Non-blind area sensing system for intelligent network connection bus

Abstract

The invention relates to a vehicle-road collaborative non-blind area sensing system for intelligent network connection buses, which belongs to the technical field of intelligent traffic and vehicle-road collaborative sensing, and is characterized in that based on multi-dimensional data of vehicle side and road side sensors, a vehicle sensor sensing range, a fixed bus route and real-time pose information are combined; a blind area target is sensed and predicted in real time, the potential risk of the blind area target to a bus is dynamically evaluated, high-risk information is transmitted back to a bus end, non-blind area global environment sensing and early warning are achieved, a reliable sensing basis is provided for follow-up functions such as intelligent scheduling, safety early warning and automatic driving, and the system is suitable for popularization and application. And the sensing safety and the risk early warning capability of the bus in the complex urban road section are obviously improved.

Description

TECHNICAL FIELD

[0001] The application belongs to the technical field of intelligent transportation, and particularly relates to a blind area-free perception system for intelligent networked public buses. BACKGROUND

[0002] With the development of automatic driving and intelligent transportation technology, the real-time, accuracy and robustness of vehicle environmental perception are constantly improving. Currently, traditional automatic driving systems mainly rely on vehicle-mounted sensors such as cameras, laser radars, millimeter wave radars, etc., to collect environmental data around the vehicle for target detection and trajectory prediction. However, due to the physical installation position of the vehicle-mounted sensors, which are usually installed on the front or top of the vehicle, their field of view and coverage range are limited. Especially in the areas such as the side and rear of the vehicle, which are easily blocked, it is difficult to meet the real-time environmental perception requirements of all-around and no dead angle, which directly affects the completeness of automatic driving environmental perception and driving safety.

[0003] To break through the limitations of single-vehicle perception, vehicle-road cooperative perception technology has emerged. This technology deploys laser radars, cameras and other multi-source sensors on the roadside, combines with vehicle-mounted sensor data, and uses V2X communication modules to exchange information in real time, which can compensate for the blind area of the vehicle end, expand the perception range, and realize the fusion of cross-time and space dimension target detection results and vehicle state information. However, the existing vehicle-road cooperative system needs to transmit all the perception data on the roadside, which has a large amount of data and high requirements for communication bandwidth, which can easily cause network congestion and transmission delay. In addition, most of the current research and landing cases are mainly for passenger car scenarios. Intelligent networked public buses, as public transportation vehicles running on fixed routes and frequently crossing complex urban traffic flow, have highly predictable driving trajectories and routes, but also face more complex shielding and intersection scenarios. Especially in the queue at the intersection during the morning and evening peak hours, the intersection of special lanes and ordinary lanes, and the multi-lane parallel driving flow, the traditional vehicle-road cooperative scheme does not fully consider the coordination of public bus route characteristics and real-time planning trajectory, and it is difficult to accurately compensate and dynamically predict the blind area targets caused by road facilities or other vehicles.

[0004] In the prior art, Chinese Patent Publication No. CN115052266A discloses a vehicle blind area detection method and application system based on blockchain in a vehicle networking environment. The method first obtains the basic information of the target vehicle and the basic information of other traffic participants in its visual blind area, and automatically transmits the basic information of other traffic participants in the blind area to the target vehicle through the intelligent contract mode using vehicle networking technology. This method realizes the trusted interaction of information through the blockchain smart contract, and improves the security of data transmission. However, this method only realizes the collection and transmission of information, and does not perform spatio-temporal fusion processing on the basic data of traffic participants in the blind area, making it difficult for the target vehicle to respond effectively based on the obtained information, lacking comprehensive processing capability for perception information, and having a single function, making it difficult to achieve the goal of vehicle-road collaborative perception.

[0005] Chinese Patent Publication No. CN114155713A discloses a road intersection blind area early warning control method and system based on roadside crowd computing. The method collects the position information and driving state of motor vehicles through the crowd computing system deployed on the roadside, and predicts and analyzes the motion trajectories of pedestrians and vehicles, then sends warning information such as speed reduction to road intersection participants through signal lights and other facilities. This technical solution can improve the traffic safety of road intersection blind areas to some extent, but it only targets specific intersection scenarios and does not consider the diversity of blind area environments under different road conditions and road structures, lacking systematic modeling of blind area characteristics and their change rules, resulting in limited application scenarios and weak universality.

[0006] Chinese Patent Publication No. CN116580555A discloses a vehicle blind area collaborative perception system and method based on roadside sensors. The system uses roadside sensor perception information to identify and dynamically divide blind area regions for vehicles in different traffic scenarios, and accurately assesses the potential risks of each blind area in real time. Meanwhile, it performs spatio-temporal alignment and fusion of multi-source perception information, and combines a backhaul compensation mechanism to realize vehicle-road collaborative perception. This technical solution can provide more accurate perception environment information for vehicles, but its method is not suitable for intelligent networked buses, and does not optimize the predictable running trajectory of fixed route buses, making it difficult to use the route rules to predict blind area risks and allocate resources in advance, thus limiting its practicality in the intelligent networked bus scenario.

[0007] Therefore, there is an urgent need for a blind area-free perception technology for intelligent networked buses that can both complete the vehicle-end blind area and reduce communication redundancy, and dynamically predict and assess blind area targets using bus route characteristics, to meet the safety perception needs of buses in complex urban traffic flows. SUMMARY

[0008] In view of the deficiencies of the prior art, the purpose of the present application is to provide a blind area free perception system for intelligent networked public buses. Based on the multi-dimensional data of the vehicle end and the roadside sensors, combined with the vehicle sensor perception range, fixed public bus routes and real-time pose information, the system can realize real-time perception, prediction and dynamic evaluation of blind area targets and their potential risks to the bus, and can return high-risk information to the vehicle end, so as to realize blind area free global environment perception and early warning, provide a reliable perception basis for subsequent functions such as intelligent scheduling, safety warning and automatic driving, and significantly improve the perception safety and risk warning capability of the bus in complex urban road sections.

[0009] To achieve the above-mentioned purpose, the present application provides a blind area free perception system for intelligent networked public buses, which performs multi-modal target detection in parallel at the vehicle end and the roadside, fuses GNSS / IMU / visual odometry information at the vehicle end, calculates and sends the pose information of the vehicle end in real time, receives the vehicle end pose information and sensor parameters at the roadside unit, maps the detected targets to the vehicle coordinate system to complete the space-time synchronization, filters the blind area targets based on the field of view model of the vehicle end sensor, and performs short-term trajectory prediction on the blind area targets, calculates the route attention weight combined with the geometric characteristics of the fixed public bus route, and calculates the minimum distance and weighting combined with the blind area target prediction trajectory and the bus planning trajectory for real-time collision risk assessment, returns the high-risk target state information to the vehicle end for fusion, and generates blind area free global perception information. The system mainly includes: vehicle end perception and state fusion, roadside perception and space-time alignment, real-time pose estimation, blind area target identification, blind area target trajectory prediction, trajectory interaction and route attention weighting, collision risk assessment, information return and vehicle-road information fusion module.

[0010] The vehicle end perception and state fusion are as follows:

[0011] Environment perception data acquisition: deploy multi-modal sensors such as cameras, millimeter wave radars and laser radars on the bus, realize millisecond-level time alignment between sensors through PPS pulse signal and GNSS time synchronization module, and real-time collect image and point cloud data of the surrounding environment;

[0012] Target detection and fusion: for image data, use the improved YOLOv5 algorithm with adaptive anchor box mechanism and CIoU loss function to output target ID, category and two-dimensional bounding box; for point cloud data, use PointPillars algorithm to directly predict target three-dimensional coordinates and size; after time series filtering, the above results are fused to generate a vehicle end target list containing target ID, category, three-dimensional coordinates, velocity vector, detection confidence and timestamp;

[0013] Pose estimation and fusion: combined with GNSS, IMU and visual odometry information, use extended Kalman filter to fuse vehicle motion state and sensor data, and output 4x4 pose transformation matrix of the bus relative to the roadside coordinate system in real time And issue to the roadside unit through the V2X communication module.

[0014] The roadside perception is aligned in space and time, specifically as follows:

[0015] Roadside multi-modal target detection: deploy laser radar and surround-view camera on key road sections along the fixed route of the bus, real-time detect traffic participants, and generate a roadside target list , including target position, speed, acceleration and category label;

[0016] Sensor parameter and pose information receiving: obtain bus ID, vehicle-end sensor parameters, real-time pose transformation matrix and timestamp through C-V2X communication, wherein the sensor parameters include the inherent installation position, orientation of each sensor (such as camera, laser radar, millimeter wave radar) in the vehicle coordinate system, as well as the inherent field of view (FOV), effective detection distance, pitch angle and other parameters of the sensor itself;

[0017] Coordinate transformation and time synchronization: use the roadside static calibration matrix and the real-time pose transformation matrix issued by the vehicle end to convert the three-dimensional coordinates of the roadside target from the roadside coordinate system to the vehicle coordinate system, and complete the space-time alignment according to the timestamp, and output the aligned roadside perception result.

[0018] The real-time pose estimation is specifically as follows:

[0019] Vehicle-end pose calculation: receive the 4x4 pose transformation matrix and the corresponding timestamp issued by the vehicle-end through V2X communication. The matrix represents the real-time pose (including rotation and translation ) of the bus relative to the roadside world coordinate system. Based on this matrix, the accurate position and orientation of the bus in the roadside world coordinate system are calculated and updated in real time;

[0020] Vehicle-end sensor field of view modeling: according to the real-time pose of the bus and the sensor parameter information, calculate the instantaneous effective sensing range boundary (i.e. sensor field of view model) of each vehicle-end sensor in the roadside world coordinate system;

[0021] Perception range parameter output: finally, the calculated real-time pose information of the bus and the field of view boundary model parameters of each vehicle-end sensor in the roadside world coordinate system are sent to the blind area identification module.

[0022] The blind area target identification is specifically as follows:

[0023] Field of view model definition: establish a "point-in-field of view" detection model for each vehicle-mounted sensor in the vehicle coordinate system, and clearly define the FOV (field of view), detection distance, pitch angle and other parameters of each sensor;

[0024] Blind area judgment: for each aligned roadside target, if its coordinates are outside the field of view of a certain sensor, it is marked as a blind area target for that sensor, otherwise it is marked as a non-blind area target;

[0025] Blind area target list generation: extract all targets marked as "blind area" and generate a list containing target ID, category, three-dimensional coordinates, velocity vector, detection confidence and timestamp.

[0026] The blind area target trajectory prediction is as follows:

[0027] History state input: construct a history state sequence for the current position, velocity and acceleration of each target in the blind area target list;

[0028] LSTM trajectory prediction model: input the history state vector into the trained LSTM network to predict the continuous trajectory in the next 2-3 seconds;

[0029] Trajectory output: generate a trajectory sequence for each blind area target at subsequent time points and send it to the collision risk assessment module.

[0030] The collision risk assessment and information feedback are as follows:

[0031] Route attention weighting: load the bus route centerline geometry data, calculate the Euclidean distance from the blind area target point to the centerline and apply a decay function to obtain the attention weight;

[0032] Minimum distance calculation: calculate the minimum distance between the target prediction trajectory and the bus planning trajectory at each time point;

[0033] Risk index calculation: based on the attention weight and the minimum distance, calculate the risk index and classify the target as "high risk target" or "low risk target" according to the preset threshold;

[0034] Information feedback: for high-risk targets, transmit the target position, velocity, prediction trajectory and risk level in real time to the vehicle end through the V2X communication module; for low-risk targets, continuously transmit monitoring information to update their state and risk level.

[0035] The vehicle-road information fusion is as follows:

[0036] Vehicle-end fusion processing: the vehicle end receives the high-risk target information returned by the roadside and unifies it with its own multi-modal detection results to obtain a more complete and high-precision target state;

[0037] Blind area-free perception output: based on the fused multi-source target state, the vehicle end generates a blind area-free environment perception result covering the entire area, providing accurate environmental input for intelligent scheduling, safety warning and autonomous driving.

[0038] The beneficial effects of the present application are:

[0039] (1) By fusing the vehicle end multi-source sensor and V2X communication technology, the real-time generation and dynamic sharing of vehicle self-position and target detection information are realized. After receiving the sensor parameters and 4x4 pose matrix issued by the vehicle end, the spatio-temporal alignment and perception field boundary calculation are completed combined with the high-precision map and external parameter information, so as to effectively identify the blind area and improve the perception accuracy and coverage ability of the vehicle-road cooperative system in complex urban road environment.

[0040] (2) The LSTM model is introduced to predict the trajectory of the blind area target, and the risk assessment is combined with the line center line and the minimum distance. Only the high-risk target information is returned to the vehicle end, avoiding redundant data transmission, reducing the communication load, improving the information transmission efficiency and decision timeliness, providing more valuable perception input for the vehicle system, and helping to realize safer and more reliable automatic driving control and behavior decision.

[0041] (3) The target area is prioritized combined with the route attention weighting mechanism, which further improves the efficiency and response speed of the system in processing dynamic blind area information. The system is not only suitable for automatic driving buses, but also can serve traditional buses, effectively filling the perception blind area through vehicle-road cooperation, realizing real-time early warning of potential dangerous targets, and significantly improving the traffic safety and operation efficiency of the bus during operation. It has good application prospect and promotion value. BRIEF DESCRIPTION OF DRAWINGS

[0042] In order to more clearly illustrate the technical solutions implemented by the present application, the following will briefly introduce the drawings needed to be used in the implementation description.

[0043] Figure 1 The system structure diagram of the present application is a blind area perception system for intelligent networked buses.

[0044] Figure 2 The method flow chart of the present application is a collision risk assessment based on trajectory interaction and route attention in the embodiment.

[0045] Figure 3 The step flow chart of the system in the embodiment of the present application is a step flow chart of the system.

[0046] Figure 4 The step flow chart of the system in the embodiment of the present application is a step flow chart of the system.

[0047] Figure 5 The application scenario of the present application is a vehicle-road cooperative blind area perception system for intelligent networked buses. DETAILED DESCRIPTION

[0048] Embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that the relative arrangement of components and steps, numerical expressions, and numerical values set forth in these embodiments are not limiting to the scope of the present application unless otherwise specifically stated.

[0049] The following description of the embodiments is merely illustrative in nature and is in no way intended to limit the scope of the application or its application or uses.

[0050] Referring to Figures 1-5 , the embodiments of the present application provide a blind area perception system for intelligent networked public transportation, aiming to monitor and evaluate blind area targets in the traffic environment in real time through the cooperative work of vehicle-side and roadside, thereby improving the safety and effectiveness of the public transportation system. The method comprises the following steps:

[0051] Step 1: Vehicle-side multi-modal target detection and state information output:

[0052] Cameras, lidar and other multi-modal sensors are deployed on the vehicle side to realize data time alignment through a synchronization mechanism (PPS+GNSS time synchronization module) to collect image data and point cloud information of the surrounding environment in real time;

[0053] The vehicle-mounted processing unit uses deep learning algorithms for target detection and feature extraction on the synchronously collected images and point clouds. Among them, the image target detection module uses an improved YOLOv5 method that introduces an adaptive anchor box mechanism and a CIoU loss function. The adaptive anchor box mechanism automatically adjusts the anchor box size according to the training data, improving the detection accuracy of targets of different sizes. The CIoU loss function not only considers the overlapping area when calculating the loss of the predicted box and the real box, but also considers the center point distance and the aspect ratio, making the detection box regression more accurate, and can output the target ID, category and two-dimensional bounding box in real time;

[0054] ;

[0055] Among them, is the target bounding box in the image;

[0056] The point cloud target detection module uses the PointPillars method to divide the three-dimensional point cloud space into multiple vertical columnar regions, encode the features of the point cloud in each columnar region, extract features through a multi-layer perception machine, and directly predict the three-dimensional coordinates and bounding box size of the target;

[0057] ;

[0058] Among them, is the three-dimensional coordinate, is the target bounding box size;

[0059] Subsequently, the detection results of both are fused by time sequence filtering to extract target ID, category, three-dimensional coordinates, speed vector, detection confidence and timestamp, and generate a bus-end target list , obtaining complete state information of each bus-end detection target;

[0060] The target list includes target ID, category label, three-dimensional coordinates (x, y, z), speed vector, detection confidence and timestamp, etc.

[0061] The bus-end combines GNSS, IMU and visual odometry information, uses an extended Kalman filter method to fuse vehicle motion state and sensor data, GNSS provides vehicle position information, IMU measures vehicle acceleration and angular velocity in real time, and visual odometry calculates vehicle relative motion by analyzing continuous frame image feature matching relationship. EKF, through a nonlinear state space model, first predicts the current time pose according to the last time pose and IMU measurement value, and then updates the GNSS measurement value and visual odometry calculation result as observation value, and outputs the 4x4 pose matrix of the bus in the vehicle coordinate system in real time ;

[0062] Among them, is a rotation matrix, is a vehicle position translation vector, which is used to describe the attitude and position of the vehicle in the vehicle coordinate system;

[0063] The 4x4 pose matrix, sensor parameters and corresponding timestamp information are sent to the roadside unit through the V2X communication module;

[0064] Step 2: Roadside target detection and space-time alignment

[0065] The roadside unit is deployed on the key sections along the fixed route of the bus, and uses the same method as the bus-end to perform real-time target detection on traffic participants through laser radar and surround-view camera, generating a roadside target list , wherein each target rj includes the position, speed, acceleration and category label of the target;

[0066] The bus ID, bus-end sensor internal parameter, real-time pose information and timestamp are obtained through C-V2X communication;

[0067] The roadside static calibration matrix is used with the real-time pose matrix sent by the bus-end to convert the three-dimensional coordinates of the roadside target from the roadside coordinate system to the vehicle coordinate system according to the following formula:

[0068]

[0069] Among them, The three-dimensional coordinates of the roadside target in the roadside coordinate system, The three-dimensional coordinates of the roadside target converted to the vehicle-mounted coordinate system;

[0070] And according to the timestamp, interpolation or buffering processing is performed to realize space-time alignment, and finally the aligned roadside perception target result is output, and the list is transmitted to the blind area identification module;

[0071] Step 3: Real-time pose estimation

[0072] The roadside unit uses the 4×4 pose transformation matrix issued by the vehicle end to estimate the rotation matrix of the bus relative to the roadside station through multi-sensor fusion algorithm And the translation vector. After receiving the matrix, the roadside system takes it as the absolute position and orientation of the current frame of the bus, which provides a basis for subsequent space alignment and sensor modeling;

[0073] Based on the received pose transformation matrix and the installation parameters of the vehicle-mounted sensor (including installation position, sensor type, horizontal and vertical field of view FOV, maximum effective detection distance, pitch angle, etc.), an instantaneous effective perception range model is constructed for each sensor;

[0074] Taking a laser radar as an example, its installation position in the vehicle-mounted coordinate system is , the field of view angle is , and the maximum detection distance is . The field boundary is mapped from the vehicle-mounted coordinate system to the world coordinate system to form the boundary area in the roadside world coordinate system. This process is also applicable to other sensors such as cameras and millimeter wave radars, forming a complete set of multi-modal joint field boundary;

[0075] The real-time pose information of the bus obtained by the above modeling and the field boundary model of each sensor in the roadside world coordinate system are uniformly packaged as a parameter set and transmitted to the blind area identification module through an internal message channel. The parameter set includes: the position and orientation of the bus in the current world coordinate system, the type and ID of each sensor, the vertex set of the field boundary of each sensor, and the timestamp and other synchronization information.

[0076] Step 4: Blind area target identification

[0077] The blind area identification module receives the real-time pose of the vehicle end and the perception range boundary information of the vehicle-mounted sensor, as well as the aligned roadside target list;

[0078] For the target in the aligned perception target list, its position coordinates in the vehicle-mounted coordinate system are The is combined with the field boundary Spatial relationship judgment, i.e. "point-in-view" comparison:

[0079] If , it is determined that the target has been covered by the on-board sensor, and is marked as a "non-blind area" target;

[0080] Otherwise, for all , if , i.e. the target is not within the perception range of the on-board sensor, it indicates that the target is located in the blind area of the overall perception system at the vehicle end, and is marked as a "blind area" target;

[0081] wherein represents the field of view model of the i-th on-board sensor in the on-board coordinate system;

[0082] For all roadside targets in the blind area of the vehicle end field of view, extract to the blind area target list , which includes target ID, category, three-dimensional coordinates , speed vector, detection confidence, timestamp and other information.

[0083] Step 5: Blind area target trajectory prediction

[0084] For the current position, speed and acceleration of each target in the blind area target list, construct a historical state sequence, taking the two-dimensional position x, speed v and acceleration a of the target at the last T time points, and splice them into a state sequence .

[0085] Input the historical state vector into the trained LSTM network , and the LSTM network processes time series data through input gate, forget gate, output gate and memory cell, learns the motion law of the target, and predicts the state at the next H time points, with the calculation formula as follows:

[0086] .

[0087] wherein and are the weight parameters and bias vectors of the network, respectively;

[0088] According to the current position, speed and acceleration of each blind area target in the blind area target list, a trajectory prediction model based on LSTM is used to generate the continuous trajectory of the target in the next 2-3 seconds, with the calculation formula as follows:

[0089] .

[0090] wherein is the blind area target at time ​The state vector contains the two-dimensional position of the target. ,speed and acceleration T represents the number of historical time steps used as input to the LSTM model.

[0091] ;

[0092] in Let represent the prediction function of the trained LSTM network, where and These are the network's weight parameters and bias vector, respectively;

[0093] By predicting the trajectory of targets in the blind spot, the system outputs the trajectory sequence of each target in the next 2-3 seconds and sends it to the collision risk assessment module.

[0094] Step 6: Collision Risk Assessment Based on Trajectory Interaction and Route Attention

[0095] The roadside unit loads the bus route centerline geometry data sent by the V2X communication module and sets the route attention weight function:

[0096] ;

[0097] in, Indicates the target point in the blind zone Euclidean distance to the center line of the bus route. This is the attenuation coefficient. The closer the distance, the higher the attention weight.

[0098] The Euclidean distance from the target point in the blind zone to the center line is calculated using the above formula, and the attention weight is obtained by applying the decay function.

[0099] Through the V2X communication module, the planned trajectory of the bus is obtained in real time from the vehicle's motion planning module. The predicted trajectory of the blind spot target is then compared with the planned trajectory of the bus by minimum distance calculation. Specifically, for each moment... ,calculate:

[0100] ;

[0101] in, For buses The planned trajectory location at any given time. For blind spot targets exist The predicted trajectory position at any given time;

[0102] Based on line attention weight minimum distance Define risk indicators , the collision risk is calculated according to the following formula:

[0103] ;

[0104] wherein, is a scale coefficient for adjusting the influence of distance on risk, generally taking a value of 10-50;

[0105] According to the preset threshold (set according to the actual traffic scene and the safety requirements of bus operation, generally taking a value of 5-15), the blind area target is evaluated for collision risk. If is greater than the set threshold , it is marked as a "high-risk" target, otherwise it is marked as a "low-risk" target.

[0106] Step 7: Information feedback and fusion

[0107] The roadside unit transmits the state information (position, speed, predicted trajectory, risk level, etc.) of the high-risk blind area target list in real time to the vehicle end through the V2X communication module, and updates the target state and risk level in real time to ensure timely response to potential risks;

[0108] The vehicle end will receive the high-risk target list information and the vehicle end target list information for vehicle-road information fusion, and finally generate a blind area-free global environment perception result, providing comprehensive and accurate environmental data for intelligent scheduling, safety warning and automatic driving of vehicles.

[0109] In summary, through the cooperation of multi-modal sensors at the vehicle end and roadside, time and space alignment, blind area identification, trajectory prediction and risk assessment, etc., the blind area target can be discovered and predicted in real time, the potential risk of the target to the bus can be dynamically evaluated, and the high-risk information can be fed back in real time, realizing accurate perception and warning of blind area targets, and building a complete and efficient vehicle-road cooperative blind area-free perception system, significantly improving the safety and stability of intelligent networked buses.

[0110] The above content only illustrates the technical idea of the present application, and cannot limit the protection scope of the present application. Any modification made according to the technical idea of the present application on the basis of the technical solution falls within the protection scope of the claims of the present application.

Claims

1. A blind area free perception system for intelligent networked public transport, characterized in that: The vehicle end deploys multi-modal sensors and realizes millisecond-level time alignment through the PPS pulse signal and GNSS time synchronization module, real-time collects surrounding image and point cloud data, and performs target detection and feature extraction, generates a vehicle end target list through time sequence filtering and fusion, and combines GNSS, IMU and visual odometry information, uses an extended Kalman filter to output a 4x4 pose transformation matrix of the bus relative to the roadside coordinate system, and sends the matrix to the roadside unit through the V2X communication module; The roadside unit deploys a laser radar and a surround-view camera on the key road sections along the fixed route of the bus, real-time detects traffic participants to generate a roadside target list, receives the pose matrix and timestamp sent by the vehicle end, combines the roadside static calibration matrix to convert the three-dimensional coordinates of the roadside target to the vehicle coordinate system and performs time-space alignment, performs "point-in-view" judgment on the aligned roadside target based on the field of view model of each sensor of the vehicle end in the vehicle coordinate system to identify blind area targets and generate a blind area target list, constructs a state sequence for the historical position, speed and acceleration of the targets in the list and inputs the trained LSTM network to predict the continuous trajectory of the future 2-3 seconds, loads the center line geometry data of the bus route, calculates the route attention weight according to the Euclidean distance and a preset attenuation function, calculates the minimum distance between the blind area target prediction trajectory and the bus planning trajectory at each time, and combines the attention weight to calculate the risk index, marks the target with a risk index greater than a preset threshold as a high-risk target, and returns it to the vehicle end in real time through V2X communication, and the vehicle end fuses it with the detection result to generate a global environment perception result without blind area. 2.The intelligent network connection oriented bus blind area perception system according to claim 1, wherein: The vehicle end perception data acquisition module of the intelligent networked bus end acquires synchronous image and point cloud data through multi-modal sensors and a PPS+GNSS time synchronization module; The target detection module uses a method based on YOLOv5 and introducing an adaptive anchor mechanism and a CIoU loss function for image data, and outputs target ID, category and two-dimensional bounding box in real time; and uses a PointPillars columnar feature encoding network for point cloud data to directly predict target three-dimensional coordinates and size; After time sequence filtering and fusion, a vehicle end target list containing target ID, category, three-dimensional coordinates, speed vector, detection confidence and timestamp is generated; The intelligent networked bus end also combines GNSS, IMU and visual odometry information, fuses vehicle motion state and sensor observation through an extended Kalman filter, and outputs a 4x4 pose transformation matrix of the bus relative to the roadside coordinate system in real time, and sends the matrix, sensor internal participation timestamp to the roadside unit through the V2X communication module. 3.The intelligent network connection oriented bus non-blind area perception system according to claim 1, characterized in that: The roadside perception data acquisition module of the roadside perception unit generates a roadside target list containing target position, speed, acceleration and category label through a laser radar and a surround-view camera; The data alignment module receives the 4x4 pose matrix and timestamp sent by the vehicle end, converts the three-dimensional coordinates of the roadside target to the vehicle coordinate system using the roadside static calibration matrix, and performs time-space alignment according to the timestamp to output the aligned roadside perception result.

4. The blind area-free perception system for intelligent networked public transportation according to claim 3, characterized in that: The trajectory interaction and route attention weighting module loads the bus route center line geometry data, and calculates the route attention weight according to the Euclidean distance and the preset attenuation function. 5.The intelligent network connection oriented bus non-blind area perception system according to claim 3, characterized in that: The real-time pose estimation module fuses GNSS, IMU and visual odometry information through extended Kalman filtering, and outputs a 4×4 pose transformation matrix of the bus relative to the roadside coordinate system. 6.The intelligent network connection oriented bus non-blind area perception system according to claim 3, characterized in that: The data alignment module synchronizes the time of the vehicle-end perception data output by the vehicle-end perception data acquisition module and the roadside perception data output by the roadside perception data acquisition module according to the time stamp issued by the vehicle end, and performs spatial coordinate conversion alignment using the 4×4 pose transformation matrix. 7.The intelligent network connection oriented bus non-blind area perception system according to claim 3, characterized in that: The blind area identification module determines the "point-in-view" of the aligned roadside targets based on the field of view model of each sensor on the vehicle-end coordinate system, marks the targets not within the field of view of any sensor as blind area targets, and generates a blind area target list. 8.The intelligent network connection oriented bus blind area free perception system according to claim 7, characterized in that: The blind area target trajectory prediction module constructs a state sequence for the historical position, speed and acceleration of each target in the blind area target list, inputs the trained LSTM network to predict the continuous trajectory of the target in the next 2-3 seconds, and generates a blind area target prediction trajectory. 9.The intelligent network connection oriented bus blind area free perception system according to claim 8, characterized in that: The collision risk assessment module calculates the minimum distance between the blind area target prediction trajectory and the bus planning trajectory at each time, combines the route attention weight output by the trajectory interaction and route attention weighting module, and calculates the risk index. The target with a risk index greater than a preset threshold is marked as a high-risk target, and the route attention weight is calculated according to the Euclidean distance and the preset attenuation function based on the bus route center line geometry data. The target with a risk index greater than a preset threshold is marked as a high-risk target.

10. The blind area-free perception system for intelligent networked public transportation according to claim 9, characterized in that: The high-risk target information feedback module feeds back the position information, speed, prediction trajectory and risk level of the high-risk blind area target to the vehicle end in real time through V2X communication. The vehicle-road information fusion module fuses its results with the vehicle-end self-detection results to generate a blind area-free global environmental perception result. The vehicle-road information fusion module fuses its results with the vehicle-end self-detection results to generate a blind area-free global environmental perception result.

Images