of a vehicle driving scenario, as illustrated in an embodiment. Figure 1 As shown in the embodiments of this application, the vehicle 100 can travel on various road surfaces. In some embodiments, during travel, the vehicle 100 may encounter impact obstacles with concave and / or convex features. Of course, it may also encounter impact obstacles with other features, which are not limited in this application.

[0022] In one possible implementation, the application scenario also includes a cloud server 200, and the vehicle 100 and the cloud server 200 communicate via a network. In one embodiment, the cloud server 200 is implemented using a virtual machine.

[0023] In this system, some or all functions of vehicle 100 are controlled by computing platform 150 (or computer system). Computing platform 150 includes at least one processor 151 that executes instructions stored in a non-transitory computer-readable medium such as memory 152. In some embodiments, computing platform 150 comprises multiple computing devices that control individual components or subsystems of vehicle 100 in a distributed manner. Processor 151 can be any conventional processor, such as a central processing unit (CPU). Optionally, processor 151 may also include one or more of the following: a graphics processing unit (GPU), a field-programmable gate array (FPGA), a system-on-chip (SoC), or an application-specific integrated circuit (ASIC).

[0024] Optionally, vehicle 100 can be a means of transportation (such as commercial vehicles, passenger cars, motorcycles, flying cars, trains, etc.), industrial vehicles (such as forklifts, trailers, tractors, etc.), engineering vehicles (such as excavators, bulldozers, cranes, etc.), agricultural equipment (such as lawnmowers, harvesters, etc.), amusement equipment, toy vehicles, etc.; the embodiments of this application do not specifically limit the type of vehicle 100.

[0025] Taking a car as an example, a car generally consists of four major components: engine, chassis, body, and electrical equipment. The engine provides power, the chassis enables driving, the body ensures space and safety, and the electrical equipment coordinates various systems. The integrated design of these four components determines the vehicle's power performance, handling experience, and level of intelligence.

[0026] An engine is used to provide power by converting fuel (e.g., gasoline / diesel) into mechanical energy. It mainly consists of two major structures: the crankshaft and connecting rod mechanism and the valve train, as well as five major systems: the fuel supply system, the cooling system, the lubrication system, the ignition system, and the starting system.

[0027] The chassis is used to transmit power, control driving direction, and ensure braking safety. It mainly includes the transmission system, running system, steering system, and braking system. For example, the transmission system includes a clutch and gearbox, responsible for power transmission. The running system includes the frame, tires, and suspension, supporting the vehicle body and enabling movement. The steering system includes the steering gear, steering wheel, and transmission mechanism, used to control the vehicle's direction. The braking system includes brake discs, anti-lock braking system (ABS), and other standard and parking brakes, ensuring parking safety.

[0028] The vehicle body is typically mounted on the chassis frame and is used for driver and passenger seating and / or cargo loading, providing operating and passenger space.

[0029] Electrical equipment is used to connect and coordinate the other three major components to achieve automated control. It mainly includes power systems (e.g., batteries, generators, etc.), electrical equipment (e.g., lighting, navigation), and power distribution devices.

[0030] In this embodiment of the application, the chassis of the vehicle 100 includes an active suspension. The wheels are connected to the vehicle body through the active suspension. The active suspension is mainly used to transmit the force and torque acting between the wheels and the vehicle body, buffer the impact force transmitted to the vehicle body from the uneven road surface, and attenuate the vibration caused therefrom, so as to ensure that the car can drive smoothly.

[0031] As an example, the vehicle 100 in this embodiment has a built-in vehicle operating system, which includes an intelligent driving system or an autonomous driving system to provide intelligent assisted driving functions or autonomous driving functions. The vehicle 100 can be equipped with various sensors. After the vehicle 100 is powered on, the various sensors work in real time to collect various perception information and provide the perception information to multiple subsystem applications of the vehicle, such as for driving strategy decision-making and comfort function adjustment.

[0032] It should be understood that in this embodiment, the vehicle 100 autonomously controls itself without direct driver intervention. For example, the vehicle 100's own in-vehicle system collects external environmental information and generates control commands based on this information. These control commands can be used to automatically control the vehicle's actuators. For instance, the aforementioned intelligent driving system / autonomous driving system is enabled / activated / started and operates, relying on technologies such as sensors, computer vision, radar, and global positioning systems to enable the vehicle 100 to perceive its surroundings and formulate corresponding control strategies. These control strategies then control the actuators to perform corresponding operations.

[0033] To facilitate understanding, the following description is provided in conjunction with the accompanying drawings and embodiments.

[0034] The vehicle control method of this application embodiment can be implemented by a vehicle control device, which can be an independent device or a... Figure 1The chips or components in the vehicle 100 shown may be software modules that can be deployed on related on-board equipment of the vehicle 100, such as intelligent driving domain control units (e.g., mobile data centers (MDC)), vehicle control units (VCU), vehicle domain controllers (VDC), motor control units (MCU), advanced driver assistance systems (ADAS) domain controllers integrated into the computing platform 150 of the aforementioned vehicle 100, or control units deployed on other components of the vehicle 100. The embodiments of this application do not limit the product form and deployment method of the vehicle control device.

[0035] Figure 2 A schematic flowchart of a vehicle control method provided in an embodiment of this application is shown. Figure 2 As shown, the vehicle control method includes steps S201 to S203.

[0036] In step S201, an image of the road surface where the vehicle is located is acquired while driving.

[0037] For example, an onboard visual perception system can be used to acquire images of the road surface where the vehicle is located during driving. This system utilizes optical devices, a positioning module, and a computing unit to work together to convert the physical road conditions into digital image data. For instance, in practical applications, a front-facing camera or LiDAR can be used to capture real-time images of the road surface in front of the vehicle (e.g., 5 meters ahead). As the vehicle moves forward, the front-facing camera or LiDAR continuously captures images at a rate of tens of frames per second, thus forming a continuous sequence of road surface images.

[0038] It should be understood that road surface images may contain obstacles, such as speed bumps and manhole covers in traffic infrastructure, which are relatively fixed in location but can cause vehicles to bump. Alternatively, they may contain road structural anomalies caused by natural aging or external damage. Common examples include potholes, ruts, subsidence, wave-like bulges, and various cracks. Alternatively, they may contain external road obstacles (foreign objects), which are highly random and temporary external obstacles. These include mudguards, tire fragments, rims, boxes, and road surface flooding or fallen trees that occur during severe weather.

[0039] In step S202, based on the similarity between the features of the target obstacle in the road image and the features of the reference obstacle, the type of the target obstacle is identified by the obstacle recognition model. The features of the reference obstacle are obtained by the obstacle recognition model through comparative learning of the reference obstacles in the training data.

[0040] The obstacle recognition model can detect the presence of obstacles in road images in real time. When an obstacle is present, the model identifies its type. For example, the model extracts features from a road image containing a target obstacle, such as its edges, shape, and texture. Then, it compares these features with previously learned obstacle features to select those with high similarity. The obstacle type identified by these learned features is then used as the obstacle type recognized by the model.

[0041] It should be understood that high feature similarity means that, in a specified feature representation space, the similarity metric between the feature vectors corresponding to two obstacles exceeds a preset threshold (or the distance metric is below a preset threshold), reflecting a high degree of consistency in the attribute dimensions encoded by the two obstacles in that feature representation space. For example, the cosine similarity between the feature vector of the target obstacle and the feature vector of the reference obstacle is calculated to be 0.92. The preset similarity threshold is 0.85. 0.92 exceeds this similarity threshold of 0.85, therefore, the feature similarity between the target obstacle and the reference obstacle is determined to be high.

[0042] By employing a contrastive learning method, the obstacle recognition model learns the features of benchmark obstacles in the training data. Optionally, the training data is divided into two groups: the first group includes benchmark obstacles, and the second group does not; that is, the second group may include other types of obstacles or not include any obstacles at all. The similarity distance between target obstacles in each road surface image of the first group of training data is reduced, while the similarity distance between target obstacles in road surface images of the second group of training data and those of the first group of training data is increased, thereby enabling the obstacle recognition model to learn the features of the benchmark obstacles.

[0043] Optionally, by using a contrastive loss function, the similarity between target obstacles in each road surface image of the first set of training data is increased, thereby narrowing the similarity distance between target obstacles. Conversely, by using a contrastive loss function, the similarity between target obstacles in road surface images of the second set of training data and those in the first set of training data is reduced, thereby widening the similarity distance between road surface images of the second set of training data and target obstacles.

[0044] Among them, the contrastive loss function is a loss function that trains the feature extraction network in the obstacle recognition model by simultaneously increasing the similarity between positive sample pairs (such as the first set of training data mentioned above) and decreasing the similarity between negative sample pairs (such as the second set of training data mentioned above), so that similar samples are clustered and dissimilar samples are dispersed in the feature space.

[0045] In step S203, the vehicle is controlled based on the type of the target obstacle.

[0046] For example, threat assessment is performed based on the type of obstacle identified (such as static roadblocks, dynamic pedestrians, or sudden hazards), and then differentiated obstacle avoidance strategies such as path replanning, predictive detours, or emergency braking are selected. For static roadblocks (such as cones or construction barriers), the fixed position and lane encroachment of the roadblock are assessed, and if deemed a low dynamic threat, path replanning or slow detour strategies can be adopted. For dynamic pedestrians, the probability of a pedestrian entering the vehicle's trajectory is assessed based on the pedestrian's direction of movement, speed, and relative distance to the vehicle, and if deemed a medium to high dynamic threat, predictive detours and continuous tracking are required. For sudden hazards (such as a vehicle suddenly stopping ahead or an obstacle suddenly appearing), they are determined to be an emergency threat based on the instantaneous rate of change, triggering the highest priority response. Subsequently, the decision-making level translates these strategies into control commands for a smooth trajectory that balances safety and comfort. The underlying control system converts the control commands into signals such as throttle, brakes, and suspension adjustments to ensure that the vehicle completes the avoidance maneuver smoothly and safely.

[0047] The vehicle control method provided in this application utilizes an obstacle recognition model obtained through comparative learning to identify target obstacles in road images. On one hand, this improves the accuracy of obstacle recognition. On the other hand, the learning process of the obstacle recognition model does not rely on manual annotation; by comparing the similarity between samples, it can learn obstacle features with high recognition accuracy, saving data and labor costs. Furthermore, the obstacle recognition model can utilize unlabeled data generated during vehicle movement for comparative learning, covering infinitely long-tail scenarios that are difficult to learn with fixed training (such as manhole covers with rare materials, speed bumps under extreme lighting, etc.), thus improving the generalization ability of the obstacle recognition model in complex environments.

[0048] In one possible implementation, the vehicle's control execution level is determined based on the type of the target obstacle. The vehicle then executes control commands corresponding to that level. Different types of target obstacles correspond to different execution levels for the actuators. The vehicle performs a risk assessment based on the obstacle type to determine the corresponding control execution level. For example, for low-risk static obstacles (such as distant construction cones, road signs, or speed bumps), the determined control execution level is the primary level, where strategies might include warnings, slight deceleration, and adjusting suspension stiffness. For obstacles with some dynamic uncertainty (such as a normally moving vehicle ahead or a non-motorized vehicle on the roadside), the trajectory prediction results are assessed. If a collision risk exists, the level is raised to intermediate control execution, employing active deceleration or path fine-tuning. When facing extremely high-risk sudden obstacles (such as a pedestrian suddenly crossing or a vehicle braking abruptly), the highest level of control execution is determined, prioritizing collision avoidance.

[0049] After determining the vehicle's control execution level, this level is mapped to specific vehicle control commands. For example, at the basic control execution level, the vehicle might output audible and visual warnings to alert the driver, or the suspension might soften. At the intermediate control execution level, control is gained over the vehicle to generate a smooth avoidance trajectory, allowing the vehicle to safely maneuver around obstacles. At the advanced control execution level, emergency braking is initiated.

[0050] For example, suppose the controlled object is the vehicle's suspension. Within the same suspension control system, the suspension's execution level is tiered based on differences in control bandwidth, energy input method, and response speed. When road conditions are normal and obstacles are common manhole covers, a low control execution level is determined. The suspension is controlled primarily using electro-hydraulic actuators connected in series with springs, intervening in a low-frequency range below 3 Hz, focusing on eliminating macroscopic attitude changes such as vehicle pitch and roll. Conversely, when road conditions change abruptly, with multiple potholes in front of the vehicle, a high control execution level is determined. The suspension actuators take over all movement between sprung and unsprung mass, achieving control up to 10-15 Hz or even wider frequency bands, reducing high-frequency vibrations when the vehicle is driving on the road.

[0051] It is easy to understand that the controlled objects also include the engine and powertrain, intelligent cockpit and comfort systems, etc., and this application does not limit this. For example, the execution level of the engine and powertrain can be divided into steps according to the type of obstacle perceived. When the detected obstacle is a low-risk target such as a small bump in the road surface, a speed bump, or a slow-moving flow of traffic ahead, a low control execution level is determined, and the engine and powertrain maintain an economical and smooth mode. When the detected obstacle is a high-risk target such as a pedestrian suddenly crossing, a large fallen obstacle, or a continuous pothole, a high control execution level is determined. At this time, the engine and powertrain control system takes over torque management, dynamically adjusts the torque distribution between the front and rear axles, and activates traction control and torque vectoring control to ensure that the vehicle obtains driving force in emergency obstacle avoidance or low-traction road surfaces.

[0052] Furthermore, the execution level of the intelligent cockpit and comfort system can be tiered according to the type of obstacle. When the identified obstacle is a construction cone, a regular manhole cover, or a road sign, a low control execution level is determined. In this case, the intelligent cockpit and comfort system operates in energy-saving mode, the air conditioning maintains a constant temperature and low airflow, the active noise cancellation system only cancels steady-state low-frequency road noise, and the seat and ambient lighting system maintains basic output, focusing on maintaining basic occupant comfort and energy consumption balance. When the identified obstacle is a dense pothole area or a road surface with large elevation changes, a high control execution level is determined. In this case, the intelligent cockpit and comfort system actively enhances comfort intervention. For example, the air conditioning automatically switches to internal circulation and increases airflow to filter dust, the active noise cancellation system dynamically emits reverse sound waves to cancel broadband road noise, and the seat side wings automatically clamp to fix the occupant's posture, thereby reducing the impact and discomfort caused by the obstacle.

[0053] In this implementation, the vehicle's control execution level is determined based on the type of the target obstacle, which can effectively reduce the false alarm rate of obstacles and the false trigger rate of vehicle control, enabling the vehicle to take reasonable countermeasures when facing different obstacles and improve ride comfort.

[0054] In one possible implementation, the vehicle control method further includes: determining the physical distance between the vehicle and the target obstacle based on a target distance prediction model using a road surface image; determining the control duration for controlling the vehicle based on the physical distance and the vehicle speed; and controlling the vehicle based on the type of the target obstacle, including: controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle.

[0055] For example, a vehicle acquires a road surface image and inputs it into a target distance prediction model. This model detects target obstacles in the road surface image and determines the physical distance between the vehicle and the target obstacle based on the road surface image containing the obstacle. After determining the physical distance, the vehicle's control duration is determined based on the physical distance and the vehicle's current speed. Then, based on the type of target obstacle, the vehicle executes control commands within the control duration.

[0056] It is easy to understand that the control time required for the vehicle to reach the target obstacle can be calculated based on the physical distance and the current vehicle speed. Within this time, a control signal is sent to the vehicle in advance, and the vehicle can perform corresponding operations in advance to deal with unpredictable dangers or discomforts that the target obstacle may bring to the passengers.

[0057] In this implementation, the control time of the vehicle is determined based on the physical distance between the vehicle and the target obstacle. The required vehicle control time under the current road conditions can be calculated in advance, giving the vehicle more time to perform actions such as obstacle avoidance, deceleration, or detour.

[0058] Optionally, the above-mentioned determination of the physical distance between the vehicle and the target obstacle based on the road surface image and the distance prediction model includes: determining the reference distance from the vehicle to the horizontal centerline of the road surface image and the position of the target obstacle in the road surface image using the distance prediction model; and obtaining the physical distance based on the reference distance and the position of the target obstacle in the road surface image.

[0059] It should be understood that the reference distance from the vehicle to the horizontal centerline of the road image refers to the longitudinal distance (i.e., the straight-line distance in the direction of the vehicle's movement) from a reference point on the vehicle to the horizontal line in the real world corresponding to the horizontal centerline of the image. The horizontal line in the real world refers to the horizontal line projected onto the actual road surface from the horizontal line at the very center of the road image.

[0060] For example, the distance prediction model constructs a spatial mapping relationship between the road image pixel coordinate system and the real-world physical coordinate system using intrinsic parameters obtained from the vehicle's camera calibration, distortion coefficients, and measured extrinsic parameters (such as camera mounting height and pitch angle). To eliminate road image distortion caused by lens distortion, the original road image is dedistorted using intrinsic parameters and distortion coefficients. Subsequently, in the dedistorted road image, the center point of the bottom edge of the road image is selected as the reference point directly in front of the vehicle. The pixel coordinates of this reference point are combined with the camera's intrinsic parameters to convert them into a direction vector in the camera coordinate system. Then, combining the extrinsic parameters and the ground plane assumption, the mapping relationship between each row of pixels in the image and the actual longitudinal distance from the vehicle to the corresponding road position is established using the principle of similar triangles. When it is necessary to determine the reference distance from the vehicle to the horizontal centerline of the road image, the row number of the pixel corresponding to the horizontal centerline of the road image is substituted into the established mapping relationship to solve for the reference distance from the vehicle's current position to the road image in the real three-dimensional world coordinate system.

[0061] A distance prediction model is used to perform real-time inference on the road surface image, outputting a two-dimensional bounding box of the target obstacle. Subsequently, the distance prediction model extracts the center point of the bottom of the two-dimensional bounding box as the projected coordinates of the target obstacle's ground point, thereby determining the position of the target obstacle in the road surface image.

[0062] After determining the baseline distance from the vehicle to the horizontal centerline of the road image, and the position of the target obstacle in the road image, the distance deviation value is calculated using the distance of the target obstacle's position in the road image relative to the horizontal centerline. The actual physical distance from the vehicle to the target obstacle is then obtained by summing the distance deviation value and the baseline distance from the vehicle to the horizontal centerline. For example, first calculate the distance L_normal from the vehicle to the horizontal centerline of the road image, and then calculate the distance deviation value L_dis using the distance of the target obstacle's position in the road image relative to the horizontal centerline. The physical distance between the vehicle and the target obstacle is then calculated by summing L = L_normal + L_dis.

[0063] In this implementation, on the one hand, the distance prediction model is used to determine the position of obstacles in the road image, which can effectively make up for the shortcomings of a single high-precision map and enhance the perception of non-fixed and low-lying long-tailed obstacles. On the other hand, the physical distance is calculated based on the geometric relationship of the horizontal centerline of the road image and the position of the obstacle in the road image. Compared with deep learning distance prediction technology, this method has the advantage of being lightweight and does not require a lot of computing power.

[0064] In one possible implementation, the vehicle includes a front suspension and a rear suspension, and the aforementioned control duration includes a first control duration. Controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: if the first control duration exceeds a control duration threshold, controlling the front and rear suspension to execute control commands within the first control duration based on the type of the target obstacle.

[0065] It should be understood that the control duration threshold can be a dynamic delay compensation. Dynamic delay compensation refers to the objective hardware response lag that exists in actual vehicle operation; that is, the actual time spent from the moment the control system issues a control command until the mechanical components inside the suspension (such as the opening of solenoid valves, the establishment of air pressure, etc.) actually begin to produce physical actions. Alternatively, the control duration threshold can also be a pre-set threshold, which is greater than the aforementioned dynamic delay compensation, to ensure that the command for one control cycle is completely and accurately executed.

[0066] For example, if the control duration threshold is dynamic delay compensation, and the dynamic delay compensation duration is 0.06 seconds (s), the first control duration determined based on the physical distance between the vehicle and the target obstacle and the current vehicle speed is 1 second. The control system sends a control command to the suspension. After the 0.06-second dynamic delay compensation duration, the suspension cannot perform the corresponding operation. Then, after the 0.06-second duration, the suspension begins to perform the corresponding operation to deal with the target obstacle in front of the vehicle.

[0067] If the initial control duration is 0.08s, the dynamic delay compensation duration of the suspension is 0.05s. However, due to potential instability in the control system (e.g., conflicts between controller computing power and task scheduling), the dynamic delay compensation duration may increase, for example, to 0.1s. After the control system sends a signal to the suspension, it must first undergo the 0.1s dynamic delay compensation period. However, after 0.1s, the vehicle may have already passed the obstacle (the initial control duration is 0.08s, which is also the time it takes for the vehicle to pass the obstacle from its current position). At this point, the suspension no longer needs to perform any action. If the suspension then performs the corresponding action, it may cause confusion in the control strategy, affecting the passenger's riding experience. Therefore, a control duration threshold greater than the dynamic delay compensation duration can be preset, such as 0.12s. It's easy to understand that if the initial control duration is 0.08s and the time is less than the preset control duration threshold of 0.12s, the suspension will not be controlled to perform any action.

[0068] In this implementation, the control strategies of the front and rear suspensions are dynamically adjusted by comparing the control duration and the control duration threshold. This can upgrade the traditional passive response to active prediction, calculate sufficient control duration in advance, and give the front and rear suspensions more time to perform corresponding actions when passing through obstacles, thereby reducing the discomfort caused to passengers when the vehicle passes through or avoids obstacles.

[0069] In one possible implementation, the vehicle includes a rear suspension, and the aforementioned control duration includes a first control duration. Controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: if the first control duration is less than or equal to a control duration threshold, determining a second control duration based on physical distance, the vehicle's wheelbase, and the vehicle's speed; and if the second control duration is greater than the control duration threshold, controlling the rear suspension to execute control commands within the second control duration based on the type of the target obstacle.

[0070] If the initial control duration is less than or equal to the control duration threshold, it can be considered that the front suspension is unable to perform the corresponding action when the vehicle's front wheels pass the obstacle. In this case, the system then assesses whether the rear suspension can perform the corresponding action when the rear wheels pass the obstacle. The physical distance from the vehicle to the target obstacle plus the vehicle's wheelbase equals the physical distance from the rear wheels to the target obstacle. Dividing this distance by the current vehicle speed gives the time it takes for the rear wheels to pass the target obstacle. If the time taken for the rear wheels to pass the target obstacle exceeds the control duration threshold, the rear suspension can perform the corresponding action. In this case, the system controls the rear suspension to perform the appropriate operation to address the target obstacle in front of the vehicle.

[0071] For example, the physical distance between the current vehicle and the target obstacle is 1 meter (m), the vehicle's wheelbase is 3m, the current vehicle speed is 5m / s, the calculated second control duration is 0.8s, that is, the time it takes for the vehicle's rear wheels to reach the target obstacle is 0.8s, the control duration threshold is 0.3s, at this time the control system judges that the time it takes for the vehicle's rear wheels to reach the target obstacle is greater than the control duration threshold, and then sends a control signal to the suspension to control the suspension to soften.

[0072] In this implementation, the control strategy of the rear suspension is dynamically adjusted by comparing the control duration and the control duration threshold. This allows for the calculation of sufficient control duration in advance when the front suspension does not have enough time to perform the action, thereby minimizing the discomfort caused to passengers when the vehicle passes over or avoids obstacles.

[0073] Figure 3 A schematic flowchart of a model training method provided in an embodiment of this application is shown. Figure 3 As shown, the model training method includes steps S301 to S303.

[0074] In step S301, multiple sets of first images and multiple sets of second images are acquired. The first images include a first obstacle, and the second images do not include the first obstacle. The first images and the second images represent the road surface information where the vehicle is located when driving.

[0075] In one possible implementation, road surface images are acquired under various road conditions and driving speeds. These images represent the road surface information the vehicle is on while driving. For the same road surface, the images are grouped according to whether they contain an obstacle to be detected. Images containing the obstacle are used as positive sample datasets (multiple sets of first images), while images without the obstacle are used as negative sample datasets (multiple sets of second images).

[0076] For example, road images containing obstacles to be detected are filtered out, and these images are used as the positive sample dataset. The obstacles to be detected can be speed bumps, manhole covers, etc. Road images without obstacles to be detected are used as the negative sample dataset. It should be understood that images without obstacles to be detected refer to road images without obstacles, or road images containing obstacles of a different type than the type of obstacle to be detected.

[0077] Optionally, negative samples can be processed to reduce the occurrence of false and simple negative samples. False negative samples may affect accurate predictions, for example, speed bumps may be marked as ordinary road surfaces. Simple negative samples may prevent the model from learning the essential features of obstacles, such as an excessively high proportion of pure road background data. In addition, noise can be labeled onto negative samples to prevent maintenance signs on the road surface from being misidentified as obstacles.

[0078] In one possible implementation, obtaining multiple sets of first images includes: obtaining multiple sets of third images, performing data augmentation on each set of third images to obtain multiple sets of first images.

[0079] To enhance the robustness of the obstacle recognition model to complex environmental factors, data augmentation strategies can be applied to road surface images containing obstacles to be detected. This involves altering the grayscale distribution of the road surface images to simulate different weather conditions. In other words, based on an original road surface image, road surface images with different grayscale levels are obtained. For example, grayscale transformations can be applied to simulate scenarios such as cloudy days, fog, rain, snow, or low light, ensuring that the different road surface images after data augmentation maintain a high degree of similarity in the feature space. This allows the obstacle recognition model to learn essential semantic features independent of weather changes. Subsequently, the similarity distance between obstacles in multiple sets of data-augmented road surface images is reduced. The method for reducing the similarity distance can be found in the embodiment in step S302, and will not be elaborated further here.

[0080] By augmenting road surface images containing target obstacles, the scale and diversity of the road surface image dataset are expanded. This allows the model to be trained on a richer dataset of road surface images, effectively improving the obstacle recognition model's adaptability and recognition accuracy to new road surface image data.

[0081] In step S302, the similarity distance between the first obstacles in multiple sets of first images is reduced to obtain a first feature; the similarity distance between the first obstacles in multiple sets of second images is increased to obtain a second feature.

[0082] For each road surface image, an encoder extracts a feature vector corresponding to an obstacle, which is then L2 normalized (the feature vector is divided by the L2 norm (i.e., the Euclidean length) to make the modulus 1). Next, a similarity matrix is ​​calculated between each pair of road surface images, where each element represents the cosine similarity of the corresponding sample pair. A higher cosine similarity value indicates that the obstacles in the two road surface images are closer in the feature space. Positive and negative sample datasets are determined based on predefined semantic judgment rules (e.g., the first image has a positive sample label, and the second image has a negative sample label). Then, a contrastive loss function is used to increase the similarity of positive sample pairs in the positive sample dataset and decrease the similarity of negative sample pairs. The contrastive loss function is as follows:

[0083] in, For the first i Features of a road surface image The characteristics of positive sample pairs are given, where τ is the temperature coefficient. 2N The total number of samples, i and j To enhance the index of samples in the subsequent batch (ranging from 1 to 2N).

[0084] By calculating the similarity between obstacles in different road surface images, the model brings the features corresponding to obstacles in different road surface images in positive samples closer together, while pulling the features corresponding to different road surface images in negative samples further apart. Finally, the feature representations corresponding to the positive sample dataset and the feature representations corresponding to the negative sample dataset are obtained.

[0085] For example, see Figure 4 The upper set of road surface images represents positive samples, and the lower set represents negative samples. In positive samples, obstacles to be detected are represented by triangles, while in negative samples, obstacles of a different type than those in the positive samples are represented by ellipses. Road surface images in negative samples may not contain obstacles.

[0086] Continue reading Figure 4 Taking three road surface images "Z1, Z2, Z3" from the positive sample as examples, the similarity between obstacles in these three road surface images "Z1, Z2, Z3" is calculated respectively, and the similarity distance between the obstacles (represented by triangles) in these three road surface images "Z1, Z2, Z3" is increased. Taking two road surface images from the negative sample as examples, the similarity between the obstacles (represented by ellipses) in the road surface image "F1" from the negative sample and the obstacles in the three road surface images "Z1, Z2, Z3" from the positive sample is calculated respectively, and the similarity distance between the obstacles in the road surface image "F1" from the negative sample and the obstacles in the three road surface images "Z1, Z2, Z3" from the positive sample is increased. Alternatively, the similarity between the road surface image "F2" from the negative sample (excluding obstacles) and the obstacles in the three road surface images "Z1, Z2, Z3" from the positive sample is calculated, and the similarity distance between the road surface image "F2" from the negative sample (excluding obstacles) and the obstacles in the three road surface images "Z1, Z2, Z3" from the positive sample is increased.

[0087] Optionally, obstacle noise is added to the multiple sets of second images before widening the similarity distance between them and the first obstacle. For negative samples, i.e., multiple sets of second images, obstacle noise can be randomly added to the road surface image corresponding to the negative sample. Obstacle noise includes, but is not limited to, local occlusion blocks, irregular masks simulating mud splashes, sensor noise, or specular reflection artifacts. After adding obstacle noise to the negative samples, the similarity distance between the negative samples with added obstacle noise and the obstacles in the positive samples is widened. Since negative samples should maintain a distance from positive samples in terms of similarity, the additional obstacle noise added to the negative samples further widens the difference in obstacle features between the negative and positive samples, enabling the obstacle recognition model to not only distinguish different types of targets but also ignore task-irrelevant stray interference. Through the obstacle noise addition strategy, the generalization performance of the obstacle recognition model in different road surface environments is improved.

[0088] In step S303, a first model is trained based on the first feature and the second feature to obtain an obstacle recognition model.

[0089] After obtaining the feature representations corresponding to the positive and negative sample datasets in step S302, the model performs backpropagation on these feature representations. Each backpropagation calculates a backpropagation gradient, and the model parameters are updated based on this gradient. Each backpropagation and parameter update brings the features corresponding to obstacles in the positive sample pairs closer to the current values, while moving the features in the negative sample pairs further away from the features corresponding to obstacles in the positive sample pairs. After multiple fine-tunings, the model parameters eventually converge to a stable state with a low loss value. At this point, the model has learned to bring the features corresponding to obstacles in the positive samples closer together and to move the features corresponding to obstacles in the negative samples further away from the features corresponding to obstacles in the positive samples. This allows the obstacle recognition model to be trained.

[0090] In this embodiment, by narrowing the similarity distance between obstacles with high similarity and widening the similarity distance between obstacles with low similarity, the quality of extracted obstacle features is improved, resulting in higher accuracy of the obstacle recognition model. Furthermore, the training process of this method does not rely on manual annotation; by narrowing the similarity between compared samples, obstacle features with high recognition accuracy can be learned, saving data and labor costs. Moreover, the obstacle recognition model can utilize unlabeled data generated during vehicle movement for comparative learning, covering infinitely long-tail scenarios that are difficult to learn from with fixed training (such as manhole covers with rare materials, speed bumps under extreme lighting, etc.), thus improving the generalization ability of the obstacle recognition model in complex environments.

[0091] In one possible implementation, each of the multiple sets of first images is divided into at least two parts, upper and lower, and the position of the target obstacle is marked on the lower part of the first image. Based on the marked sets of first images, a second model is trained to obtain a distance prediction model. The distance prediction model is used to predict the distance between the vehicle and the target obstacle based on the position of the target obstacle in the road surface image.

[0092] It should be understood that the upper part of a road image may include elements such as sky, buildings, or treetops. This information contributes little to determining the distance to obstacles on the road and may even introduce noise. Obstacles are labeled in the lower part of the road image because this portion covers the road surface the vehicle is traveling on. This lower part is a crucial perception area for the vehicle to make emergency obstacle avoidance and collision avoidance decisions. The lower half of the road image directly reflects obstacles within the currently drivable area. By focusing attention on the lower part of the road image, interference from the environmental background can be effectively eliminated, allowing the model to concentrate on analyzing obstacles in front of the vehicle that may affect driving.

[0093] For example, each road surface image containing the target obstacle is divided into at least two parts, upper and lower. After obtaining the lower part of the road surface image, the target obstacle in the road surface image is labeled. The specific location of the target obstacle is marked in the lower part of the road surface image. For example, a 2D bounding box localization method that closely follows the actual outline of the object is used to label the specific location of the target obstacle. During labeling, an axis-aligned rectangle that closely follows the actual outline of the obstacle is assigned to each obstacle to accurately define the object boundary. Subsequently, based on the labeled road surface images, the model is trained to obtain a distance prediction model. The distance prediction model can determine the position of the target obstacle in a newly input road surface image.

[0094] During training, the model continuously attempts to learn the non-linear relationship between road image features and distance values. Through multiple iterative calculations, the model's internal parameters are optimized, enabling it to understand the size, position, and spatial meaning of texture changes representing obstacles in the image. After multiple training and validation iterations, a well-trained distance prediction model can be obtained.

[0095] It should be understood that distance prediction models are used to predict the distance between a vehicle and a target obstacle based on the target obstacle's position in a road surface image.

[0096] First, a reference distance from the vehicle to the horizontal centerline of the road image is determined using a distance prediction model. Then, the distance deviation is calculated based on the distance of the target obstacle's position in the road image relative to the horizontal centerline. Finally, the actual physical distance from the vehicle to the target obstacle is obtained by summing the distance deviation and the reference distance from the vehicle to the horizontal centerline of the road image. The method for determining the reference distance from the vehicle to the horizontal centerline of the road image can be found in the embodiment of step S203, and will not be repeated here.

[0097] In this implementation, dividing the image into at least upper and lower parts and labeling obstacle locations in the lower half improves the efficiency of distance calculation. Since road surfaces and obstacles in a vehicle's driving environment are mainly concentrated in the lower half of the field of view, this strategy effectively filters out interference from irrelevant background areas such as the sky and distant buildings, significantly reducing the amount of computation on invalid data. Furthermore, focusing on labeling key areas not only reduces manual labeling costs but also guides the model to more effectively learn the pixel features and spatial distribution patterns of target obstacles in the road image, thereby improving the accuracy and robustness of the distance prediction model in complex road scenarios.

[0098] The above embodiments describe in detail the vehicle control method provided by the embodiments of this application. In other embodiments, this application also provides a vehicle control device 500. Figure 5 This is a schematic diagram of the structure of a vehicle control device provided in an embodiment of this application, as shown below. Figure 5 As shown, the vehicle control device 500 includes: The acquisition module 510 is used to acquire images of the road surface where the vehicle is driving.

[0099] The recognition module 520 is used to identify the type of target obstacle based on the similarity between the features of the target obstacle in the road image and the features of a reference obstacle, using an obstacle recognition model. The features of the reference obstacle are obtained by the obstacle recognition model through comparative learning of reference obstacles in the training data.

[0100] Control module 530 is used to control the vehicle based on the type of target obstacle.

[0101] In one possible implementation, the control module 530 is used to determine the control execution level of the vehicle based on the type of the target obstacle; and to control the vehicle to execute control commands corresponding to the control execution level.

[0102] In one possible implementation, the vehicle control device 500 further includes a distance determination module, which determines the physical distance between the vehicle and the target obstacle based on a road surface image and a distance prediction model; and determines the control duration of the vehicle based on the physical distance and the vehicle speed. The control module 530 is also used to control the vehicle to execute control commands within the control duration based on the type of the target obstacle.

[0103] In one possible implementation, the aforementioned distance determination module is used to determine the position of the target obstacle in the road image and the reference distance from the vehicle to the horizontal centerline of the road image through a distance prediction model; and to obtain the physical distance based on the reference distance and the position of the target obstacle in the road image.

[0104] In one possible implementation, the control module 530 is used to control the front and rear suspensions of the vehicle to execute control commands within the first control duration, based on the type of the target obstacle, if the first control duration is longer than a control duration threshold.

[0105] In another possible implementation, the control module 530 is used to determine a second control duration based on physical distance, vehicle wheelbase, and vehicle speed when the first control duration is less than or equal to a control duration threshold; and to control the vehicle's rear suspension to execute control commands within the second control duration based on the type of the target obstacle when the second control duration is greater than the control duration threshold.

[0106] In other embodiments, this application also provides a model training apparatus 600. Figure 6 This is a schematic diagram of a model training device structure provided in an embodiment of this application, as shown below. Figure 6 As shown, the model training device 600 includes: The acquisition module 610 is used to acquire multiple sets of first images and multiple sets of second images. The first images include a first obstacle, and the second images do not include the first obstacle. The first images and the second images are used to characterize the road surface information where the vehicle is located when driving.

[0107] The feature extraction module 620 is used to narrow the similarity distance between the first obstacles in multiple sets of first images to obtain the first feature, and to widen the similarity distance between the first obstacles in multiple sets of second images to obtain the second feature.

[0108] The first training module 630 is used to train the first model based on the first feature and the second feature to obtain the obstacle recognition model.

[0109] In one possible implementation, the acquisition module 610 is used to acquire multiple sets of third images, and perform data augmentation on each set of third images to obtain multiple sets of first images.

[0110] In one possible implementation, the model training device 600 further includes a second training module, which is used to divide each of the multiple sets of first images into at least two parts, and to mark the position of the first obstacle in the lower part of the first image; and to train the second model based on the marked multiple sets of first images to obtain a distance prediction model, which is used to predict the distance between the vehicle and the target obstacle based on the position of the target obstacle in the road image.

[0111] It should be understood that the above Figure 5 or Figure 6The provided device, in implementing its functions, is only illustrated by the division of the above-described functional modules. In practical applications, the functions can be assigned to different functional modules as needed, that is, the internal structure of the device can be divided into different functional modules to complete all or part of the functions described above. Furthermore, the device and method embodiments provided in the above embodiments belong to the same concept, and their specific implementation processes are detailed in the method embodiments, and will not be repeated here.

[0112] According to one aspect of the embodiments of this application, a vehicle 700 is also provided, such as Figure 7 As shown, the vehicle 700 includes one or more processors 710. The processors 710 are coupled to a memory 720. The memory 720 stores one or more program instructions, which are executed by the one or more processors 710 to perform the aforementioned... Figure 2 or Figure 3 The methods provided.

[0113] Furthermore, the processor 710 may include one or more processing cores. The processor 710 runs or executes instructions, programs, code sets, or instruction sets stored in the memory 720, and retrieves data stored in the memory 720. Optionally, the processor 710 may be implemented using at least one hardware form selected from Digital Signal Processing (DSP), Field-Programmable Gate Array (FPGA), and Programmable Logic Array (PLA). The processor 710 may integrate one or a combination of several of the following: Central Processing Unit (CPU), Graphics Processing Unit (GPU), and modem. The CPU primarily handles the operating system, user interface, and applications; the GPU is responsible for rendering and drawing the displayed content; and the modem handles wireless communication. It is understood that the modem may also not be integrated into the processor and may be implemented using a separate communication chip.

[0114] According to one aspect of this application, a computer-readable storage medium is also provided, which may be included in the cloud server described in the above embodiments; or it may exist independently and not assembled into the cloud server. The aforementioned computer-readable storage medium carries computer-readable instructions that, when executed by a processor, implement the methods in any of the above embodiments.

[0115] It should be noted that the computer-readable medium shown in the embodiments of this application can be a computer-readable signal medium or a computer-readable storage medium, or any combination of the two. Computer-readable storage media can be, for example, but not limited to: electrical, magnetic, optical, electromagnetic, infrared, or semiconductor systems, apparatuses, or devices, or any combination thereof. More specific examples of computer-readable storage media may include, but are not limited to: electrical connections with one or more wires, portable computer disks, hard disks, random access memory (RAM), read-only memory (ROM), erasable programmable read-only memory (EPROM), flash memory, optical fiber, portable compact disc read-only memory (CD-ROM), optical storage devices, magnetic storage devices, or any suitable combination thereof. In this application, a computer-readable storage medium can be any tangible medium containing or storing a program that can be used by or in conjunction with an instruction execution system, apparatus, or device. In this application, a computer-readable signal medium can include data signals propagated in baseband or as part of a carrier wave, carrying computer-readable program code. Such transmitted data signals can take various forms, including but not limited to electromagnetic signals, optical signals, or any suitable combination thereof. The computer-readable signal medium can also be any computer-readable medium other than a computer-readable storage medium, which can send, propagate, or transmit a program for use by or in connection with an instruction execution system, apparatus, or device. The program code contained on the computer-readable medium can be transmitted using any suitable medium, including but not limited to wireless, wired, etc., or any suitable combination thereof.

[0116] The units described in the embodiments of this application can be implemented in software or hardware, and the described units can also be located in a processor. The names of these units do not necessarily limit the specific unit itself.

[0117] It should be noted that all information (including but not limited to user device information, user personal information, etc.), data (including but not limited to data used for analysis, stored data, displayed data, etc.), and signals involved in this application have been authorized by the user or fully authorized by all parties, and the collection, use, and processing of related data must comply with the relevant laws, regulations, and standards of the relevant countries and regions. For example, the object information and detailed information involved in this application were obtained with full authorization.

[0118] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the embodiments disclosed herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein.

[0119] It should be understood that this application is not limited to the precise structure described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A vehicle control method, characterized in that, The method includes: Acquire images of the road surface where the vehicle is located while driving; Based on the similarity between the features of the target obstacle in the road image and the features of the reference obstacle, the type of the target obstacle is identified by an obstacle recognition model; the features of the reference obstacle are obtained by the obstacle recognition model through comparative learning of the reference obstacles in the training data. The vehicle is controlled based on the type of the target obstacle.

2. The method according to claim 1, characterized in that, Controlling the vehicle based on the type of the target obstacle includes: The control execution level of the vehicle is determined based on the type of the target obstacle; Control the vehicle to execute control commands corresponding to the control execution level.

3. The method according to claim 1 or 2, characterized in that, The method further includes: Based on the road surface image, the physical distance between the vehicle and the target obstacle is determined by a distance prediction model; The control duration of the vehicle is determined based on the physical distance and the vehicle speed. Controlling the vehicle based on the type of the target obstacle includes: Based on the type of the target obstacle, the vehicle is controlled to execute control commands within the control duration.

4. The method according to claim 3, characterized in that, The step of determining the physical distance between the vehicle and the target obstacle based on the road surface image using a distance prediction model includes: The distance prediction model is used to determine the position of the target obstacle in the road image, as well as the reference distance from the vehicle to the horizontal centerline of the road image; The physical distance is obtained based on the reference distance and the position of the target obstacle in the road surface image.

5. The method according to claim 3 or 4, characterized in that, The vehicle includes a front suspension and a rear suspension, the control duration includes a first control duration, and controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: If the first control duration exceeds the control duration threshold, the front suspension and the rear suspension are controlled to execute control commands within the first control duration based on the target obstacle type.

6. The method according to claim 3 or 4, characterized in that, The vehicle includes a rear suspension, the control time includes a first control duration, and controlling the vehicle to execute control commands within the control duration based on the type of the target obstacle includes: If the first control duration is less than or equal to the control duration threshold, the second control duration is determined based on the physical distance, the wheelbase of the vehicle, and the vehicle speed. If the second control duration exceeds the control duration threshold, the rear suspension is controlled to execute control commands within the second control duration based on the type of the target obstacle.

7. A model training method, characterized in that, The method includes: Acquire multiple sets of first images and multiple sets of second images, wherein the first images include a first obstacle and the second images do not include the first obstacle, and the first images and the second images represent the road surface information where the vehicle is located when driving; By narrowing the similarity distance between the first obstacles in the multiple sets of first images, a first feature is obtained; by widening the similarity distance between the multiple sets of second images and the first obstacles, a second feature is obtained. The first model is trained based on the first feature and the second feature to obtain the obstacle recognition model.

8. The method according to claim 7, characterized in that, The acquisition of multiple sets of first images includes: Multiple sets of third images are acquired, and data augmentation is performed on each set of third images to obtain the multiple sets of first images.

9. The method according to claim 7 or 8, characterized in that, The method further includes: Each of the multiple sets of first images is divided into at least two parts, upper and lower, and the position of the first obstacle is marked in the lower part of the first image; Based on multiple labeled first images, a second model is trained to obtain a distance prediction model. The distance prediction model is used to predict the distance between a vehicle and a target obstacle based on the position of the target obstacle in the road surface image.

10. A vehicle, characterized in that, It includes one or more processors; the processors are coupled to a memory; the memory stores one or more program instructions, which are executed by the one or more processors to implement the method as described in any one of claims 1-6 or 7-9.

Images