Road obstacle recognition system based on image processing

Abstract

The invention discloses a road obstacle recognition system based on image processing, which relates to the field of road obstacle recognition systems and comprises a camera group of an image acquisition module, an image preprocessing module, a built-in processor, an algorithm denoising module, a stretching comparison module, a data processing module and a data processing module, the method comprises the following steps: combining a vehicle data correction image, a feature extraction module, a leading-edge CNN variant model, sample training, feature extraction by means of a technology, accurate target feature capture, an obstacle identification and classification module, and a vehicle and target dynamic, trajectory prediction, decision information providing, warning and feedback module based on the extracted features by using an integrated learning classifier. A flexible screen is arranged to display information in real time, sound and light alarms are graded according to risks, danger instructions are quickly transmitted to an ECU, and the vehicle is driven to operate emergently. The system has remarkable advantages, image acquisition and preprocessing adapt to the environment, feature extraction, recognition and classification are accurate and efficient, and driving safety is guaranteed.

Description

Technical Field

[0001] The present invention relates to the field of road obstacle recognition systems, and in particular to a road obstacle recognition system based on image processing. Background Art

[0002] In today's increasingly complex road traffic environment, road obstacle recognition technology plays a vital role in driving safety. With the continuous increase in the number of cars, the traffic conditions on the road are becoming more and more changeable, and various potential obstacles may threaten the safety of life and property of drivers and passengers at any time.

[0003] Traditional methods of identifying road obstacles have many limitations. Early simple visual recognition systems based on a single camera had low camera resolution and narrow viewing angles, making it difficult to capture all-round road scene information. In complex road conditions, such as curves, intersections, or overtaking scenes, it is very easy to miss obstacles in blind spots. Moreover, such systems lack effective image preprocessing capabilities. Faced with severe weather conditions such as direct sunlight, fog, and dust, the collected images are blurry and have extremely low contrast, which greatly reduces the subsequent recognition accuracy and cannot provide reliable warnings to drivers.

[0004] Furthermore, the previous feature extraction and recognition algorithms were relatively crude. Most of them only roughly classified objects based on their basic outlines. It was difficult to accurately distinguish obstacles with similar shapes but different properties, such as fallen bicycles and roadblock signs, and the misjudgment rate was high. At the same time, the system was weak in dealing with obstacles of different sizes, from nails and potholes on the road to large construction vehicles occupying the road, and it was unable to identify them at the same time, posing a hidden danger to driving safety.

[0005] In addition, the warning feedback link is not perfect. Some systems can only emit monotonous alarm sounds, and it is difficult for drivers to intuitively judge the degree of danger from the alarm; the display device screen is small and the information is updated slowly, and it is impossible to present detailed dynamic information of obstacles in time, such as movement trajectory, speed, etc., making it difficult for drivers to make accurate avoidance decisions in time.

[0006] In general, the existing road obstacle recognition technology can no longer meet the growing demand for traffic safety. There is an urgent need for a new, high-precision, high-reliability, and complex environment-adaptive road obstacle recognition system based on image processing to fill this technical gap and improve road driving safety in all aspects. Summary of the invention

[0007] The road obstacle recognition system based on image processing proposed by the present invention is to solve the problems mentioned in the above-mentioned prior art.

[0008] In order to achieve the above-mentioned purpose, the present invention adopts the following technical solution: a road obstacle recognition system based on image processing, comprising:

[0009] Image acquisition module: Using a camera group, based on the principle of binocular stereo vision, the object distance information is calculated through the parallax between the two cameras to build a three-dimensional scene model. The camera has an intelligent light adaptation system that automatically switches the filter, adjusts the aperture and shutter under a light intensity of 20-15000 lux, and satisfies the exposure formula E=I×S×A based on the light intensity I, shutter speed S, and aperture value A, where E is the exposure amount;

[0010] The camera group is installed on the front grille, side mirrors and rear spoiler of the vehicle, and is fixed by an active shock-absorbing device. The device has a built-in acceleration sensor to monitor the vibration of the vehicle in real time. According to the vibration frequency f and amplitude a, electromagnetic force or hydraulic damping is used to offset the vibration, where the damping force F = -c × v, c is the damping coefficient, and v is the relative speed;

[0011] Image preprocessing module: Built-in computing image processor, equipped with 1GB or more cache, uses artificial intelligence-based noise elimination algorithm to distinguish the real texture and noise of the image, removes noise while retaining edge details, improves image smoothness, and calculates the grayscale variance σ 2 Measure, after processing σ 2 Reduce; Adaptive histogram stretching technology is used to dynamically adjust the stretching amplitude according to the brightness distribution of different areas of the image;

[0012] Combine the vehicle GPS navigation information with the inertial measurement unit IMU technology to obtain the vehicle position, pitch angle α, yaw angle β, and roll angle γ. Use the affine transformation matrix M to correct the image. The original image coordinates (x o ,y o ) is transformed to (x c ,y c ), satisfying (x c ,y c )=M×(x o ,y o );

[0013] Feature extraction module: Use the convolutional neural network (CNN) variant model, including the residual neural network (ResNet) combined with the attention mechanism, and train with samples of road obstacles in different environments and types;

[0014] Relying on the feature pyramid to integrate FPN and dilated convolution technology, when constructing feature maps, different layers focus on information of different scales according to the preset scale range;

[0015] Obstacle recognition and classification module: Based on feature extraction, it uses an integrated learning classifier and combines the advantages of support vector machine (SVM) and random forest (RF) algorithms to achieve an obstacle classification accuracy of 92%-98%.

[0016] Combined with the vehicle's real-time speed v c , acceleration a c And the target dynamic information, use the Kalman filter prediction model to predict the obstacle trajectory, the prediction time is 4-6 seconds, assess the collision risk in advance, and give timely warnings. According to the kinematic formula x t (T) = x t (0)+v t ×T+0.5×a t ×T 2 y t (T) = y t (0)+v t ×T+0.5×a t ×T 2 , where ((x t (0), y t (0)) is the initial coordinate of the target, T is the prediction time, and the future position is calculated;

[0017] Warning and feedback module: equipped with a display screen, fitted to the curved surface inside the vehicle, installed directly in front of the driver's field of vision, and displays the obstacle's location, type, distance, movement trajectory, and risk level in real time;

[0018] When the system determines the risk of collision, it triggers sound and light alarms according to the risk level. Low risk sound pressure level is 75-85dB, and light flashing frequency is 1.5-2.5Hz; medium risk is 90-100dB, 3-4Hz; high risk is 110-120dB, 4.5-5.5Hz. At the same time, the danger command is sent to the vehicle's electronic control unit ECU through the on-board high-speed bus, driving the vehicle to start braking within 0.3 seconds.

[0019] Furthermore, in the image acquisition module, the camera group shell is made of self-cleaning nanomaterials. The surface microstructure of the nanomaterial has a special hydrophobic, oleophobic and antistatic design, which removes dust from the lens surface through high-frequency vibration.

[0020] Furthermore, in the image preprocessing module, a physical model-based image defogging algorithm is introduced for foggy and hazy scenes. The fog concentration C is estimated according to the atmospheric scattering model. By inverting the light propagation model in the fog and using dark channel prior knowledge and technical means, a clear image is restored. The visibility of the processed image is improved by 2-3 times.

[0021] Furthermore, in the feature extraction module, CNN model training adopts a distributed training architecture, using multiple GPU servers to collaborate. The servers are connected through the InfiniBand network to accelerate the training process and shorten the training time by 50%-70%. During the training process, adaptive gradient clipping technology is used to prevent gradient explosion and dynamically adjust the clipping threshold according to the norm of the gradient.

[0022] Furthermore, in the obstacle recognition and classification module, in order to update the recognition knowledge in real time, the system has the ability of online learning. When the vehicle encounters a new type of obstacle during driving, it automatically collects samples, uses idle computing power to learn in the background, updates the model parameters, and adopts a small sample incremental learning strategy.

[0023] Furthermore, in the warning and feedback module, the display screen supports voice interaction function, adopts microphone array technology, has 4-8 microphones, realizes 360° all-round sound pickup, has a voice recognition accuracy rate of more than 95%, and a response time of less than 0.5 seconds. Through deep learning voice model, the driver's voice commands are analyzed in real time.

[0024] Furthermore, a method for applying the road obstacle recognition system based on image processing comprises the following steps:

[0025] Image acquisition steps: When the vehicle is powered on, the camera group of the image acquisition module is immediately started, continuously capturing road images at a set frame rate of 60 frames per second, and simultaneously sensing the ambient light intensity I env , Temperature T env Parameters, automatically adapt shooting parameters, captured images are transmitted to the image pre-processing module in real time via high-speed optical fiber;

[0026] Image preprocessing steps: When the image preprocessing module receives the image, the processor runs at full speed and uses intelligent noise elimination and adaptive histogram stretching technology to process the image, which takes no more than 30ms; the image is corrected in combination with the vehicle position and posture data, which takes less than 20ms, and the processed image is quickly pushed to the feature extraction module;

[0027] Feature extraction step: The feature extraction module loads the pre-trained and optimized CNN variant model, scans the input image to extract features, and the processing time of a single frame image does not exceed 80ms. It extracts shape, texture, and motion features, constructs a multi-scale feature map, and provides an accurate basis for recognition and classification;

[0028] Obstacle recognition and classification step: Based on the feature extraction step, the obstacle recognition and classification module uses an integrated learning classifier for recognition and classification. Each frame takes about 70ms to recognize and classify. After determining the obstacle type, it combines the vehicle dynamics and the target motion prediction trajectory to estimate the collision risk 4-6 seconds in advance and output the result.

[0029] Warning and feedback steps: The warning and feedback module monitors the recognition results in real time and displays the obstacle information on the display screen at a refresh rate of 3Hz, so that the driver can control it at any time.

[0030] Furthermore, in the image acquisition step, if the camera group detects that the light intensity exceeds 8000 lux directly and lasts for more than 2 seconds, the multi-layer coated neutral density filter is automatically activated with a light reduction rate of 40%-60%. The multi-layer coating is made of materials with different refractive indices stacked alternately, and effectively reduces the light intensity through multiple reflections and refraction principles.

[0031] Furthermore, in the feature extraction step, when it is found that the similarity between the target feature and the known obstacle feature is less than 40%, the target image and related feature information are automatically uploaded to the cloud server via the 5G network. The cloud uses an ultra-large-scale database and a super-powerful computing cluster for secondary identification, and the cloud server adopts a parallel computing architecture.

[0032] Furthermore, in the warning and feedback steps, a dynamic color-changing warning light design is adopted for collision risks of different urgency levels, and the light brightness and flashing interval are controlled through pulse width modulation technology; a double flash mode is used to enhance the warning effect.

[0033] Compared with the prior art, the present invention has the following beneficial effects:

[0034] First, the image acquisition module has outstanding advantages. The high-resolution, ultra-wide-angle and depth-sensing camera group, combined with the intelligent light adaptation system and active vibration reduction device, can stably obtain clear and detailed images in strong light, weak light or bumpy road conditions, build accurate three-dimensional scene models, greatly broaden the field of view and reduce visual blind spots.

[0035] Secondly, the image preprocessing module uses advanced algorithms and high-precision vehicle posture data to effectively remove noise, enhance contrast and accurately correct images, laying a solid foundation for subsequent recognition, ensuring that the system can operate stably in different weather and road conditions, and significantly improving recognition accuracy.

[0036] Furthermore, the feature extraction and recognition module relies on cutting-edge CNN variant models and multi-scale feature extraction technology to achieve an accuracy rate of over 96% in feature extraction of common and rare obstacles, a classification accuracy rate of 92%-98%, and fine classification types, providing drivers with key decision-making information and effectively avoiding misjudgments.

[0037] Finally, the warning and feedback module's graded sound and light alarms are combined with a large-size, high-refresh rate display, a dynamic color-changing warning light design, and the display's voice interaction function, allowing drivers to intuitively judge risks, conveniently obtain information, and make avoidance maneuvers in a timely manner to ensure driving safety to the greatest extent. BRIEF DESCRIPTION OF THE DRAWINGS

[0038] Figure 1 A schematic block diagram of a road obstacle recognition system based on image processing proposed by the present invention;

[0039] Figure 2 This is a schematic block diagram of a road obstacle recognition method based on image processing proposed by the present invention. DETAILED DESCRIPTION

[0040] 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 only part of the embodiments of the present invention, not all of the embodiments. Based on 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 present invention.

[0041] In the description of the present invention, it should be understood that the terms "center", "longitudinal", "lateral", "length", "width", "thickness", "up", "down", "front", "back", "left", "right", "vertical", "horizontal", "top", "bottom", "inside", "outside", "clockwise", "counterclockwise" and the like indicate orientations or positional relationships based on the orientations or positional relationships shown in the accompanying drawings, and are only for the convenience of describing the present invention and simplifying the description, rather than indicating or implying that the referred device or element must have a specific orientation, be constructed and operated in a specific orientation, and therefore should not be understood as limiting the present invention.

[0042] In addition, the terms "first" and "second" are used for descriptive purposes only and cannot be understood as indicating or implying relative importance or implicitly indicating the number of technical features indicated. Therefore, the features defined as "first" and "second" may explicitly or implicitly include one or more of the features. In the description of the present invention, the meaning of "multiple" is two or more, unless otherwise clearly and specifically defined. In addition, the terms "installed", "connected" and "connected" should be understood in a broad sense. For example, it can be a fixed connection, a detachable connection, or an integral connection; it can be a mechanical connection or an electrical connection; it can be a direct connection, or it can be indirectly connected through an intermediate medium, and it can be the internal connection of two elements. For ordinary technicians in this field, the specific meanings of the above terms in the present invention can be understood according to specific circumstances. The present invention will be further described in detail below in conjunction with the accompanying drawings.

[0043] Reference Figure 1-2 :A road obstacle recognition system based on image processing, comprising:

[0044] Image acquisition module: It uses a high-resolution, ultra-wide-angle camera group with depth perception function. The camera pixel is not less than 20 million, the horizontal viewing angle is 220°, and the vertical viewing angle is 150°. Based on the principle of binocular stereo vision, the parallax between the two cameras can accurately calculate the distance information of the object and build a three-dimensional scene model. The camera has an intelligent light adaptation system that can automatically switch the filter, adjust the aperture and shutter under the light intensity of 20-15000lux. According to the light intensity I, shutter speed S, and aperture value A, the exposure formula E=I×S×A (E is the exposure) is satisfied to ensure that the image is clear, bright and rich in details without overexposure or underexposure.

[0045] It is installed on the front grille, side mirrors, rear spoiler and other locations of the vehicle and fixed by an active shock-absorbing device. The device has a built-in acceleration sensor to monitor the vibration of the vehicle in real time. According to the vibration frequency f and amplitude, it uses electromagnetic force or hydraulic damping (damping force F = -c×v, c is the damping coefficient, v is the relative speed) to offset the vibration, so that the camera displacement deviation is controlled within 0.1mm to ensure image stability.

[0046] Image preprocessing module: Built-in super computing image processor, core frequency over 2GHz, equipped with 1GB or more cache, using artificial intelligence-based noise removal algorithm, can distinguish the real texture and noise of the image, remove noise while retaining more than 95% of edge details, and improve image smoothness by 50%-70%. 2 Difference measurement, after processing σ 2 The adaptive histogram stretching technology is used to dynamically adjust the stretching range according to the brightness distribution of different areas of the image, which increases the image contrast by 3-6 times, and clearly presents the details of dark and bright areas.

[0047] Combine the vehicle's high-precision GPS navigation information with the inertial measurement unit (IMU) technology to obtain the vehicle's precise position and attitude (pitch angle α, yaw angle β, and roll angle γ are accurate to 0.05°). Use the affine transformation matrix M to correct the image, and the original image coordinates x o ,y o After transformation, it is x c ,y c , satisfying (x c ,y c )=M×(x o ,y o ), the image edge straightness error after correction is less than 0.2%, which accurately matches the actual scene.

[0048] Feature extraction module: Using cutting-edge convolutional neural network (CNN) variant models, such as residual neural network (ResNet) combined with attention mechanism, after hundreds of millions of road obstacle samples of different environments and types are trained, the accuracy of feature extraction for common and rare obstacles exceeds 96%. It can automatically capture multi-dimensional features of the target, such as identifying body movements and clothing texture features for pedestrians; for vehicles, it can extract unique body logos and component movement states (such as wheel rotation and wiper swing), etc., laying the foundation for accurate classification.

[0049] Relying on feature pyramid fusion (FPN) and dilated convolution technology, when constructing feature maps, different layers focus on information of different scales according to the preset scale range. For example, the bottom layer focuses on the feature extraction of tiny objects, and can accurately capture road screws that occupy 0.3% of the image area. The upper layers focus on large objects, ensuring efficient extraction of multi-scale target features from tiny road screws to large trucks, without missing any potential dangerous objects.

[0050] Obstacle recognition and classification module: Based on feature extraction, it uses an integrated learning classifier and integrates the advantages of algorithms such as support vector machine (SVM) and random forest (RF). The obstacle classification accuracy rate reaches 92%-98%. It not only distinguishes pedestrians, vehicles, fixed roadblocks and other major categories, but also subdivides them, such as pedestrians are classified by age and walking status (running, staggering), and vehicles are subdivided by purpose (fire trucks and ambulances have priority passage requirements). Different icons and colors are highlighted on the display interface to provide key decision-making information for drivers.

[0051] Combined with the vehicle's real-time speed vc, acceleration ac and target dynamic information (speed v t , acceleration a t 、Motion directionθ t ), using the Kalman filter prediction model to predict the trajectory of obstacles, with a prediction time of 4-6 seconds, to assess the risk of collision in advance and to issue a timely warning, based on the kinematic formula x t (T) = x t (0)+v t ×T+0.5×a t ×T 2 ,y t (T) = y t (0)+v t ×T+0.5×a t ×T 2 ,((x t (0), y t (0)) is the initial coordinate of the target, T is the predicted time) calculate the future position, determine the intersection with the vehicle trajectory,

[0052] Warning and feedback module:

[0053] Equipped with a large-size, high-brightness, flexible and bendable display screen, the size is not less than 10 inches, the resolution is 2560×1600 pixels, the viewing angle is 180°, and it is installed in the best position of the driver's field of vision to fit the curved surface inside the car. It can display the obstacle position, type, distance, movement trajectory, risk level, etc. in real time, with a refresh rate of 3Hz, and the information is updated in time.

[0054] When the system determines the risk of collision, it triggers the sound and light alarm according to the risk level. The low risk sound pressure level is 75-85dB and the light flashing frequency is 1.5-2.5Hz; the medium risk is 90-100dB and 3-4Hz; the high risk is 110-120dB and 4.5-5.5Hz. At the same time, the danger command is sent to the vehicle electronic control unit (ECU) through the on-board high-speed bus (rate 1Gbps and above), driving the vehicle to start emergency operations such as braking and avoidance within 0.3 seconds to avoid accidents.

[0055] In the present invention, in the image acquisition module, the camera group housing adopts self-cleaning nanomaterials. The surface microstructure of the nanomaterials is specially designed to be hydrophobic, oleophobic and antistatic. When encountering water or oil droplets, the contact angle is greater than 150°, so that it automatically slides off. When the dust on the surface accumulates to a certain extent (covering area exceeds 5%), ultrasonic vibration dust removal is started. The ultrasonic frequency is 20kHz-50kHz. The dust is separated from the lens surface through high-frequency vibration, keeping the lens clean, and the collection is not hindered in harsh environments.

[0056] In the present invention, in the image preprocessing module, for low visibility scenes such as foggy days and haze days, an image defogging algorithm based on a physical model is introduced, the fog concentration C is estimated according to the atmospheric scattering model, and a clear image is restored by inverting the light propagation model in the fog and using dark channel prior knowledge and other technical means. After processing, the visibility of the image is improved by 2-3 times, ensuring the accuracy of special weather recognition. Specifically, in the dark channel prior algorithm, it is assumed that the pixel points in a certain local area of ​​the image have a very low intensity value on at least one color channel, and the atmospheric light value is estimated on this basis, and then the transmittance is calculated according to the fog concentration and the light transmission model, and finally the image defogging is achieved.

[0057] In the feature extraction module of the present invention, the CNN model training adopts a distributed training architecture, using multiple GPU servers to collaborate, and the servers are connected through a high-speed InfiniBand network with a bandwidth of more than 100Gbps, which accelerates the training process and shortens the training time by 50%-70%. During the training process, the adaptive gradient clipping technology is used to prevent gradient explosion, and the clipping threshold is dynamically adjusted according to the norm of the gradient to ensure stable convergence of the model, quickly iterate the model, and adapt to complex and changeable road environments.

[0058] In the obstacle recognition and classification module of the present invention, in order to update the recognition knowledge in real time, the system has the ability of online learning. When the vehicle encounters a new type of obstacle during driving, it automatically collects samples, uses idle computing power to learn in the background, updates the model parameters, and adopts a small sample incremental learning strategy. Only a small number of new samples (generally 10-50) are needed each time to effectively update the model. After the update, the recognition accuracy rate of the next time is improved by 10%-20%. The mechanism based on memory playback is used to avoid new samples causing the model to forget old knowledge, ensuring that the recognition accuracy of common obstacles is not affected.

[0059] In the present invention, in the warning and feedback module, the display screen supports voice interaction function, adopts advanced microphone array technology, has 4-8 microphones, can achieve 360° omnidirectional sound pickup, voice recognition accuracy of more than 95%, response time within 0.5 seconds, through deep learning voice model, the driver's voice command is analyzed in real time, such as "query the details of the nearest obstacle" and "check the system status" and other commands, can quickly make accurate responses, the operation is convenient and efficient,

[0060] The present invention also discloses a road obstacle recognition method based on image processing, which comprises the following steps:

[0061] Image acquisition steps:

[0062] When the vehicle is powered on, the camera group of the image acquisition module starts immediately, continuously capturing road images at a set frame rate of 60 frames per second, and simultaneously sensing the ambient light intensity I env , Temperature T env The captured images are transmitted to the image preprocessing module in real time via high-speed optical fiber (transmission rate 2Gbps and above), with a transmission delay of less than 5ms, ensuring the timeliness and integrity of the images.

[0063] Image preprocessing steps: When the image preprocessing module receives the image, the processor runs at full speed and uses intelligent noise elimination and adaptive histogram stretching technology to process it, which takes no more than 30ms; the image is corrected in combination with the vehicle position and posture data, which takes less than 20ms. The processed image is quickly pushed to the feature extraction module to ensure the efficiency of subsequent processes.

[0064] Feature extraction step: The feature extraction module loads the pre-trained and optimized CNN variant model, scans the input image to extract features, and the processing time of a single frame image does not exceed 80ms. It extracts features such as shape, texture, and motion, and constructs a multi-scale feature map to provide an accurate basis for recognition and classification.

[0065] Obstacle recognition and classification steps: Based on the features, the obstacle recognition and classification module uses an integrated learning classifier for recognition and classification. Each frame takes about 70ms to classify. After determining the obstacle type, the module combines the vehicle dynamics and the target motion prediction trajectory to estimate the collision risk 4-6 seconds in advance and output the results in a timely manner.

[0066] Warning and feedback steps: The warning and feedback module monitors the recognition results in real time and displays the obstacle information on the display screen at a refresh rate of 3Hz, so that the driver can control it at any time; once there is a risk of collision, the sound and light alarm is triggered according to the risk level, and the danger command is transmitted to the ECU within 0.3 seconds, and the vehicle responds urgently to ensure driving safety.

[0067] In the present invention, in the image acquisition step, if the camera group detects strong direct light (light intensity exceeds 8000 lux) and lasts for more than 2 seconds, the multi-layer coating neutral density filter is automatically activated, with a light reduction rate of 40%-60%. The multi-layer coating is made of materials with different refractive indices stacked alternately, and effectively reduces the light intensity through multiple reflections and refraction principles, avoids overexposure, and stabilizes image acquisition.

[0068] In the present invention, in the feature extraction step, when it is found that the similarity between the target feature and the known obstacle feature is less than 40%, the target image and related feature information are automatically uploaded to the cloud server via the 5G network. The cloud uses an ultra-large-scale database and a super-powerful computing cluster for secondary identification. The cloud server adopts a large-scale parallel computing architecture with thousands of computing nodes and a floating-point computing capacity of more than 10 billion times per second. The results are returned within 3 seconds, expanding the ability to identify unknown obstacles.

[0069] In the present invention, in the warning and feedback steps, for collision risks of different urgency levels, a dynamic color-changing warning light design is adopted. A low-risk light is displayed in blue with a flashing frequency of 1.5-2.5Hz, and the light brightness and flashing interval are accurately controlled by pulse width modulation technology; a medium-risk light is displayed in yellow at 3-4Hz, and a double flash mode is adopted to enhance the warning effect; a high-risk red light is displayed in red at 4.5-5.5Hz, with the brightness increased to the maximum and flashing rapidly, to assist the driver in intuitively judging the risk and enhance the warning effect.

[0070] The above are only preferred specific implementation modes of the present invention, but the protection scope of the present invention is not limited thereto. Any technician familiar with the technical field can make equivalent replacements or changes according to the technical solutions and inventive concepts of the present invention within the technical scope disclosed by the present invention, which should be covered by the protection scope of the present invention.

Claims

1. A road obstacle recognition system based on image processing, characterized in that: include: Image acquisition module: Using a camera group, based on the principle of binocular stereo vision, the object distance information is calculated through the parallax between the two cameras to build a three-dimensional scene model. The camera has an intelligent light adaptation system that automatically switches the filter, adjusts the aperture and shutter under a light intensity of 20-15000 lux, and satisfies the exposure formula E=I×S×A based on the light intensity I, shutter speed S, and aperture value A, where E is the exposure amount; The camera group is installed on the front grille, side mirrors and rear spoiler of the vehicle, and is fixed by an active shock-absorbing device. The device has a built-in acceleration sensor to monitor the vibration of the vehicle in real time. According to the vibration frequency f and amplitude a, electromagnetic force or hydraulic damping is used to offset the vibration, where the damping force F = -c × v, c is the damping coefficient, and v is the relative speed; Image preprocessing module: Built-in computing image processor, equipped with 1GB or more cache, uses artificial intelligence-based noise elimination algorithm to distinguish the real texture and noise of the image, removes noise while retaining edge details, improves image smoothness, and calculates the grayscale variance σ 2 Measure, after processing σ 2 Reduce; Adaptive histogram stretching technology is used to dynamically adjust the stretching amplitude according to the brightness distribution of different areas of the image; Combine the vehicle GPS navigation information with the inertial measurement unit IMU technology to obtain the vehicle position, pitch angle α, yaw angle β, and roll angle γ. Use the affine transformation matrix M to correct the image. The original image coordinates (x o ,y o ) is transformed to (x c ,y c ), satisfying (x c ,y c )=M×(x o ,y o ); Feature extraction module: Use the convolutional neural network (CNN) variant model, including the residual neural network (ResNet) combined with the attention mechanism, and train with samples of road obstacles in different environments and types; Relying on the feature pyramid to integrate FPN and dilated convolution technology, when constructing the feature map, different layers focus on information of different scales according to the preset scale range; Obstacle recognition and classification module: Based on feature extraction, it uses an integrated learning classifier and combines the advantages of support vector machine (SVM) and random forest (RF) algorithms to achieve an obstacle classification accuracy of 92%-98%. Combined with the vehicle's real-time speed v c , acceleration a c And the target dynamic information, use the Kalman filter prediction model to predict the obstacle trajectory, the prediction time is 4-6 seconds, assess the collision risk in advance, and give timely warnings. According to the kinematic formula x t (T) = x t (0)+υ t ×T+0.5×a t ×T 2 y t (T) = y t (0)+υ t ×T+0.5×a t ×T 2 , where ((x t (0), y t (0)) is the initial coordinate of the target, T is the prediction time, and the future position is calculated; Warning and feedback module: equipped with a display screen, fitted to the curved surface inside the vehicle, installed directly in front of the driver's field of vision, and displays the obstacle's location, type, distance, movement trajectory, and risk level in real time; When the system determines the risk of collision, it triggers sound and light alarms according to the risk level. Low risk sound pressure level is 75-85dB, and light flashing frequency is 1.5-2.5Hz; medium risk is 90-100dB, 3-4Hz; high risk is 110-120dB, 4.5-5.5Hz. At the same time, the danger command is sent to the vehicle's electronic control unit ECU through the on-board high-speed bus, driving the vehicle to start braking within 0.3 seconds.

2. The road obstacle recognition system based on image processing according to claim 1, characterized in that: In the image acquisition module, the camera group shell is made of self-cleaning nanomaterials. The surface microstructure of the nanomaterial has a special hydrophobic, oleophobic and anti-static design, which removes dust from the lens surface through high-frequency vibration.

3. The road obstacle recognition system based on image processing according to claim 1, characterized in that: In the image preprocessing module, a physical model-based image defogging algorithm is introduced for foggy and hazy scenes. The fog concentration C is estimated according to the atmospheric scattering model. The light propagation model in the fog is inverted and the dark channel prior knowledge and technical means are used to restore the clear image. The visibility of the processed image is improved by 2-3 times.

4. The road obstacle recognition system based on image processing according to claim 1, characterized in that: In the feature extraction module, CNN model training adopts a distributed training architecture, using multiple GPU servers to collaborate. The servers are connected through the InfiniBand network to accelerate the training process and shorten the training time by 50%-70%. During the training process, adaptive gradient clipping technology is used to prevent gradient explosion and dynamically adjust the clipping threshold according to the norm of the gradient.

5. The road obstacle recognition system based on image processing according to claim 1, characterized in that: In the obstacle recognition and classification module, in order to update the recognition knowledge in real time, the system has the ability of online learning. When the vehicle encounters a new type of obstacle during driving, it automatically collects samples, uses idle computing power to learn in the background, updates the model parameters, and adopts a small sample incremental learning strategy.

6. The road obstacle recognition system based on image processing according to claim 1, characterized in that: In the warning and feedback module, the display screen supports voice interaction function, adopts microphone array technology, has 4-8 microphones, realizes 360° all-round sound pickup, has a voice recognition accuracy rate of more than 95%, and a response time of less than 0.5 seconds. Through deep learning voice model, it can analyze the driver's voice commands in real time.

7. A method for applying the road obstacle recognition system based on image processing according to any one of claims 1 to 6, characterized in that: The following steps are involved: Image acquisition steps: When the vehicle is powered on, the camera group of the image acquisition module is immediately started, continuously capturing road images at a set frame rate of 60 frames per second, and simultaneously sensing the ambient light intensity I enυ , Temperature T evυ Parameters, automatically adapt shooting parameters, captured images are transmitted to the image pre-processing module in real time via high-speed optical fiber; Image preprocessing steps: When the image preprocessing module receives the image, the processor runs at full speed and uses intelligent noise elimination and adaptive histogram stretching technology to process the image, which takes no more than 30ms; the image is corrected in combination with the vehicle position and posture data, which takes less than 20ms, and the processed image is quickly pushed to the feature extraction module; Feature extraction step: The feature extraction module loads the pre-trained and optimized CNN variant model, scans the input image to extract features, and the processing time of a single frame image does not exceed 80ms. It extracts shape, texture, and motion features, constructs a multi-scale feature map, and provides an accurate basis for recognition and classification; Obstacle recognition and classification step: Based on the feature extraction step, the obstacle recognition and classification module uses an integrated learning classifier for recognition and classification. Each frame takes about 70ms to recognize and classify. After determining the obstacle type, it combines the vehicle dynamics and the target motion prediction trajectory to estimate the collision risk 4-6 seconds in advance and output the result. Warning and feedback steps: The warning and feedback module monitors the recognition results in real time and displays the obstacle information on the display screen at a refresh rate of 3Hz, so that the driver can control it at any time.

8. According to the method for identifying road obstacles based on image processing according to claim 7, in the image acquisition step, if the camera group detects that the light intensity exceeds 8000 lux and lasts for more than 2 seconds, the multi-layer coated neutral density filter is automatically activated, and the neutral density rate is 40%-60%. The multi-layer coating is formed by alternating stacking of materials with different refractive indices, and effectively reduces the light intensity through multiple reflections and refraction principles.

9. According to the road obstacle identification method based on image processing as described in claim 7, in the feature extraction step, when it is found that the similarity between the target feature and the known obstacle feature is less than 40%, the target image and related feature information are automatically uploaded to the cloud server via the 5G network, and the cloud uses an ultra-large-scale database and a super-powerful computing cluster for secondary identification, and the cloud server adopts a parallel computing architecture.

10. According to the method for identifying road obstacles based on image processing as described in claim 7, in the warning and feedback step, for collision risks of different urgency, a dynamic color-changing warning light design is adopted, and the light brightness and flashing interval are controlled by pulse width modulation technology; a double flash mode is adopted to enhance the warning effect.

Images