An AI-based train obstacle detection system and method

Abstract

This invention discloses an artificial intelligence-based train obstacle detection system and method, relating to the field of obstacle recognition and management technology. The invention monitors the phenomenon of a multi-sensor fusion model continuously outputting different target information for the same obstacle, analyzes the changes in effective warning data that can trigger train warnings or braking commands from the changing target information, and identifies and judges the sensing sensors that are more sensitive to the collection of feature information that can trigger train warnings or braking commands. Simultaneously, based on the aforementioned sensitive feature information collection phenomena of each sensing sensor, the invention evaluates and calculates the important indices of each sensing sensor in the train obstacle detection process, assisting management personnel in formulating and carrying out corresponding performance maintenance work for the sensing sensors installed on the train.

Description

Technical Field

[0001] This invention relates to the field of obstacle recognition and management technology, specifically an artificial intelligence-based train obstacle detection system and method. Background Technology

[0002] As a high-capacity public transportation system that ensures the operation of cities, urban rail transit carries millions of passengers every day. Its safe and reliable operation is related to the safety of public life and property and is the core cornerstone of the stable operation of urban transportation networks. Among them, perception sensors, as the "sensory nerves" of obstacle detection systems, provide the basis for triggering train warnings and braking commands.

[0003] With the rapid popularization of fully automated driverless technology in rail transit, there is a greater reliance on autonomous obstacle detection. This trend further amplifies the importance of maintaining the performance of sensing sensors. However, current traditional train obstacle detection technology and its associated sensor management models are insufficient in "extending sensor lifespan and providing continuous data support for train safety." Specific problems are as follows: First, in the traditional model, the maintenance of sensing sensors often adopts a passive approach of "uniform periodic inspection" (such as monthly full disassembly and inspection) or "repair after failure," without differentiated management based on the actual role of the sensors in obstacle detection. Second, in train obstacle detection systems, the "importance" of sensors should be directly linked to their "actual contribution to safety warnings," which needs to be quantitatively evaluated through indicators such as "data contribution frequency." However, the traditional model has not established a relevant evaluation system. Summary of the Invention

[0004] The purpose of this invention is to provide an artificial intelligence-based train obstacle detection system and method to solve the problems raised in the prior art.

[0005] To achieve the above objectives, the present invention provides the following technical solution: a train obstacle detection method based on artificial intelligence, the method comprising:

[0006] Step S1: Establish an electronic map of the train route, pre-record point cloud maps and feature maps of the train route, and achieve real-time positioning of the entire train line using SLAM technology; install several perception sensors at the front of the train, construct a multi-sensor fusion model, and use the multi-sensor fusion model to identify and warn of obstacles in real time within the track clearance area in front of the train; the perception sensors include lidar, millimeter-wave radar, inertial navigation sensors, and video sensors;

[0007] Step S2: Collect each historical warning message that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of characteristic warning messages of the train;

[0008] Step S3: If the multi-sensor fusion model detects that there are two different target information outputs for an obstacle within the track area boundary in front of the train, extract the two historical operation records generated by the multi-sensor fusion model when it outputs two different target information, and construct a target operation node.

[0009] Step S4: Based on the trend of information similarity changes between the sets of different target information and feature warning information continuously output by the multi-sensor fusion model for the same obstacle at each target operation node, the corresponding comprehensive warning coefficient is evaluated and calculated for each target operation node.

[0010] Step S5: Based on the changes in the comprehensive warning coefficient presented at each target operating node, select the characteristic operating nodes, and based on each characteristic operating node, perform feature marking processing on the corresponding sensing sensors that have data impact on issuing warnings or braking commands to the train.

[0011] Step S6: Based on all feature operation nodes, analyze the distribution of each sensing sensor marked with features, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

[0012] Furthermore, step S2 includes: extracting features from each historical warning information to obtain several feature information sets; extracting the intersection between the several feature information sets, or extracting the set formed by feature information that appears more than once in the several feature information sets, and setting the intersection or set as the train's feature warning information set; the feature warning information set contains high-frequency feature information that can trigger the train to issue a warning or braking command; that is, if the above feature information is extracted from the corresponding target information output by the multi-sensor fusion model, the probability of triggering the train to issue a warning or braking command is relatively high.

[0013] Furthermore, step S4 includes:

[0014] Step S4-1: If target information A and B are extracted from a historical running record P(A) and a historical running record P(B) within a target running node respectively, and the historical running record P(A) was generated earlier than the historical running record P(B); feature extraction is performed on target information A and B respectively to obtain feature information sets D(A) and D(B) for target information A and B respectively;

[0015] Step S4-2: Extract the first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and the second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for a target running node; for each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B), obtain the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn;

[0016] Step S4-3: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for a certain target running node:

[0017] β 1 =θ1×θ2×...×θ n ;

[0018] Where θ1, θ2, ..., θ n Let F(A) represent the 1st, 2nd, ..., nth features in the first distinguishing feature information set F(A), respectively, and the highest similarity values ​​obtained after calculating the similarity with each feature in the feature warning information set. For a target running node, calculate the second comprehensive warning coefficient based on the second distinguishing feature information set F(B):

[0019] β 2 =θ1'×θ2'×...×θ m ';

[0020] Among them, θ1', θ2',..., θ m 'Represents the highest similarity value presented after calculating the similarity between the 1st, 2nd, ..., mth features in the second distinguishing feature information set F(B) and each feature in the feature warning information set.

[0021] Furthermore, step S5 includes:

[0022] Step S5-1: When β is satisfied in a certain target running node 2 >β 1 , and β 2 -β 1 When the threshold value is greater than η, a target running node is determined to be a characteristic running node; where η represents the threshold value; indicating that in the target running node, the target information output by the multi-sensor fusion model for the same obstacle is increasingly close to the historical warning information that can trigger the train to issue a warning or braking command.

[0023] Step S5-2: Extract the datasets input by each sensor to the multi-sensor fusion model from the two historical operation records contained in each feature operation node before the multi-sensor fusion model continuously outputs two not completely identical target information based on the same obstacle. This shows that in the target operation node, the target information continuously output by the multi-sensor fusion model for the same obstacle is getting closer and closer to the historical warning information that can trigger the train to issue a warning or braking command. That is, there is a process of increasing effective data collection for the same obstacle among several sensor sensors. The effective data here mainly refers to the feature data of the obstacle that objectively exists and poses a danger to the operation of the train.

[0024] Step S5-3: Suppose that dataset Q is extracted from two historical operation records contained in a certain feature operation node for a certain sensing sensor. 1 and dataset Q 2 When dataset Q 1 and dataset Q 2 When the similarity between two points is less than the similarity threshold, a feature label is made on a certain sensing sensor based on a certain feature running node. This means that the above process of increasing the effective data collected for the same obstacle is likely due to the fact that the feature-labeled sensing sensor inputs more effective data into the multi-sensor fusion model. In other words, the sensing sensor is more sensitive to the collection of feature information that can trigger the train to issue a warning or braking command.

[0025] Furthermore, step S6 includes:

[0026] Step S6-1: The total number of feature running nodes is K. The total number of feature markings for each sensing sensor is G. The interval between each two adjacent feature markings for each sensing sensor is monitored to obtain the average interval T between feature markings for each sensing sensor.

[0027] Among them, the longer the T corresponding to the sensing sensor, the higher the probability that the phenomenon of the sensing sensor feeding more effective data into the multi-sensor fusion model among all sensing sensors is accidental.

[0028] Step S6-2: Calculate the importance index δ=(G / K)×(1 / T) of each sensing sensor in the process of detecting obstacles on the train; sort all sensing sensors from largest to smallest according to their respective importance indices, generate a sensor sequence, and prompt the management personnel terminal to formulate the interval cycle for carrying out corresponding performance maintenance work according to the ranking value of each sensing sensor in the sensor sequence.

[0029] To better implement the above methods, a train obstacle detection system is also proposed. The system includes: a multi-sensor fusion model management module, a feature warning information extraction management module, a target operation node construction management module, a comprehensive warning coefficient evaluation and calculation module, a feature label processing module, and an importance index evaluation module.

[0030] The multi-sensor fusion model management module is used to create an electronic map of the train route, pre-record point cloud maps and feature maps of the train route, and achieve real-time positioning of the entire train line by using SLAM technology; several perception sensors are installed at the front of the train to build a multi-sensor fusion model, and obstacle identification and early warning are performed in real time within the track area boundary in front of the train through the multi-sensor fusion model.

[0031] The feature warning information extraction and management module is used to collect each historical warning information that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of feature warning information of the train.

[0032] The target operation node construction and management module is used to extract two historical operation records corresponding to the two different target information output by the multi-sensor fusion model when it detects that the multi-sensor fusion model outputs two different target information in a certain obstacle within the track area limit range in front of the train, and constructs and generates a target operation node.

[0033] The comprehensive early warning coefficient evaluation and calculation module is used to evaluate and calculate the corresponding comprehensive early warning coefficient for each target operation node based on the trend of information similarity changes between the multi-sensor fusion model's continuous output of different target information and feature early warning information sets for the same obstacle at each target operation node.

[0034] The feature labeling processing module is used to filter out feature operating nodes based on the changes in the comprehensive early warning coefficient presented in each target operating node, and to perform feature labeling processing on the corresponding sensing sensors that have data impact on issuing early warning or braking commands to the train based on each feature operating node.

[0035] The important index evaluation module is used to analyze the distribution of each sensing sensor marked with features based on all feature operation nodes, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

[0036] Furthermore, the target running node construction and management module includes: a target information output monitoring and management unit and a target running node construction and management unit;

[0037] The target information output monitoring and management unit is used to monitor the phenomenon that the multi-sensor fusion model outputs two different target information continuously for an obstacle within the track area boundary in front of the train.

[0038] The target running node construction and management unit is used to extract two historical running records generated when the multi-sensor fusion model outputs two not completely identical target information, and to construct and generate a target running node.

[0039] Furthermore, the feature labeling processing module includes a feature running node filtering unit and a feature label judgment processing unit;

[0040] The feature operation node filtering unit is used to filter out feature operation nodes based on the changes in the comprehensive early warning coefficient presented in each target operation node;

[0041] The feature labeling and processing unit is used to analyze each feature operation node and perform feature labeling processing on the corresponding sensing sensors that have data impact on issuing warning or braking commands to the train.

[0042] Compared with existing technologies, the beneficial effects of this invention are as follows: By monitoring the phenomenon of a multi-sensor fusion model continuously outputting different target information for the same obstacle, the invention analyzes the changes in effective warning data that can trigger train warnings or braking commands from the changing target information. It identifies and judges the sensing sensors that are more sensitive to the collection of feature information that can trigger train warnings or braking commands. At the same time, based on the above-mentioned sensitive feature information collection phenomenon of each sensing sensor, the invention evaluates and calculates the important index of each sensing sensor in the process of train obstacle detection. This assists the management terminal in formulating and carrying out corresponding performance maintenance work for the sensing sensors installed on the train, improving the service life of the sensing sensors, and providing data support for train operation safety. Attached Figure Description

[0043] Figure 1 This is a flowchart illustrating an artificial intelligence-based train obstacle detection method according to the present invention.

[0044] Figure 2 This is a schematic diagram of the structure of an artificial intelligence-based train obstacle detection system according to the present invention. Detailed Implementation

[0045] Based on the embodiments of this invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this invention.

[0046] Example: Figures 1-2As shown, this invention provides a train obstacle detection method based on artificial intelligence, the method comprising:

[0047] Step S1: Establish an electronic map of the train route, for example, an electronic map of the entire train route on the ground, in tunnels, and within the factory area; pre-record point cloud maps and feature maps of the train route, and use SLAM technology to achieve real-time positioning of the entire train line; install several perception sensors at the front of the train, construct a multi-sensor fusion model, and use the multi-sensor fusion model to identify and warn of obstacles in real time within the track clearance area in front of the train; the perception sensors include lidar, millimeter-wave radar, inertial navigation sensors, and video sensors;

[0048] For example, creating an "electronic map" of the entire train route, including ground surfaces, tunnels, and factory areas.

[0049] Step S2: Collect each historical warning message that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of characteristic warning messages of the train;

[0050] The multi-sensor fusion model can identify obstacles within the clearance range ahead of the train, including the train ahead, pedestrians, and small obstacles, and output information such as the type of obstacle ahead and the distance between the train and the obstacle.

[0051] After the multi-sensor fusion model identifies and outputs foreign objects on the train track, the train's travel control terminal will comprehensively evaluate the train's status, such as braking force, running speed, running direction and scene status. If it confirms that there is danger, it will issue different warnings or braking commands according to different distances.

[0052] Step S2 includes: extracting features from each historical warning information to obtain several feature information sets; extracting the intersection between the several feature information sets, or extracting the set formed by feature information that appears more than once in the several feature information sets, and setting the intersection or set as the train's feature warning information set.

[0053] Furthermore, it should be noted that:

[0054] Data characteristics: The set contains three types of warning characteristics, all of which are core parameters that trigger train safety responses:

[0055] (1) Obstacle type characteristics: such as "pedestrians", "track-left equipment (such as bolts, tools)" and "large foreign objects (such as billboard fragments, animals)" and other target category labels that can threaten driving safety;

[0056] (2) Distance threshold features: such as "the distance between the train and the obstacle is ≤50 meters (high-speed operation scenario)" and "the distance between the train and the obstacle is ≤30 meters (low-speed station entry scenario)" and other distance parameters (dynamically adapted according to the train's operating speed);

[0057] (3) Size threshold characteristics: such as "obstacle height ≥ 30cm (exceeds track clearance)" and "obstacle lateral width ≥ 50cm (occupies effective track space)" and other physical size parameters.

[0058] Inclusion criteria: Following the extraction rules in step S2 of the instruction manual, namely:

[0059] (1) If the feature sets of multiple historical early warning information have an intersection (e.g., all information that triggers braking contains "distance ≤ 30 meters"), then the intersection features are included;

[0060] (2) If there is no clear intersection, extract the feature “repeated occurrences > number of occurrences threshold (supplementary threshold: ≥3 times)” (e.g., if a “pedestrian intrusion” warning appears 4 times in the history of a certain line, its “pedestrian type + distance ≤ 40 meters” feature is included in the set).

[0061] The role of the "highest similarity value" is clarified: The "highest similarity value" is an indicator that measures the "matching degree between the target information output by the multi-sensor fusion model and the early warning features." Its role is to support step S4, "calculation of comprehensive early warning coefficient," and step S5, "screening of feature operation nodes." Specific functions are supplemented as follows:

[0062] (1) Calculation logic: For each feature in the first distinguishing feature set F(A) and the second distinguishing feature set F(B) (such as “distance = 60 meters” in F(A) and “distance = 45 meters” in F(B)), calculate the similarity with the features in the “feature warning information set” (such as “distance ≤ 50 meters”), and take the highest matching value of each feature (i.e., “highest similarity value”).

[0063] (2) Function: Calculate β by multiplying the "highest similarity value". 1 (The first comprehensive early warning coefficient of F(A)) and β 2 (The second comprehensive early warning coefficient of F(B)) can quantify whether "the change of target information from A to B is closer to the early warning condition"—for example, the similarity between "distance = 60 meters" and the early warning feature "≤ 50 meters" in F(A) is 0.3, while the similarity between "distance = 45 meters" in F(B) is 0.8. The difference in their highest similarity values ​​directly leads to β 2 >β 1 This provides data support for subsequent judgments on whether "target information is approaching the warning state".

[0064] Step S3: If the multi-sensor fusion model detects that there are two different target information outputs for an obstacle within the track area boundary in front of the train, extract the two historical operation records generated by the multi-sensor fusion model when it outputs two different target information, and construct a target operation node.

[0065] Step S4: Based on the trend of information similarity changes between the sets of different target information and feature warning information continuously output by the multi-sensor fusion model for the same obstacle at each target operation node, the corresponding comprehensive warning coefficient is evaluated and calculated for each target operation node.

[0066] Step S4 includes:

[0067] Step S4-1: If target information A and B are extracted from a historical operation record P(A) and a historical operation record P(B) within a target operation node respectively, and the historical operation record P(A) was generated earlier than the historical operation record P(B), this time order is the design of the present invention to capture "obstacle information dynamically iterates as the train approaches", which affects the logical validity of feature difference extraction, similarity calculation and warning coefficient comparison; feature extraction is performed on target information A and B respectively to obtain feature information sets D(A) and D(B) of target information A and B respectively;

[0068] Step S4-2: Extract the first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and the second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for a target running node; for each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B), obtain the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn;

[0069] Step S4-3: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for a certain target running node:

[0070] β 1 =θ1×θ2×...×θ n ;

[0071] Where θ1, θ2, ..., θ n Let F(A) represent the 1st, 2nd, ..., nth features in the first distinguishing feature information set F(A), respectively, and the highest similarity values ​​obtained after calculating the similarity with each feature in the feature warning information set. For a target running node, calculate the second comprehensive warning coefficient based on the second distinguishing feature information set F(B):

[0072] β 2 =θ1'×θ2'×...×θ m ';

[0073] Among them, θ1', θ2',..., θ m 'Represents the highest similarity value presented after calculating the similarity between the 1st, 2nd, ..., mth feature information in the second distinguishing feature information set F(B) and each feature information in the feature warning information set;

[0074] Furthermore, it should be noted that:

[0075] For example: When a train is running along a track and continuously detects the same obstacle, the sensor fusion model outputs two records successively:

[0076] Historical operation record P(A) (generation time: T0, train 80 meters from obstacle): Target information A is "suspected obstacle (type ambiguous), distance 78-82 meters (error ±4 meters), size about 20-40cm (low accuracy)";

[0077] Historical operation record P(B) (Generation time: T1, T1=T0+2 seconds, train 50 meters from obstacle): Target information B is "track leftover tool (type clearly defined), distance 49 meters (error ±1 meter), size 35cm (high precision)".

[0078] Impact on feature difference extraction:

[0079] Since P(A) occurs earlier than P(B), the time difference between the two corresponds to the physical process of the train approaching the obstacle (from 80 meters to 50 meters). Therefore, the feature differences (such as "type ambiguity → clarity", "distance error reduction", "size accuracy improvement") reflect the rule that "obstacle information becomes more accurate as the distance decreases" rather than random differences. If the time order is reversed (P(A) occurs later than P(B)), the feature differences may manifest as "information accuracy decreases", which contradicts the actual physical process and makes subsequent calculations unreasonable.

[0080] Impact on similarity calculation:

[0081] The similarity calculation result between the feature set F(A) of P(A) and the "feature warning information set" is β. 1 =0.3 (due to ambiguity in information, the matching degree is low); the similarity calculation result of the feature set F(B) of P(B) is β. 2 =0.8 (due to accurate information and high matching degree), the time sequence ensures β 2 >β 1 The difference (0.5) can reflect the "obstacle information approaching warning condition" and provide a basis for subsequent judgment.

[0082] Impact on feature running node selection:

[0083] β determined based on time sequence 2 -β 1 =0.5>threshold η=0.2, the node is determined to be a feature running node, and then the sensors with improved performance during T0 to T1 (such as LiDAR, because the point cloud accuracy is improved after the distance is shortened) are selected. If the time order is ignored, nodes with "decreased information accuracy" may be mistakenly determined as valid nodes, resulting in the error of sensor maintenance strategy.

[0084] Another example: If the target information for P(A) (T0, 100 meters) is "no obstacles" and the target information for P(B) (T1, 70 meters) is "pedestrian intrusion", the time sequence ensures that the difference between the two ("no → present") reflects the actual process of the obstacle's appearance. Its β... 2 -β 1 =0.9>η, triggering an alert; if the time is reversed, it may be misjudged as "obstacle disappearance", leading to a missed alert.

[0085] Step S5: Based on the changes in the comprehensive warning coefficient presented at each target operating node, select the characteristic operating nodes, and based on each characteristic operating node, perform feature marking processing on the corresponding sensing sensors that have data impact on issuing warnings or braking commands to the train.

[0086] Step S5 includes:

[0087] Step S5-1: When β is satisfied in a certain target running node 2 >β 1 , and β 2 -β 1 When the threshold value is greater than η, a target running node is determined to be a feature running node; where η represents the threshold value.

[0088] Step S5-2: Extract the datasets input by each sensor to the multi-sensor fusion model from the two historical running records contained in each feature running node, before the multi-sensor fusion model outputs two not completely identical target information based on the same obstacle.

[0089] Step S5-3: Suppose that dataset Q is extracted from two historical operation records contained in a certain feature operation node for a certain sensing sensor. 1 and dataset Q 2 , where Q 1 To perceive the raw data collected by the sensor within 1 second before the multi-sensor fusion model outputs target information A, Q 2The data represents the raw data acquired by the sensor within one second before the model outputs target information B. Both data reflect the data changes of the sensor during the obstacle information iteration process. When the dataset Q... 1 and dataset Q 2 When the similarity between two nodes is less than the similarity threshold, a feature label is applied to a certain sensor based on a certain feature node.

[0090] Furthermore, it should be noted that:

[0091] Dataset Q 1 Q 2 Specific definitions and data content

[0092] (a) Meaning and source of data

[0093] Q 1 Q 2 All data are raw data collected by sensors targeting the same obstacle; the only difference is the time of data collection.

[0094] Q 1 : Corresponding to the historical operation record P(A), which is the continuous data collected by the sensor within 1 second before the model outputs target information A (early fuzzy information);

[0095] Q 2 : Corresponds to the historical operation record P(B), which is the continuous data collected by the sensor within 1 second before the model outputs the target information B (later accurate information).

[0096] (ii) Dataset content and quantity for different sensors

[0097] The contents and quantity of the LiDAR sensor dataset are shown below:

[0098] Data meaning: Includes parameters such as the three-dimensional spatial coordinates (x, y, z) of the obstacle, point cloud reflection intensity, and point cloud density;

[0099] Data quantity: The sampling rate is 10Hz (1 frame is collected every 0.1 seconds), with a total of 10 frames of data per second. Each frame contains 150,000 to 200,000 point cloud data (the data size of a single frame is about 500KB to 800KB).

[0100] Example: Q 1 (1 second before output A) is "10 frames of point cloud, with a point cloud density of 80 points / m2 per frame (point cloud is missing in some areas due to dust obstructing the lens)", Q 2 (1 second before output B) is "10 frames of point cloud, with a point cloud density of 200 points / m2 per frame (after cleaning the lens, the point cloud completely covers the obstacle)".

[0101] The dataset contents and quantity of the millimeter-wave radar sensor are shown below:

[0102] Data meaning: Includes parameters such as distance to obstacles (radial distance), radial velocity, angles (azimuth and pitch), and target signal-to-noise ratio;

[0103] Data quantity: The sampling rate is 50Hz (1 set of data is collected every 0.02 seconds), with a total of 50 sets of data per second. Each set of data contains complete parameters of 32 target detection points (the size of a single set of data is about 1KB).

[0104] Example: Q 1 For "50 sets of data, the average target signal-to-noise ratio is 15dB (due to radar calibration offset, some data have a distance error of ±0.5 meters)", Q 2 The data consists of 50 sets of data, with an average target signal-to-noise ratio of 28 dB (after radar calibration, the range error was reduced to ±0.1 meters).

[0105] The content and quantity of the video sensor dataset are shown below:

[0106] Data meaning: Includes parameters such as RGB image frames of obstacles, target contour coordinates, color histogram, and motion trajectory;

[0107] Data quantity: Frame rate is 25fps (1 frame per 0.04 seconds), a total of 25 frames per second, and the resolution of a single frame is 1920×1080 (data volume is about 3MB / frame).

[0108] Example: Q 1 For "25 frames of image, the grayscale value of the obstacle outline is blurred due to lens backlighting (grayscale difference ≤ 20)", Q 2 The requirement is "25 frames of images, with backlight compensation enabled, the grayscale difference of the obstacle outline is ≥50 (the outline is clearly distinguishable)";

[0109] Core threshold range, setting basis and examples

[0110] (a) η(β) 2 -β 1 (difference threshold)

[0111] Parameter range: 0.15-0.3

[0112] Setting basis: β 1 β 2 This is the product of the highest similarity values ​​(ranging from 0 to 1). This range has been experimentally verified: when the difference is ≥ 0.15, it can distinguish the "target information significantly approaching warning condition" (such as β). 2 =0.8, β 1 =0.6, difference 0.2) and "small fluctuations in information" (such as β) 2 =0.65, β 1=0.55, difference 0.1); the upper limit is set at 0.3 because the maximum difference of "effective early warning approaching scenario" in the experiment did not exceed 0.3 (such as β). 2 =0.9, β 1 =0.6, difference 0.3), exceeding this value may be abnormal data (such as misjudgment caused by sudden sensor failure).

[0113] For example:

[0114] In a certain target running node, β 1 =0.12 (Comprehensive early warning coefficient of F(A)), β 2 =0.504 (Comprehensive early warning coefficient of F(B)), β 2 -β 1 =0.384. Since 0.384 > 0.3 (the upper limit of η), it is necessary to check whether there is any abnormal data in the sensor.

[0115] In another target running node, β 1 =0.4, β 2 =0.6, β 2 -β 1 =0.2 (within the range of 0.15-0.3), it is determined to be a feature running node and enters the subsequent sensor marking process.

[0116] (ii) Similarity threshold (Q) 1 With Q 2 (Similarity threshold)

[0117] Parameter range: 0.6-0.8

[0118] Basis for setting: This threshold is set based on the acquisition accuracy characteristics of different types of sensors: the similarity of raw data from high-precision sensors such as LiDAR and millimeter-wave radar needs to be ≥0.8 (high data stability requirements), while the similarity threshold for video sensors is set to ≥0.6 due to the greater influence of light; experimental data shows that when Q 1 With Q 2 When the similarity is less than the threshold, more than 90% of the scenarios are "effective increments in sensor-collected data" (such as more complete point cloud data after lens cleaning), which can be used as a basis for sensor labeling.

[0119] For example:

[0120] LiDAR sensor: Q 1 Point cloud data (with dust on the lens) and Q 2 The similarity (after cleaning) is 0.75, which is lower than 0.8 (the similarity threshold for LiDAR). Therefore, the sensor is determined to be a "sensor that affects the data of early warning" and is marked as a feature.

[0121] Video sensor: Q1 Image data and Q (backlit scene) 2 The similarity (after enabling backlight compensation) is 0.58, which is lower than 0.6 (the similarity threshold for video sensors), so feature labeling is also performed on this sensor.

[0122] (iii) Frequency threshold (the threshold for the number of repetitions of the feature warning information set)

[0123] Parameter range: ≥3 times

[0124] Setting basis: Based on historical data, the threshold for the number of occurrences is set to ≥3 times.

[0125] For example:

[0126] In the historical early warning data of a certain line, the feature "pedestrian entering the track area + distance ≤ 40 meters" appeared 5 times, exceeding the threshold of 3 times, and was included in the "feature early warning information set";

[0127] The feature “small stones (size ≤ 5cm) + distance ≤ 20 meters” appeared only twice, which did not reach the threshold and was therefore not included in the set (because small stones do not pose a safety threat).

[0128] The above threshold range can be fine-tuned according to specific application scenarios (such as high-speed trains, subways, and light rail in factory areas).

[0129] Step S6: Based on all feature operation nodes, analyze the distribution of each sensing sensor marked with features, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

[0130] Step S6 includes:

[0131] Step S6-1: The total number of feature running nodes is K. The total number of feature markings for each sensing sensor is G. The interval between each two adjacent feature markings for each sensing sensor is monitored to obtain the average interval T between feature markings for each sensing sensor.

[0132] Step S6-2: Calculate the importance index δ=(G / K)×(1 / T) of each sensing sensor in the process of detecting obstacles on the train; sort all sensing sensors from largest to smallest according to their respective importance indices, generate a sensor sequence, and prompt the management personnel terminal to formulate the interval cycle for carrying out corresponding performance maintenance work according to the ranking value of each sensing sensor in the sensor sequence.

[0133] To better implement the above methods, a train obstacle detection system is also proposed. The system includes: a multi-sensor fusion model management module, a feature warning information extraction management module, a target operation node construction management module, a comprehensive warning coefficient evaluation and calculation module, a feature label processing module, and an importance index evaluation module.

[0134] The multi-sensor fusion model management module is used to create an electronic map of the train route, pre-record point cloud maps and feature maps of the train route, and achieve real-time positioning of the entire train line by using SLAM technology; several perception sensors are installed at the front of the train to build a multi-sensor fusion model, and obstacle identification and early warning are performed in real time within the track area boundary in front of the train through the multi-sensor fusion model.

[0135] The feature warning information extraction and management module is used to collect each historical warning information that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of feature warning information of the train.

[0136] The target operation node construction and management module is used to extract two historical operation records corresponding to the two different target information output by the multi-sensor fusion model when it detects that the multi-sensor fusion model outputs two different target information in a certain obstacle within the track area limit range in front of the train, and constructs and generates a target operation node.

[0137] The target running node construction and management module includes: a target information output monitoring and management unit and a target running node construction and management unit.

[0138] The target information output monitoring and management unit is used to monitor the phenomenon that the multi-sensor fusion model outputs two different target information continuously for an obstacle within the track area boundary in front of the train.

[0139] The target running node construction and management unit is used to extract two historical running records generated when the multi-sensor fusion model outputs two not completely identical target information, and to construct and generate a target running node.

[0140] The comprehensive early warning coefficient evaluation and calculation module is used to evaluate and calculate the corresponding comprehensive early warning coefficient for each target operation node based on the trend of information similarity changes between the multi-sensor fusion model's continuous output of different target information and feature early warning information sets for the same obstacle at each target operation node.

[0141] The feature labeling processing module is used to filter out feature operating nodes based on the changes in the comprehensive early warning coefficient presented in each target operating node, and to perform feature labeling processing on the corresponding sensing sensors that have data impact on issuing early warning or braking commands to the train based on each feature operating node.

[0142] The feature labeling processing module includes a feature running node filtering unit and a feature labeling judgment processing unit.

[0143] The feature operation node filtering unit is used to filter out feature operation nodes based on the changes in the comprehensive early warning coefficient presented in each target operation node;

[0144] The feature labeling and processing unit is used to analyze each feature operation node and perform feature labeling processing on the corresponding sensing sensors that have data impact on issuing warning or braking commands to the train.

[0145] The important index evaluation module is used to analyze the distribution of each sensing sensor marked with features based on all feature operation nodes, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

[0146] Finally, it should be noted that the above descriptions are merely preferred embodiments of the present invention and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art can still modify the technical solutions described in the foregoing embodiments or make equivalent substitutions for some of the technical features. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A train obstacle detection method based on artificial intelligence, characterized in that: The method includes: Step S1: Establish an electronic map of the train route, pre-record the point cloud map and feature map of the train route, and achieve real-time positioning of the entire train line by using SLAM technology; install several perception sensors at the front of the train, construct a multi-sensor fusion model, and use the multi-sensor fusion model to identify and warn of obstacles in real time within the track area boundary in front of the train. Step S2: Collect each historical warning message that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of characteristic warning messages of the train; Step S3: If the multi-sensor fusion model detects that there are two different target information outputs for an obstacle within the track area boundary in front of the train, extract the two historical operation records generated by the multi-sensor fusion model when it outputs two different target information, and construct a target operation node. Step S4: Based on the trend of information similarity changes between the sets of different target information and feature warning information continuously output by the multi-sensor fusion model for the same obstacle at each target operation node, the corresponding comprehensive warning coefficient is evaluated and calculated for each target operation node. Step S5: Based on the changes in the comprehensive warning coefficient presented at each target operating node, select the characteristic operating nodes, and based on each characteristic operating node, perform feature marking processing on the corresponding sensing sensors that have data impact on issuing warnings or braking commands to the train. Step S6: Based on all feature operation nodes, analyze the distribution of each sensing sensor marked with features, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

2. The train obstacle detection method based on artificial intelligence according to claim 1, characterized in that: Step S2 includes: extracting features from each of the historical warning information to obtain several feature information sets; extracting the intersection between the several feature information sets, or extracting a set of feature information that appears more than once in the several feature information sets, and setting the intersection or the set as the train's feature warning information set.

3. The train obstacle detection method based on artificial intelligence according to claim 2, characterized in that: Step S4 includes: Step S4-1: If target information A and B are extracted from a historical running record P(A) and a historical running record P(B) within a target running node respectively, and the historical running record P(A) was generated earlier than the historical running record P(B); feature extraction is performed on target information A and B respectively to obtain feature information sets D(A) and D(B) for target information A and B respectively; Step S4-2: Extract a first set of distinguishing feature information F(A) = D(A) - D(A) ∩ D(B) and a second set of distinguishing feature information F(B) = D(B) - D(A) ∩ D(B) for the target running node; for each feature information in the first set of distinguishing feature information F(A) and the second set of distinguishing feature information F(B), obtain the highest similarity value presented after calculating the similarity with each feature information in the feature warning information set in turn; Step S4-3: Calculate the first comprehensive early warning coefficient based on the first distinguishing feature information set F(A) for the target running node: b 1 =θ1×θ2×...×θ n ; Where θ1, θ2, ..., θ n Let F(A) represent the 1st, 2nd, ..., nth features in the first distinguishing feature information set F(A), respectively, and let F(A) represent the highest similarity values ​​obtained after calculating the similarity between F(A) and each feature in the feature warning information set; let F(A) represent the second comprehensive warning coefficient based on the second distinguishing feature information set F(B) for the target running node. b 2 =θ1'×θ2'×...×θ m '; Among them, θ1', θ2',..., θ m 'Represents the highest similarity value presented after calculating the similarity between the 1st, 2nd, ..., mth feature information in the second distinguishing feature information set F(B) and each feature information in the feature warning information set.

4. The train obstacle detection method based on artificial intelligence according to claim 3, characterized in that, Step S5 includes: Step S5-1: When β is satisfied in the target running node 2 >β 1 , and β 2 -β 1 When the threshold value is greater than η, the target running node is determined to be a feature running node; where η represents the threshold value. Step S5-2: Extract the datasets input by each sensor to the multi-sensor fusion model from the two historical running records contained in each feature running node, before the multi-sensor fusion model continuously outputs two not completely identical target information based on the same obstacle. Step S5-3: Suppose that dataset Q is extracted from two historical operation records contained in a certain feature operation node for a certain sensing sensor. 1 and dataset Q 2 When dataset Q 1 and dataset Q 2 When the similarity between two nodes is less than a similarity threshold, the node performs a feature labeling on the sensor based on the feature.

5. The train obstacle detection method based on artificial intelligence according to claim 4, characterized in that: Step S6 includes: Step S6-1: The total number of feature running nodes is K. The total number of feature markings for each sensing sensor is G. The interval between each two adjacent feature markings for each sensing sensor is monitored to obtain the average interval T between feature markings for each sensing sensor. Step S6-2: Calculate the importance index δ=(G / K)×(1 / T) of each sensing sensor in the process of detecting obstacles on the train; sort all sensing sensors from largest to smallest according to their respective importance indices, generate a sensor sequence, and prompt the management personnel terminal to determine the interval cycle for carrying out corresponding performance maintenance work according to the ranking value of each sensing sensor in the sensor sequence.

6. A train obstacle detection system, used to execute the artificial intelligence-based train obstacle detection method according to any one of claims 1-5, characterized in that, The system includes: a multi-sensor fusion model management module, a feature early warning information extraction management module, a target operation node construction management module, a comprehensive early warning coefficient evaluation and calculation module, a feature labeling processing module, and an important index evaluation module; The multi-sensor fusion model management module is used to establish an electronic map of the train route, pre-record point cloud maps and feature maps of the train route, and achieve real-time positioning of the entire train line by using SLAM technology; several perception sensors are installed at the front of the train to construct a multi-sensor fusion model, and obstacle identification and early warning are performed in real time within the track area boundary in front of the train through the multi-sensor fusion model. The feature warning information extraction and management module is used to collect each historical warning information that triggers the train to issue a warning or braking command from the output of the multi-sensor fusion model, and extract the set of feature warning information of the train. The target running node construction and management module is used to extract two historical running records corresponding to the multi-sensor fusion model when it detects that the multi-sensor fusion model outputs two different target information for an obstacle within the track area boundary in front of the train, and constructs a target running node. The comprehensive early warning coefficient evaluation and calculation module is used to evaluate and calculate the corresponding comprehensive early warning coefficient for each target operation node based on the trend of information similarity change between the different target information and feature early warning information sets continuously output by the multi-sensor fusion model for the same obstacle in each target operation node. The feature marking processing module is used to filter out feature operating nodes based on the changes in the comprehensive early warning coefficient presented in each target operating node, and to perform feature marking processing on the corresponding sensing sensors that have data impact on issuing early warning or braking commands to the train based on each feature operating node. The important index evaluation module is used to sort out the distribution of each sensing sensor marked with features based on all feature operation nodes, evaluate and calculate the important indices of each sensing sensor in the process of train obstacle detection, and assist the management terminal in formulating corresponding performance maintenance strategies for each sensing sensor.

7. A train obstacle detection system according to claim 6, characterized in that: The target running node construction and management module includes: a target information output monitoring and management unit and a target running node construction and management unit; The target information output monitoring and management unit is used to monitor the phenomenon that the multi-sensor fusion model outputs two different target information continuously for an obstacle within the track area boundary in front of the train. The target running node construction and management unit is used to extract two historical running records generated when the multi-sensor fusion model outputs two not completely identical target information, and to construct and generate a target running node.

8. A train obstacle detection system according to claim 6, characterized in that: The feature marking processing module includes a feature running node filtering unit and a feature marking judgment processing unit; The feature running node filtering unit is used to filter out feature running nodes based on the changes in the comprehensive early warning coefficient presented in each target running node. The feature labeling and processing unit is used to analyze each feature operation node and perform feature labeling processing on the corresponding sensing sensors that have data impact on issuing warnings or braking commands to the train.

Images