Safe obstacle avoidance method applied to automatic driving and based on obstacle material detection

Abstract

The invention provides a safe obstacle avoidance method applied to automatic driving and based on obstacle material detection. The safe obstacle avoidance method comprises the steps that millimeter wave echo signals of different obstacles are collected through a vehicle-mounted millimeter wave radar of an automatic driving vehicle; performing wavelet transform on the millimeter wave echo signal to obtain a signal after noise removal; performing fast Fourier transform on the denoised signal to obtain frequency domain information; performing feature extraction on the frequency domain information to obtain 10 feature values; inputting the 10 feature values into a convolutional neural network to obtain a classification result of the signals; inputting a classification result into an Apoll automatic driving system, and optimizing an obstacle avoidance scheme output by the system through a planning module; extracting a dangerous scene from the INTERACTION real scene data set by using a dangerous scene extraction method based on a geometric lane relationship; the extracted scene is simulated and reproduced on the Apollo platform so as to verify the performance of the optimized obstacle avoidance method; according to the method, the Apoll automatic driving decision is changed, and when the soft obstacle is detected, the system directly passes through the obstacle, so that the collision risk estimation level of the vehicle is reduced, and the collision accident of the vehicle is avoided.

Description

Technical Field

[0001] This invention belongs to the field of intelligent driving system safety, specifically relating to a safety obstacle avoidance method based on obstacle material detection for autonomous driving. Background Technology

[0002] Recently, with the rapid development of artificial intelligence technology, intelligent driving systems, with their precise environmental perception capabilities and intelligent decision-making, are leading the future of transportation and driving a new round of industrial revolution. As application scenarios continue to expand, intelligent driving systems need to face more complex and variable traffic environments on actual roads. During driving, the system typically relies on perception devices such as millimeter-wave radar, lidar, and cameras, combined with multi-sensor fusion and intelligent decision-making methods, to identify and model the surrounding environment and generate corresponding driving strategies.

[0003] When the perception module of an intelligent driving system detects a potential obstacle, avoidance methods are invoked, triggering actions such as braking or lane changing based on environmental conditions to achieve obstacle avoidance control. However, typical obstacle avoidance decisions rely solely on the geometric and motion characteristics of obstacles, such as position, distance, and speed. When an obstacle is detected in the current driving direction, braking or lane changing is often made, which can easily lead to rear-end collisions. In reality, obstacles of different materials have significantly different impacts on vehicle safety. For example, soft materials such as paper bags and plastic bags cause almost no substantial damage to vehicles; driving directly over such obstacles is actually safer than sudden braking or forced lane changes, helping to reduce the probability of accidents. Because millimeter waves have different reflection characteristics on different materials, the material information of different obstacles can be identified using the characteristics of millimeter wave echo signals. The obstacle material identification results can then be used as an important basis for obstacle avoidance decisions, improving the obstacle avoidance safety performance of autonomous driving systems. Summary of the Invention

[0004] In view of the above, this invention proposes a safety-enhancing obstacle avoidance method. It utilizes millimeter-wave radar to perceive obstacle material information and incorporates the perception results into the obstacle avoidance decision-making system of an autonomous driving system. This allows the system to make smarter and safer decisions based on obstacle material information, thereby effectively reducing the occurrence of collisions. To achieve the above objective, this invention provides a safety obstacle avoidance method based on obstacle material detection for autonomous driving. This method includes: using millimeter-wave radar modules to collect millimeter-wave echo signals from different obstacles; performing wavelet transform on the millimeter-wave echo signals to obtain noise-removed signals; performing fast Fourier transform on the denoised signals to obtain frequency domain information; extracting features from the frequency domain information to obtain 10 feature values; inputting the 10 feature values ​​into a convolutional neural network to obtain signal classification results; inputting the classification results into the Apollo autonomous driving system for obstacle material identification, and optimizing the obstacle avoidance scheme through a planning module based on the material identification results; extracting hazardous scene features from the INTERACTION real-world scene dataset using a hazardous scene extraction method based on geometric lane relationships; and then simulating and reproducing the extracted scenes on the Apollo platform to verify the performance of the optimized obstacle avoidance method.

[0005] The beneficial effects of this invention are:

[0006] This invention uses a CNN model to classify obstacle materials into soft and hard categories based on the frequency statistical characteristics of millimeter-wave reflection. The classification results are then integrated into the Apollo perception module, and the obstacle avoidance decision logic in the Apollo planning module is modified. A hazard scene extraction method based on geometric lane relationships is presented to extract hazard scenes from the INTERACTION real-world scenario dataset, and simulations are performed on the Apollo platform to verify the performance of the optimized obstacle avoidance decision method. Using the obstacle avoidance decision method presented in this invention, the Apollo autonomous driving decision-making is altered. When a soft obstacle is detected, the system directly traverses the obstacle, reducing the vehicle collision risk prediction level and preventing collision accidents. Attached Figure Description

[0007] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0008] Figure 1 This is an overall flowchart of the present invention;

[0009] Figure 2 This is the millimeter-wave echo signal acquisition scenario of the present invention;

[0010] Figure 3 The image shows a portion of the obstacle to be collected in this invention.

[0011] Figure 4 This is a hazardous scene extraction image from the INTERACTION dataset based on geometric lane relationships, as presented in this invention.

[0012] Figure 5 This is a comparison chart showing the decision-making effects of the obstacle avoidance method before and after optimization in two typical scenarios of the present invention.

[0013] Figure 6 This is a CRELM effect diagram using the original OAA for the present invention;

[0014] Figure 7 This is a CRELM effect diagram of ADOAA according to the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] A safety obstacle avoidance method based on obstacle material detection for autonomous driving is disclosed. The method includes: using millimeter-wave echo signals from different obstacles via a millimeter-wave radar module; performing wavelet transform on the millimeter-wave echo signals to obtain noise-removed signals; performing fast Fourier transform on the denoised signals to obtain frequency domain information; extracting features from the frequency domain information to obtain 10 feature values; inputting the 10 feature values ​​into a convolutional neural network to obtain signal classification results; inputting the classification results into the Apollo autonomous driving system for obstacle material identification, and optimizing the obstacle avoidance scheme through a planning module based on the material identification results; extracting hazardous scene features from the INTERACTION real-world scene dataset using a hazardous scene extraction method based on geometric lane relationships; and then simulating and reproducing the extracted scenes on the Apollo platform to verify the performance of the optimized obstacle avoidance method.

[0017] In this embodiment, a specific implementation of a safety obstacle avoidance method based on obstacle material detection for autonomous driving is provided, including:

[0018] S1. Collect millimeter-wave echo signals from different obstacles using the millimeter-wave radar module of the autonomous vehicle. Establish a database of millimeter-wave echo signals from obstacles.

[0019] S2. Denoising the acquired signal using wavelet transform, with the wavelet type being db8, and performing a 5-level decomposition on the signal.

[0020] S3. Use FFT to convert the audio signal to the frequency domain for analysis and processing.

[0021] S4. Statistical feature extraction is performed on the frequency domain information to obtain 10 typical statistical feature values ​​of the spectrum. The specific parameters include: center frequency, spectral bandwidth, spectral skewness, spectral kurtosis, dominant frequency, average frequency, median frequency, spectral entropy, mean square frequency, and frequency standard deviation.

[0022] S5. Design a Convolutional Neural Network (CNN) model to classify obstacle types. The CNN model mainly consists of two convolutional units, each containing a batch normalization layer, a ReLU activation function, and a convolutional layer. The first convolutional layer contains 16 filters, and the second convolutional layer contains 32 filters. The fully connected layer contains 512 hidden units and uses the sigmoid activation function for binary classification.

[0023] S6. The classification results are integrated into the Apollo autonomous driving system, and its perception and planning modules are modified to optimize its obstacle avoidance decision-making method. By writing code, the input parameter dimensions of the perception module's "RadarSignalProcessing" function are expanded to include not only speed and distance information but also obstacle material information. The material type of obstacles is marked in the "ObjectDetection" function, and obstacle information containing material characteristics is passed from the perception module to the prediction and planning module. In the planning module's "PathPlanning" function, a conditional judgment function is added. When the planning module receives the material type of a soft obstacle from the perception module, the "PathPlanning" function will no longer consider obstacle information for path planning, and the "DecisionMaking" function will not use obstacle information for decision-making. For hard obstacles, the functionality of the "PathPlanning" and "DecisionMaking" functions remains unchanged. Through these modifications to Apollo's perception and planning modules, when a soft obstacle is detected, the autonomous vehicle can directly pass through it.

[0024] S7. Using a hazard scene extraction method based on geometric lane relationships, two types of hazard scenes, sudden changes and emergency braking, are extracted from the INTERACTION real-world scene dataset to verify the optimized obstacle avoidance method.

[0025] S8. The optimized obstacle avoidance method was verified using the Apollo simulation platform, and the collision risk estimation level was calculated in two typical scenarios to illustrate the improvement of the system safety performance by the optimized method.

[0026] In this embodiment, when the perception system of an autonomous vehicle detects an obstacle ahead, it typically initiates braking or steering to avoid it using obstacle avoidance methods. However, if the avoidance strategy is based solely on the geometric and motion parameters of the obstacle, such as its position, distance, and speed, it can easily lead to overly aggressive or sudden avoidance behaviors, such as sudden braking or forced lane changes, which may actually increase the risk of collision. In fact, obstacles of different materials have drastically different effects on driving safety. For example, soft obstacles such as paper bags and plastic bags cause almost no substantial damage to the vehicle. Compared to emergency braking or forced steering, driving directly over such objects may be a safer choice, helping to reduce the overall likelihood of accidents. Therefore, this invention provides an improved obstacle avoidance method (ADOAA), which uses millimeter-wave radar to detect the material information of obstacles and uses it as an important parameter for driving decisions, improving the safety performance of autonomous vehicles. The overall workflow of ADOAA is as follows: Figure 1 As shown.

[0027] First, such as Figure 2 and Figure 3 As shown, millimeter-wave echo signals from different obstacles are collected using the millimeter-wave radar module of an autonomous vehicle, and a database of obstacle millimeter-wave echo signals is established. During the acquisition process, because the millimeter-wave echo signals are affected by environmental noise and interference during propagation, wavelet transform is first used to analyze the collected millimeter-wave echo signals from the obstacles. Denoising to improve the signal-to-noise ratio of the echo signal can be expressed as:

[0028]

[0029] in, This is the millimeter-wave echo signal after wavelet transform. To control the scaling parameters of wavelet frequencies, The position parameter is used to control the time displacement of the wavelet. Denotes the wavelet complex conjugate function. It is a millimeter-wave echo signal.

[0030] Based on the denoised millimeter-wave reflection signal, FFT is used to convert the signal to the frequency domain for analysis, thereby obtaining the spectral information of the millimeter-wave echo signal.

[0031]

[0032] in, The sampling points are used. However, the frequency differences between different obstacle materials cannot be directly observed from the spectrum. Therefore, it is necessary to calculate 10 typical spectral statistical features, including center frequency, spectral bandwidth, spectral skewness, spectral kurtosis, dominant frequency, average frequency, median frequency, spectral entropy, mean square frequency, and frequency standard deviation, and use them as classification features.

[0033] Assume the sampling frequency is The actual frequency corresponding to the k-th frequency point It can be represented as:

[0034]

[0035] Corresponding power spectrum It can be represented as:

[0036]

[0037] The center frequency of the echo signal is expressed as:

[0038]

[0039] Spectrum bandwidth Represented as:

[0040]

[0041] The spectral skewness S is expressed as:

[0042]

[0043] The spectral kurtosis F is expressed as:

[0044]

[0045] clock speed Represented as:

[0046]

[0047] average frequency Represented as:

[0048]

[0049] The median frequency is the frequency at which half of the power spectrum accumulates, i.e., when... hour, Take the corresponding This is the median frequency.

[0050] Spectral entropy Represented as:

[0051]

[0052] Mean square frequency Represented as:

[0053]

[0054] Frequency standard deviation Represented as:

[0055]

[0056] The 10 spectral feature parameters obtained by the above formula are used as feature vector inputs.

[0057] In this embodiment, common road obstacles are divided into two categories: hard obstacles, mainly metallic objects such as metal boxes, computer mainframes, and mechanical parts; and soft obstacles, mainly non-metallic objects such as paper bags, plastic bags, and clothing. A CNN model is used for classification. The CNN model design mainly includes two convolutional units, each containing a batch normalization layer, a ReLU activation function, and a convolutional layer. The application of the batch normalization layer reduces dependence on initial values ​​and improves the model's generalization ability. The ReLU activation function is adopted due to its low computational complexity and ability to mitigate the vanishing gradient problem in deep neural networks. The main function of the convolutional layer is to sense local features by applying filters to the input data. To enhance the model's ability to learn high-level feature representations, 16 filters are used in the first convolutional layer, increasing to 32 filters in the second convolutional layer. After the two convolutional units, a flattening layer is introduced to convert the multidimensional feature map into a one-dimensional vector. This flattened vector is then passed to a fully connected layer. The fully connected layer contains 512 hidden units and uses the Sigmoid activation function for binary classification.

[0058] After classifying soft and hard obstacles using a CNN model, the results are fed into the Apollo autonomous driving method. Apollo is a well-known open-source, scalable autonomous driving platform that provides a complete autonomous driving software stack. The Apollo framework consists of four main modules: perception, prediction, planning, and control. The perception module is responsible for collecting and processing raw data of the vehicle's driving environment; the prediction module uses the output of the perception module to predict the future behavior of obstacles; the planning module is the core of Apollo, generating driving paths based on the inputs from the perception and prediction modules; and the control module manipulates the vehicle's accelerator, brakes, and steering according to the path information provided by the planning module to follow the planned path.

[0059] To implement flexible obstacle avoidance decisions based on material properties within the Apollo autonomous driving framework, the results of a CNN model are embedded into the perception module. In the perception module, Apollo primarily uses the "RadarSignalProcessing" function to process raw data from millimeter-wave radar and the "ObjectDetection" function to detect obstacles. Therefore, this study extends the "RadarSignalProcessing" function to output not only speed and distance information but also material information.

[0060] Using a CNN model to analyze obstacles For material identification, the classification result M is represented as:

[0061]

[0062] For obstacle material categories, These are the parameters for the CNN model.

[0063] In the "ObjectDetection" function, the material type of the obstacle is marked. The obstacle information containing material features is then passed from the perception module to the prediction and planning module. The expanded perception output is as follows:

[0064]

[0065] in, Indicates the first The comprehensive characteristics of each obstacle, including the original obstacle information and location from Apollo. ,speed ,distance And extended material types .

[0066] The obstacle avoidance decision-making in the planning module mainly involves the "PathPlanning" and "DecisionMaking" functions. The "PathPlanning" function generates the vehicle's future driving path based on obstacle, vehicle, and environmental information; the prediction module maintains the original Apollo logic, predicting the future behavior trajectory of obstacles based on perception results.

[0067]

[0068] in, This represents the set of obstacle positions and motion states at future moments.

[0069] The "DecisionMaking" function outputs specific vehicle control commands based on the information provided by the "PathPlanning" function. To change the decision-making process from avoiding soft obstacles to directly passing them, this study modified the logical structure of these functions by adding a conditional judgment function. Specifically, when the planning module receives information from the perception module regarding the material type of a soft obstacle, the "PathPlanning" function will no longer consider obstacle information for path planning, and the "DecisionMaking" function will not use obstacle information for decision-making. For hard obstacles, the functionality of the "PathPlanning" and "DecisionMaking" functions remains unchanged. By making these modifications to Apollo's perception and planning modules, when a soft obstacle is detected, the autonomous vehicle can directly pass through it. Therefore, driving decision-making... It can be represented as:

[0070]

[0071] in, This is a function in Apollo that maintains the current driving trajectory. This is the obstacle avoidance function for Apollo.

[0072] To verify the performance of the optimized Apollo obstacle avoidance driving decision-making method, this patent further presents a hazard scene extraction method based on geometric lane relationships, extracting real-world hazard scenarios caused by emergency braking and sudden lane changes from the INTERACTION traffic dataset. INTERACTION is a widely used autonomous driving dataset that includes diverse traffic scenarios such as lane changes, intersections, and roundabouts. This dataset records time-frame information of traffic participants at a sampling frequency of 10Hz. In each time frame, information such as vehicle position, speed, and heading angle is recorded.

[0073] To extract typical scenarios from the INTERACTION traffic dataset, the dataset is divided into multiple time segments, each consisting of 10 frames. A method for extracting hazardous scenarios based on geometric lane relationships is presented. The method for extracting emergency braking scenarios first identifies two vehicles traveling in the same lane within a time segment. One vehicle is defined as the ego vehicle (EV), and the other as the following vehicle (FV). Since the INTERACTION dataset does not directly record the lane IDs of the vehicles, the line connecting the center points (LCCP) of the two vehicles is used to determine if they are in the same lane. Within each time segment, if the LCCP of the EV and FV is parallel to the y-axis, the two vehicles are considered to be traveling in the same lane. Then, the horizontal acceleration of the EV is calculated. and vertical acceleration .if and This indicates that the EV is performing emergency braking.

[0074] The method for extracting sudden lane change scenarios is to first identify two vehicles traveling in different lanes. For example, if the angle between the LCCP and the y-axis is between 15° and 45°, we consider these vehicles to be in different lanes. Furthermore, if, over a period of time, the horizontal and vertical distances between the EV and the adjacent lane vehicle (ALV) continuously decrease, and... This indicates that the EV is suddenly changing lanes. Therefore, methods for extracting hazardous scenes from the INTERACTION dataset based on geometric lane relationships, such as... Figure 4 As shown.

[0075] These collision scenarios were reproduced on the Apollo platform. The performance of Apollo's original obstacle avoidance decision-making method and ADOAA was compared under both emergency braking and collision scenarios. The simulation results are as follows: Figure 5 As shown. Figure 5 (a) shows a situation where Apollo detected an obstacle and made an emergency braking action, resulting in the vehicle rear-ending the vehicle in front. Figure 5 (b) It demonstrates that after using the ADOAA method to detect that the obstacle ahead was a soft obstacle, a decision was made to directly pass through the obstacle, thus avoiding a rear-end collision. Figure 5 (c) shows that after Apollo detected an obstacle, its original obstacle avoidance decision made a sudden lane change, resulting in a collision with a vehicle in the adjacent lane. Figure 5(d) demonstrates that after the ADOAA method detects that the obstacle ahead is a soft obstacle, the decision is made to directly pass through the obstacle, avoiding turning to adjacent vehicles.

[0076] The Collision Risk Estimation Level Map (CRELM) is used to further illustrate the improvement of autonomous driving safety performance using ADOAA. CRELM uses current vehicle driving information to predict future driving information, thereby providing a collision risk estimation level. Assumptions... Let be the distance traveled within a certain time interval, assuming that the speed and acceleration remain constant. It can be represented as:

[0077]

[0078] in, This indicates the distance traveled within each time interval. and These are the relative acceleration and velocity between the electric vehicle and other vehicles. Therefore, the time required to reach a specific location is... It can be represented as:

[0079]

[0080] generally These methods can only provide accurate short-term predictions. For unforeseen events, the accuracy of these predictions decreases. Therefore, advanced placeholder time is introduced, defined as:

[0081]

[0082] Here, q is an adjustable parameter. It can predict when a vehicle will arrive at a specific location and take into account current and future risk conditions. Therefore, collision risk estimation levels... Defined as The reciprocal of is expressed as:

[0083]

[0084] As can be seen from the above formula, when hour, , and when Reducing it to 0 indicates a collision has occurred; substituting this into the equation yields... For the sake of convenience, the maximum risk value is set to 5, i.e., when... At that time, the risk collision level was The most dangerous and safest spaces are assigned risk levels of 5 and 0, respectively. The risk level within the two-dimensional space is determined using the following formula, taking into account the dimensions of obstacles and surrounding vehicles:

[0085]

[0086]

[0087]

[0088] in, This indicates the width of the vehicle.

[0089] CRELM when making driving decisions using the original OAA, such as Figure 6 As shown in the figure, the collision risk estimation level increases as the AV approaches the obstacle. Increase. In sudden lane change scenarios, the EV increases lateral acceleration to initiate a lane change to avoid obstacles. At this time, the ALV's driving area is precisely at... Within this range, it indicates a high level of collision risk between the two vehicles. In an emergency braking scenario, since the obstacle, the EV, and the preceding vehicle (FV) are in the same lane, the collision only occurs in the vertical direction. The distance between the EV and the obstacle... Between the EV and the car in front Therefore, when an EV brakes suddenly to avoid an obstacle, the vehicle in front will inevitably collide with the EV.

[0090] CRELM when using ADOAA for driving decisions, such as Figure 7 As shown. After ADOAA, the collision risk estimation level is [not specified] in both lane change and emergency braking scenarios. This indicates that there is no risk of collision for the vehicle in either of these scenarios, effectively improving the system's safety performance.

[0091] The above-described embodiments further illustrate the purpose, technical solution, and advantages of the present invention. It should be understood that the above-described embodiments are merely preferred embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made to the present invention within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A safety obstacle avoidance method based on obstacle material detection for autonomous driving, characterized in that, include: Millimeter-wave echo signals from different obstacles are collected using onboard millimeter-wave radar in autonomous vehicles. Wavelet transform is performed on the millimeter-wave echo signal to obtain the noise-removed signal; The denoised signal is subjected to a Fast Fourier Transform to obtain frequency domain information; feature extraction is performed on the frequency domain information to obtain 10 feature values; Ten feature values ​​are input into a convolutional neural network to obtain the classification result of the signal; The classification results are input into the Apollo autonomous driving system for obstacle material identification, and the obstacle avoidance scheme is optimized through the planning module based on the material identification results; hazardous scene features are extracted from the INTERACTION real scene dataset using a hazardous scene extraction method based on geometric lane relationships. The extracted scene was simulated on the Apollo platform to verify the performance of the optimized obstacle avoidance method.

2. The safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, The signal after wavelet transform of the millimeter-wave echo signal is: ； in, This is the millimeter-wave echo signal after wavelet transform. To control the scaling parameters of wavelet frequencies, The position parameter is used to control the wavelet's displacement in time. Denotes the wavelet complex conjugate function. It is a millimeter-wave echo signal.

3. A safety obstacle avoidance method based on obstacle material detection for autonomous driving, as described in claim 2, is characterized in that, The wavelet transform uses the db8 wavelet type and performs a 5-level decomposition of the signal during the wavelet transform.

4. The safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, The denoised signal is then subjected to a Fast Fourier Transform: ； Where N is the number of sampling points. This is the millimeter-wave echo signal after wavelet transform.

5. A safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, The 10 characteristic values ​​are: center frequency, spectral bandwidth, spectral skewness, spectral kurtosis, dominant frequency, average frequency, median frequency, spectral entropy, mean square frequency, and frequency standard deviation.

6. A safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, A convolutional neural network contains two convolutional units, each containing a batch normalization layer, a ReLU activation function, and a convolutional layer; the first convolutional layer contains 16 filters, the second convolutional layer contains 32 filters, the fully connected layer contains 512 hidden units, and uses the sigmoid activation function for binary classification tasks.

7. A safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, Optimizing the obstacle avoidance decision-making module of the Apollo autonomous driving system based on classification results includes: inputting the classification results into the Apollo autonomous driving system, and using the ObjectDetection function of the Apollo perception module to label the obstacle material type; passing obstacle information containing material characteristics from the perception module to the prediction and planning module, whereby the prediction and planning module determines the material of the obstacle. If the obstacle is a soft material, the obstacle information is not used for vehicle operation decisions; if the obstacle is a hard material, the obstacle information is used for vehicle operation decisions.

8. A safety obstacle avoidance method based on obstacle material detection for autonomous driving, as described in claim 7, is characterized in that, Vehicle operation decisions are made using the PathPlanning and DecisionMaking functions.

9. A safety obstacle avoidance method based on obstacle material detection for autonomous driving according to claim 1, characterized in that, The hazardous scene extraction method based on geometric lane relationships extracts two types of hazardous scenes from the INTERACTION real-world scene dataset: collisions caused by sudden lane changes and emergency braking. The INTERACTION dataset was divided into multiple time segments, each containing 10 frames of information. Emergency braking scene extraction included: identifying two vehicles traveling in the same lane within a time segment, defining one vehicle as the primary vehicle (EV) and the other as the following vehicle (FV); determining whether the two vehicles were in the same lane by using the line of convergence distance (LCCP) connecting the center points of the two vehicles; specifically, if the LCCP of the EV and FV was parallel to the y-axis within each time segment, the two vehicles were considered to be traveling in the same lane; otherwise, they were not. The horizontal acceleration of the EV was then calculated. and vertical acceleration ;if and If the signal is positive, it means the EV is performing emergency braking; otherwise, the EV is driving normally. Extracting sudden lane change scenarios includes: identifying two vehicles traveling in different lanes; defining vehicles with an angle between the LCCP and the y-axis between 15° and 45° as vehicles in different lanes; and determining that the horizontal and vertical distances between the EV and vehicles in adjacent lanes continuously decrease over a period of time. This indicates that the EV is making a sudden lane change.

Images