Active perception system for dual-axle steering cabless mining vehicles

Abstract

This invention discloses an active perception system for dual-axle steering, cabless mining vehicles. The system comprises an environmental perception and mapping module, a single-frame target detection module, a multi-frame data association module, a cloud control platform, a decision-making and planning module, and a vehicle control module. The dual-axle steering, autonomous mining vehicle can actively adjust between forward and backward driving modes based on the combined driving mode and sensor information. This system improves mining efficiency and addresses the large blind spots and high risk inherent in existing autonomous driving perception systems when turning, ascending, and descending slopes.

Description

Technical Field

[0001] The present invention relates to the field of autonomous driving technology, and in particular to an active perception system suitable for a dual-axle steering cabless mining vehicle. Background Art

[0002] In recent years, the rapid development of artificial intelligence and next-generation information technology has driven further improvement in key autonomous driving technologies. Open-pit mines, with their closed roads and relatively simple environments, have become a prime candidate for rapid implementation of autonomous driving. However, existing autonomous driving processes in mines present the following challenges: First, the cab occupies space in the mine car, making it difficult to increase the number of operations per trip. Second, mining vehicles require frequent U-turns in loading and unloading areas, reducing operational efficiency. Furthermore, the large size of mine cars creates numerous blind spots when driving uphill, downhill, and around corners, reducing the performance of the vehicle's autonomous driving perception system.

[0003] Key autonomous driving technologies can be categorized into three key areas: environmental perception, planning and decision-making, and tracking control. The environmental perception system, as the foundation of autonomous driving systems, is an essential component. Dual-axle steering vehicles, with two front and rear steering axles, eliminate the need for reverse and facilitate steering. This makes them suitable for unmanned driving systems in mining areas, improving operational efficiency.

[0004] Dual-axle steering vehicles do not distinguish between the front and rear directions of the vehicle. The perception system of autonomous driving vehicles suitable for dual-axle steering should include at least two driving modes: front-forward mode and rear-forward mode. Currently, there is no relevant research on the perception system of dual-axle steering vehicles.

[0005] At the same time, most current autonomous driving perception systems are passive perception, and the detection range of fixed sensors is limited. The method proposed by this system combines the vehicle turning angle with the drivable area based on decision-making planning feedback and adaptively changes the sensor perception angle based on the slope sensor, which effectively reduces the perception blind spot and improves the perception ability of the perception system.

[0006] In addition, the target fusion module evaluates risks through the Bayesian estimation algorithm and uses feedback control to calculate the information entropy to reduce the driving risk of autonomous vehicles and improve the reliability and safety of the perception system.

[0007] Chinese patent publication number CN113665500A, titled "Environmental Perception System and Method for Unmanned Transport Vehicles Operating Around the Clock," primarily provides an environmental perception system and method for unmanned transport vehicles operating around the clock, achieving an environmental perception system for unmanned transport vehicles operating around the clock through a multi-sensor fusion method. The system primarily constructs a multi-sensor two-dimensional grid map for structured roads, and generates a two-dimensional fused grid map through decision-level fusion. This invention only perceives and maps the structured road environment through a fusion algorithm, and does not optimize the perception hardware system. After the drivable area is extracted, only the road in front of the vehicle can be perceived, which lacks practicality and real-time performance for the perception of multiple turns and uphill and downhill environments in mining areas. The adaptive perception system proposed in this invention adaptively changes the sensor detection direction to fill in blind spots in mining areas when turning and going uphill with a full load, thereby ensuring safe and reliable autonomous driving in mining areas.

[0008] Chinese patent publication number CN113551662A, the invention name is "Perception method, dynamic perception device, and autonomous driving vehicle for autonomous driving vehicles", which mainly provides a dynamic perception system that improves perception performance by fusing comprehensive data with data of the area of ​​interest. The dynamic perception system determines the area of ​​interest through comprehensive data, thereby determining the rotation direction of the movable sensor. This method is mainly aimed at the dynamic perception optimization of urban structured roads, and the sensor rotation based on comprehensive data has limitations and cannot be adapted to the operation of dual-axle steering autonomous driving vehicles in mining areas on mining roads. The adaptive perception system proposed in the present invention accurately controls the rotation direction of the sensor through the drivable area and planning feedback information, and fills the blind spots when turning in the mining area and going up and downhill with full load, so as to ensure the safety and reliability of autonomous driving in the mining area. Summary of the Invention

[0009] The purpose of this invention is to overcome the shortcomings and deficiencies of the existing technology and provide an autonomous driving perception system suitable for dual-axle steering vehicles, thereby improving mining operation efficiency and resolving the problems of existing autonomous driving perception systems such as large blind spots and high risks when turning, going uphill or downhill. The present invention adopts the following technical solutions:

[0010] An active perception system for dual-axle steering cabless mining vehicles, comprising: an environment perception and mapping module, a single-frame target detection module, a multi-frame data association module, a cloud control platform, a decision-making and planning module, and a vehicle control module;

[0011] The environmental perception and mapping module builds a 3D perception map and extracts regions of interest by marking unstructured road boundaries;

[0012] The single-frame target detection module performs detection work on single-frame data using various sensor detection modules, including but not limited to millimeter-wave radar detection using Euclidean clustering, lidar detection using deep learning algorithms, and camera detection using deep learning algorithms.

[0013] The multi-frame data association module calculates the information entropy of each sensor and collects indicators such as the mark, speed, and distance of each detected target to predict the risk index of the target. The risk index is fed back into the information entropy to ultimately determine the obstacle status.

[0014] The cloud control platform is used to receive the perceived environmental data and transmit it to the decision-making and planning module;

[0015] The decision-making and planning module determines the vehicle speed, slope, and turning angle through a corresponding planning algorithm, and transmits this information to the vehicle control module for vehicle control. At the same time, this information is fed back to adaptively control the rotation direction of each sensor to increase the effective sensing detection range.

[0016] The vehicle control module actively adjusts to forward-forward mode and rearward-forward mode based on the driving mode and comprehensive information from the sensors.

[0017] Furthermore, the dual-axle steering automatic driving mining vehicle has a front-forward mode and a rear-forward mode: the front-forward mode uses the front sensor of the vehicle body as the main sensor, and when the vehicle is fully loaded and travels from the loading area to the unloading area, the front-forward mode is activated; the rear-forward mode uses the rear sensor of the vehicle body as the main sensor, and when the vehicle is empty and travels from the unloading area to the loading area, the rear-forward mode is activated.

[0018] Furthermore, it includes a forward sensor component and a rearward sensor component of the same configuration, and the sensor component includes at least a slope sensor, a GNSS / IMU, a lidar, a camera, a millimeter-wave radar, and a rotation angle receiving module and a rotation angle control module of each sensor.

[0019] Furthermore, the environment perception and mapping module includes a device installation and calibration module, a point cloud map generation module, and a drivable area extraction module:

[0020] Equipment installation and calibration module: Two sets of lidar, millimeter-wave radar, and camera are installed on the front and rear of the dual-axle steering autonomous mining vehicle respectively. GNSS / IMU and slope sensor are installed on the vehicle body. The lidar and GNSS / IMU are calibrated to obtain the positional relationship between the vehicle body coordinate system and the world coordinate system.

[0021] Point cloud map generation module: This module uses the GPS of the first waypoint as the origin to create a local world map. GPS and LiDAR data are input, and the conversion matrix is ​​determined using the LiDAR and vehicle GNSS / IMU calibration information. The point cloud GPS information at each moment is then converted to the world coordinate system. The point cloud is then optimized through front-end ICP matching and back-end EKF filtering to ultimately form a local point cloud map.

[0022] Drivable area extraction module: Extracts unstructured road boundaries based on point cloud normal vectors, distinguishes between relatively vertical ground normal vectors and relatively inclined retaining wall normal vectors, extracts driving boundaries through the orientation characteristics of the normal vectors, and then generates a drivable area.

[0023] Furthermore, the single-frame target detection module includes a data acquisition module, an adaptive angle rotation module, a Euclidean clustering module for extracting millimeter-wave radar detection target information, a deep learning algorithm module for extracting lidar detection target information, and a deep learning algorithm module for extracting camera detection target information:

[0024] Data acquisition module: receives point cloud information of the drivable area, vehicle speed, slope and angle information, and obtains data output by millimeter-wave radar, lidar, and camera;

[0025] Adaptive Angle Rotation Module: Determines the sensor rotation angle based on the vehicle's turning angle information combined with the point cloud information of the drivable area. The rotation angle of each sensor is controlled by the rotation angle control module. When the vehicle body tilts or the center of gravity shifts, the sensor rotation angle is determined based on the output value of the slope sensor to compensate for the blind spot caused by the vehicle's pitch angle.

[0026] Euclidean clustering module extracts target information detected by millimeter-wave radar: After removing ground information from the drivable area, Euclidean clustering is performed on the point cloud on the road. Specifically, target points are randomly selected. If the distance to the current point is less than a threshold, it is placed in the cluster. Otherwise, another point is selected for judgment. This process is repeated until no points with a distance less than the threshold exist. The clustering results are then output.

[0027] The deep learning algorithm extracts target information from lidar detection. The module first cuts the original point cloud into a grid from a top-down perspective, pillarizes the 3D point cloud, and represents the point cloud within each pillar as a 9-dimensional vector. The point cloud is then fed into the PointNet feature extraction network to generate a 2D pseudo-image. This pseudo-image is then fed into a 2D CNN to further extract point cloud features. Finally, the SSD detection head performs BoundingBox regression to output the location and orientation of the obstacle.

[0028] The deep learning algorithm extracts camera detection target information module: First, mosaic data enhancement is used to change the size and position of the true value target. The enhanced image is input into the network and the Focus module is used to unify the image size. Then, the improved residual network is used to extract the deep semantic information and shallow spatial information of the image. Finally, the feature pyramid network is used to accurately detect the target category and position at three scales.

[0029] Furthermore, the multi-frame data association module includes a target fusion module and a risk level prediction module:

[0030] Target fusion module: Generates a detection target list based on the single-frame target detection information, fuses the target features output by each sensor through the Bayesian method, and obtains a consistent explanation of the target; calculates the given hypothesis H i When the sensor's observation feature E is true, the likelihood probability P(E|H i ), compare the observed value with the predetermined threshold th i , the number of expected targets, feature matching and the quality of feature matching, analyze the above data to determine the value of the likelihood function, and use this value to represent the target a j confidence level;

[0031] Risk level prediction module: Using target speed, own vehicle speed, label value, and target relative position to the vehicle as evaluation criteria, the expert system is used to input various relevant factors into the inference engine. The inference engine repeatedly infers through the algorithm and relevant rules in the knowledge base. When the expert system outputs that the risk score of the target is greater than the risk level threshold, it is considered that the detection risk level is high and the target a is actively updated. j confidence level.

[0032] Compared with the prior art, the present invention has the following beneficial effects:

[0033] An active perception system suitable for dual-axle steering cabless mining vehicles is proposed. The mode design is carried out according to the characteristics of dual-axle steering vehicles and the mining environment. Two modes are proposed for mining operating vehicles. By setting the steering mechanism in each mode, the working efficiency and stability of the mining area can be improved.

[0034] A perception method based on decision-making planning feedback information is proposed. The feedback information mainly includes vehicle angle information and slope sensor, which effectively reduces the perception blind spot of the vehicle when turning, going uphill and downhill, and improves the perception ability of the perception system.

[0035] A Bayesian fusion algorithm is proposed. It uses an expert system to perform risk assessment. According to the obtained risk level, it actively perceives and detects the target and improves the confidence of the target, thereby reducing the driving risk of autonomous vehicles and improving the reliability and safety of the perception system.

[0036] Other features and advantages of the embodiments of the present invention will be described in detail in the subsequent detailed description. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] The accompanying drawings are used to provide further understanding of the embodiments of the present invention and constitute a part of the specification. Together with the following specific embodiments, they are used to explain the embodiments of the present invention, but do not constitute a limitation of the embodiments of the present invention.

[0038] Figure 1 : Overall architecture diagram of active perception system;

[0039] Figure 2 : Schematic diagram of the impact of vehicle pitch angle on sensor sensing range;

[0040] Figure 3 : Ackermann steering geometry principle diagram;

[0041] Figure 4 : Schematic diagram of the impact of vehicle yaw angle on sensor sensing range;

[0042] Figure 5 : Stepper motor control system diagram;

[0043] Figure 6 : Bayesian algorithm flow chart;

[0044] Figure 7 : Flowchart of the multi-frame target association module algorithm. DETAILED DESCRIPTION

[0045] During vehicle travel, the perception system can proactively adjust front and rear wheel steering based on driving patterns and integrated sensor information. For example, when a fully loaded vehicle is traveling uphill, the weight of the cargo moves downhill, the weight of the front of the vehicle is less than the weight of the rear, and the vehicle tilts backward, which can easily lead to oversteering. The system then adaptively adjusts to rear-wheel steering. When traveling downhill, the weight of the cargo moves forward, the weight of the front of the vehicle is greater than the weight of the rear, and understeering can easily lead to front-wheel steering. The adaptive perception system for dual-axle steering autonomous mining vehicles is configured with the following modules:

[0046] (1) Perception and Mapping Module:

[0047] Step 1: Equipment installation and calibration

[0048] Two sets of lidar, millimeter-wave radar, and cameras are installed on the front and rear of the dual-axle steering autonomous driving operation vehicle respectively. GNSS / IMU and slope sensor are installed on the vehicle body. The lidar and GNSS / IMU are calibrated to obtain the positional relationship between the vehicle body coordinate system and the world coordinate system.

[0049] Step 2: Point cloud map generation

[0050] A local world map is created, using the GPS of the first waypoint as the origin. GPS and LiDAR data are simultaneously input, and the vehicle's GPS at each moment is converted to a world coordinate system. The conversion matrix is ​​determined using LiDAR and vehicle IMU calibration information. The GPS point cloud information at each moment is then converted to a world coordinate system. The point cloud at each moment is then subjected to front-end ICP matching and back-end EKF filtering optimization processing to ultimately form a local point cloud map.

[0051] Step 3: Extracting the drivable area

[0052] The unstructured road boundary is extracted based on the point cloud normal vector. The ground normal vector is relatively vertical, while the normal vector of the retaining wall is relatively inclined. The boundary is obtained through the orientation characteristics of the normal vector to generate the drivable area.

[0053] (2) Single-frame target detection module:

[0054] Step 1: Get data

[0055] Receive point cloud information of the drivable area, vehicle turning angle information, slope sensor information, and obtain data output by millimeter-wave radar, lidar, and camera.

[0056] Step 2: Adaptive Angle Rotation

[0057] 1) When the planning module continuously outputs vehicle turning angle information, it compares it with the point cloud information of the drivable area to determine the sensor rotation angle, and controls the sensor rotation angle through the sensor rotation controller;

[0058] When the matching degree between the current drivable area and the expected drivable area is continuously less than the threshold, that is:

[0059]

[0060]

[0061] Among them, i is the current moment, M i M is the matching degree between the current drivable area and the expected drivable area, x is the matching threshold, S n is the current drivable area, S p is the expected drivable area. When the above formula is satisfied, it is considered that the forward sensor meets the steering condition and can turn.

[0062] The turning angle derivation method based on the vehicle's safe braking distance is related to the road surface adhesion coefficient, weather conditions and other conditions. Due to the bumpy ground and unpredictable weather in the mining area, the turning angle deduced by this method fluctuates greatly and is not applicable. The patented invention uses a turning angle derivation method based directly on geometric relationships:

[0063] Assume that when the vehicle turns right, if the left wheel angle is θ l , the right wheel turning angle is θ r , the angle from the sensor to the intersection of the left and right front wheel axes is θ x K is the ground track of the left and right front wheel axles, and L is the ground track of the front and rear wheel axles on the same side of the vehicle. The distance from the left and right front wheel axles to the sensor is l, and the sensor should rotate by an angle of θ. According to Ackermann's steering law:

[0064]

[0065] like Figure 3 As shown:

[0066]

[0067]

[0068] The angle from the sensor to the intersection of the left and right front wheel axes is θ x for:

[0069]

[0070] The corresponding sensor emission source turning radius is:

[0071]

[0072] A, B, C - distances from the sensor emission source to the front and rear wheel axes, distances from the rear and rear wheel axes, and the location of the sensor emission source;

[0073] x——the distance from the sensor emission source to the right wheel;

[0074] l——The distance from the left and right front wheel axles to the sensor;

[0075] L - locomotive axle spacing;

[0076] O——the instantaneous turning center of the locomotive;

[0077] H - the distance from the locomotive's instantaneous turning center to the right wheel;

[0078] D - the point of tangency between the sensor axis and the corresponding wheel's intended driving trajectory.

[0079] Depend on Figure 4 It can be seen that:

[0080]

[0081] From the quadrilateral interior angle theorem we know that:

[0082] ∠BCD+∠BOD=π

[0083] and:

[0084] ∠BCD+θ=π

[0085] so:

[0086] θ=∠BOD=θ x +∠AOD=θ x +∠AOC+∠COD

[0087] From the Pythagorean theorem we know that:

[0088] OB 2 =OA 2 -AB 2 =R x 2 -L 2

[0089] OC 2 =OB 2 +BC 2 =R x 2 -L 2 +(L+l) 2 =R x 2 +l 2 +2lL

[0090] Among them, R x is the radius from the center of the locomotive's clockwise turning circle to the sensor's emission source. For ∠COD:

[0091]

[0092] For ∠AOC, according to the cosine theorem, we know that: AC=l;

[0093]

[0094] Then the sensor angle θ is:

[0095]

[0096] By changing the angle in advance according to the perception system, reducing the delay error and introducing the time variable t, the function of the adaptive rotation angle of the sensor at that location can be obtained:

[0097] R x θ(t)=∫V(t)dt

[0098] Among them, V(t) is the speed function output by the decision planning module;

[0099] Substitute this into the final sensor adaptive rotation angle function:

[0100]

[0101] 2) When the road encounters slightly larger rocks or puddles, or when the vehicle is fully loaded and traveling up or downhill, it may cause the vehicle to tilt forward or shift its center of gravity backward. The slope sensor outputs the slope value, and the sensor rotation angle is determined based on the output slope value. The sensor rotation controller controls the sensor rotation to compensate for the blind spot caused by the vehicle's pitch angle.

[0102] Depend on Figure 2 It can be seen that θ y is the angle that the vehicle's forward tilt sensor should be adjusted (positive values ​​are defined as upward adjustment and negative values ​​are defined as downward adjustment). α is the vehicle's forward tilt angle output by the slope sensor. When the center of gravity of the vehicle body sinks, the sensor's sensing position drops accordingly. β is the angle between the sensor's sensed drop point and the intended sensing position due to the drop in the vehicle's center of gravity:

[0103] θ y =-α-β

[0104] Since α is small, we can approximately get:

[0105]

[0106] Where h is the height change of the vehicle's forward tilt sensor, and w is the horizontal component of the distance between the sensor and the vehicle's center of gravity;

[0107] From the geometric relationship we can get:

[0108]

[0109] W is the effective sensing distance of the sensor, so we can get:

[0110]

[0111] The adaptive rotation sensor is driven by a small stepper motor. The stepper motor control system is considered to be a second-order inertia link plus a pure lag link. The stepper motor control system is shown in the figure below. Figure 5 As shown;

[0112] Figure 5 In the equation, N(s) is the number of input pulses of the stepper motor control system, and Y(s) is the output angle of the stepper motor control system. The transfer function of the stepper motor control system is:

[0113]

[0114] Where K is the static gain of the system, T is the system time constant, and τ is the pure lag time.

[0115] The mathematical model of the front sensor is established, and the least square method is used for system identification. The mathematical model of the time-invariant dynamic process is assumed to be:

[0116] A(z -1 )z(k)=B(z -1 )u(k)+n(k)

[0117] Where u(k) and z(k) are the input and output of the system; n(k) is the noise, A(z -1 ) and B(z -1 ) is a polynomial, where:

[0118]

[0119] Use the sequence {u(k),z(k);k=1,···,L} to estimate the polynomial A(z -1 ) and B(z -1 ) is the unknown coefficient in .

[0120] definition:

[0121]

[0122] The mathematical model of the initial time-invariant dynamic process can be transformed into a standard least squares format:

[0123] z(k)=h T (k)θ+n(k)k=1,2,···,L

[0124] The above formula can be written as a linear equation system, and its matrix format is:

[0125] Z L =H L θ+N L

[0126] Where:

[0127]

[0128]

[0129] Then the least squares criterion function J(θ) can be taken as:

[0130] J(θ)=(Z L -H L θ)I L (Z L -H L θ)

[0131] Among them, I L is a weighting matrix, which is generally a positive definite diagonal matrix. By minimizing J(θ), the estimated value of the coefficient θ can be obtained

[0132]

[0133] in is the estimated value of the least square method. Once the input and output values ​​of the system are obtained, the above formula can be used to obtain the estimated value of the corresponding coefficient at once.

[0134] To establish the mathematical model of the sensor angle control system, the least squares method is required for system identification. Therefore, the transfer function of the stepper motor control system needs to be Laplace transformed. Assuming that the input of the vehicle control system is a constant value x, the Laplace transformed relationship can be obtained:

[0135]

[0136] Where y(t) is the inverse Laplace transform of Y(s), and n is the number of input pulses of the control system. To separate the three unknown coefficients K, τ, and T, integrate the two ends to obtain:

[0137]

[0138] make The above formula can be simplified to:

[0139]

[0140] Where T s is the sampling period, k is a positive integer;

[0141] Arranged:

[0142]

[0143] make have to:

[0144]

[0145]

[0146] BX=A

[0147] X=(B T B) -1 B T A

[0148] In the above formula, the right side of the equal sign is only related to the input and output of the control system and the sampling time. These quantities can be obtained based on experiments, and the left side of the equal sign is the unknown coefficient.

[0149] A state space model is established. The transfer function of the stepper motor control system used in the model is:

[0150]

[0151] The procedure for establishing the state space model is as follows:

[0152] Nump=[0,0,2.93];

[0153] Denp = [1.77, 1, 0];

[0154] Gp=tf(nump,denp);

[0155] [nump,denp]=pade(0.06,1); %The pure delay part uses the first-order pade approximation

[0156] G=Gp*tf(nump,denp); %Reconstruct approximate function

[0157] [num,den] = tfdata(G,'v'); % Extract the numerator and denominator polynomial coefficient vectors of the approximate function

[0158] [A,B,C,D]=tf2ss(num,den); %Convert to state space model

[0159]

[0160] Y=[0 -1.6554 55.1789]X

[0161] Step 3: Euclidean clustering to extract target information detected by millimeter-wave radar

[0162] After removing ground information from the drivable area, Euclidean clustering is performed on the road point cloud. Specifically, a target point is randomly selected. If the distance to the current point is less than a threshold, it is added to the cluster. Otherwise, another point is selected for evaluation. This process is repeated until no points with distances less than the threshold remain. The clustering results are then output.

[0163] Step 4: Deep learning algorithm extracts lidar detection target information

[0164] First, the original point cloud is cut into a grid from a top-down perspective. The 3D point cloud is pilled and the point cloud within each pillar is represented as a 9-dimensional vector. PointNet is used to extract features from the point cloud and generate a 2D pseudo image. This pseudo image is then fed into a 2D CNN to further extract point cloud features. Finally, the SSD detection head performs BoundingBox regression to output the location and orientation of obstacles.

[0165] Step 5: Deep learning algorithm extracts camera detection target information

[0166] First, mosaic data augmentation is used to change the size and position of ground-truth objects. The mosaic-enhanced image is fed into the network, where it is resized using the Focus module. The improved residual network, DarkNet, extracts both deep semantic information and shallow spatial information. Finally, a feature pyramid network accurately detects object categories and positions at three scales.

[0167] (3) Multi-frame target association

[0168] Step 1: Target Fusion

[0169] Generate a list of detection targets based on the single-frame target detection information, fuse the target features output by each sensor through the Bayesian method, and obtain a consistent explanation of the target; calculate the given hypothesis H i When the sensor's observation feature E is true, the likelihood probability P(E|H i ), compare the observed value with the predetermined threshold th i , the number of expected targets, feature matching and the quality of feature matching, analyze these data to determine the value of the likelihood function, and use this value to represent the target a j confidence level;

[0170] Step 2: Risk level prediction

[0171] The evaluation criteria are based on factors such as target speed, own vehicle speed, tag value, and the target's relative position to the vehicle. First, a knowledge-based expert system is established. The expert system described in this invention comprises a knowledge base, a global database, and an inference engine. The knowledge base contains established facts, algorithms, and heuristic rules; the global database temporarily caches input, intermediate, and output results; and the inference engine infers the target's danger level.

[0172] The expert system inputs various relevant factors into the inference engine, which then repeatedly infers the relevant target risk level through the algorithms and relevant rules in the knowledge base, and finally outputs the optimal target risk level. Set the risk level threshold h th When the expert system outputs a risk score greater than the threshold, the detection risk level is considered high. According to the risk score of the target, the target a is actively updated. j The probability feedback acts on step 2 to realize the target active risk perception function;

[0173] When the target's risk level is greater than the risk threshold:

[0174] h i >h th

[0175] Among them, h i is the target risk level, h th is the risk level threshold;

[0176] The target risk level is input into the expert system, which proactively analyzes targets that are greater than the risk threshold and increases the confidence level of the target:

[0177] a j =a j +(1-a j )h i

[0178] Among them, a j is the confidence of the target input, h i To increase the risk level coefficient, the expert system continuously absorbs prior knowledge and updates iteratively.

[0179] Furthermore, an embodiment of the present invention also provides a computer-readable storage medium having a computer program stored thereon, which implements the method described in the embodiment of the present invention when the program is executed by a processor.

[0180] The above describes in detail the optional implementation methods of the embodiments of the present invention in conjunction with the accompanying drawings. However, the embodiments of the present invention are not limited to the specific details in the above implementation methods. Within the technical concept of the embodiments of the present invention, various simple modifications can be made to the technical solutions of the embodiments of the present invention, and these simple modifications all fall within the scope of protection of the embodiments of the present invention.

[0181] It should also be noted that the various specific technical features described in the above specific embodiments can be combined in any appropriate manner without contradiction. To avoid unnecessary repetition, the embodiments of the present invention will not further describe various possible combinations.

[0182] Those skilled in the art will understand that all or part of the steps in the above-mentioned embodiments can be implemented by instructing the relevant hardware through a program, which is stored in a storage medium and includes a number of instructions for causing a single-chip microcomputer, chip or processor to execute all or part of the steps of the methods described in each embodiment of the present application. The aforementioned storage medium includes: a USB flash drive, a mobile hard disk, a read-only memory (ROM), a random access memory (RAM), a magnetic disk or an optical disk, and other media that can store program code.

[0183] In addition, various implementations of the embodiments of the present invention may be arbitrarily combined, and as long as they do not violate the concept of the embodiments of the present invention, they should also be regarded as the contents disclosed in the embodiments of the present invention.

Claims

1. An active sensing system suitable for a dual-axle steering cabless mining vehicle, characterized in that: include: Environmental perception and mapping module, single-frame target detection module, multi-frame data association module, cloud control platform, decision-making and planning module, and vehicle control module; The environmental perception and mapping module builds a 3D perception map and extracts regions of interest by marking unstructured road boundaries; The single-frame target detection module performs detection work on single-frame data using various sensor detection modules, including but not limited to millimeter-wave radar detection using Euclidean clustering, lidar detection using deep learning algorithms, and camera detection using deep learning algorithms. The multi-frame data association module calculates the information entropy of each sensor and collects the mark, speed, and distance of each detected target to predict the risk index of the target. The risk index is fed back into the information entropy to ultimately determine the obstacle status. The cloud control platform is used to receive the perceived environmental data and transmit it to the decision-making and planning module; The decision-making and planning module determines the vehicle speed, slope, and turning angle through a corresponding planning algorithm, and transmits this information to the vehicle control module for vehicle control. At the same time, this information is fed back to adaptively control the rotation direction of each sensor to increase the effective sensing detection range. The vehicle control module actively adjusts to forward-forward and backward-forward modes based on the combined information of the driving mode and sensors; The single-frame target detection module includes a data acquisition module, an adaptive angle rotation module, a Euclidean clustering module for extracting millimeter-wave radar detection target information, a deep learning algorithm module for extracting lidar detection target information, and a deep learning algorithm module for extracting camera detection target information: Data acquisition module: receives point cloud information of the drivable area, vehicle speed, slope and angle information, and obtains data output by millimeter-wave radar, lidar, and camera; Adaptive Angle Rotation Module: Determines the sensor rotation angle based on the vehicle's rotation angle information combined with the point cloud information of the drivable area, and controls the rotation angle of each sensor through the rotation angle control module; When the vehicle body tilts or the center of gravity shifts, the sensor's rotation angle is determined based on the output value of the slope sensor to compensate for the blind spot caused by the vehicle's pitch angle; Euclidean clustering module extracts target information detected by millimeter-wave radar: After removing ground information from the drivable area, Euclidean clustering is performed on the point cloud on the road. Specifically, target points are randomly selected. If the distance to the current point is less than a threshold, it is placed in the cluster. Otherwise, another point is selected for judgment. This process is repeated until no points with a distance less than the threshold exist. The clustering results are then output. The deep learning algorithm extracts target information from lidar detection. The module first cuts the original point cloud into a grid from a top-down perspective, pillarizes the 3D point cloud, and represents the point cloud within each pillar as a 9-dimensional vector. The point cloud is then fed into the PointNet feature extraction network to generate a 2D pseudo-image. This pseudo-image is then fed into a 2D CNN to further extract point cloud features. Finally, the SSD detection head performs BoundingBox regression to output the location and orientation of the obstacle. The deep learning algorithm extracts camera detection target information module: First, mosaic data enhancement is used to change the size and position of the true value target. The enhanced image is input into the network and the Focus module is used to unify the image size. Then, the improved residual network is used to extract the deep semantic information and shallow spatial information of the image. Finally, the feature pyramid network is used to accurately detect the target category and position at three scales.

2. The active sensing system according to claim 1, characterized in that: The dual-axle steering autonomous mining vehicle has a forward-forward mode and a rear-forward mode. In the forward-forward mode, the front sensor of the vehicle body is used as the main sensor. When the vehicle is fully loaded and travels from the loading area to the unloading area, the forward-forward mode is activated. The rear-end forward mode uses the rear sensor of the vehicle body as the main sensor. When the vehicle is unloaded and travels from the unloading area to the loading area, the rear-end forward mode is activated.

3. The active sensing system according to claim 1, characterized in that: It includes a forward sensor component and a backward sensor component with the same configuration. The sensor component includes at least a slope sensor, a GNSS / IMU, a lidar, a camera, a millimeter-wave radar, and a rotation angle receiving module and a rotation angle control module of each sensor.

4. The active sensing system according to claim 1, characterized in that: The environment perception and mapping module includes a device installation and calibration module, a point cloud map generation module, and a drivable area extraction module: Equipment installation and calibration module: Two sets of lidar, millimeter-wave radar, and camera are installed on the front and rear of the dual-axle steering autonomous mining vehicle respectively. GNSS / IMU and slope sensor are installed on the vehicle body. The lidar and GNSS / IMU are calibrated to obtain the positional relationship between the vehicle body coordinate system and the world coordinate system. Point cloud map generation module: This module uses the GPS of the first waypoint as the origin to create a local world map. GPS and LiDAR data are input, and the conversion matrix is ​​determined using the LiDAR and vehicle GNSS / IMU calibration information. The point cloud GPS information at each moment is then converted to the world coordinate system. The point cloud is then optimized through front-end ICP matching and back-end EKF filtering to ultimately form a local point cloud map. Drivable area extraction module: Extracts unstructured road boundaries based on point cloud normal vectors, distinguishes between relatively vertical ground normal vectors and relatively inclined retaining wall normal vectors, extracts driving boundaries through the orientation characteristics of the normal vectors, and then generates a drivable area.

5. The active sensing system according to claim 1, characterized in that: The multi-frame data association module includes a target fusion module and a risk level prediction module: Target fusion module: Generates a detection target list based on single-frame target detection information, fuses the target features output by each sensor through the Bayesian method, and obtains a consistent interpretation of the target; Calculate when given hypothesis H i When the sensor's observation feature E is true, the likelihood probability P(E|H i ), compare the observed value with the predetermined threshold th i , the number of expected targets, feature matching and the quality of feature matching, analyze the above data to determine the value of the likelihood function, and use this value to represent the target a j confidence level; Risk level prediction module: Using target speed, own vehicle speed, label value, and target relative position to the vehicle as evaluation criteria, the expert system is used to input various relevant factors into the inference engine. The inference engine repeatedly infers through the algorithm and relevant rules in the knowledge base. When the expert system outputs that the risk score of the target is greater than the risk level threshold, it is considered that the detection risk level is high and the target a is actively updated. j confidence level.

Images