Efficient and low-bandwidth unmanned cooperative sensing method

Abstract

In development of a 6G-V2V collaborative awareness-oriented advanced auxiliary driving system, how to balance limited network bandwidth and improve detection accuracy and real-time performance of automatic driving is a main challenge of unmanned driving. According to the invention, high-speed and low-delay switching is carried out among multi-source data streams by using an optical wireless communication technology, the bandwidth of network transmission is improved, the flexibility and reliability of unmanned communication are ensured, a lightweight network based on intermediate-level feature fusion is developed in combination with a deep learning algorithm, and the system has a wide application prospect. Feature extraction coding and a dynamic fusion model are utilized to aggregate 3D point cloud features, a dynamic sensing area selection algorithm is researched, and the balance between bandwidth energy consumption and sensing coverage of a cooperative sensing system of multiple unmanned vehicles is optimized. The invention provides a practical and efficient method for realizing the collaborative awareness advanced auxiliary driving system with low energy consumption and high precision.

Description

Technical Field

[0001] The present invention belongs to the technical field of driverless vehicles, and particularly relates to an efficient and low-bandwidth driverless cooperative perception method. Background Art

[0002] As one of the main sensors for environmental perception of autonomous vehicles, lidar is of great significance for precise 3D target detection in the point cloud it generates. However, in the actual scenario of point cloud target detection, target detection algorithms are usually applied to complex background information. When using point cloud for three-dimensional target detection, the detection accuracy of small targets and occluded targets under complex backgrounds is a prominent problem. Therefore, it is very important to infer and evolve the reliability of vehicle-to-vehicle cooperative perception algorithms and intelligently fuse different types of sensor data (lidar, images, maps) in 3D object recognition. In addition, due to the blind spot recognition accuracy caused by occlusion and complex traffic environments, in driving scenarios such as limited spatial resolution, bad weather conditions, and busy intersections, the transmission of a large amount of data will cause network congestion, making it difficult to improve the robustness and real-time performance of autonomous driving. To address the above problems, the present invention proposes an efficient and low-bandwidth driverless cooperative perception method. Summary of the Invention

[0003] The present invention aims to solve problems such as the influence of redundant sensing data and overloaded total data volume on perception accuracy during the data fusion stage of single-vehicle and multi-vehicle driverless, and proposes an efficient and low-bandwidth driverless cooperative perception method, thereby effectively solving the problems of low single-vehicle recognition accuracy and overloaded multi-vehicle cooperative perception data volume.

[0004] To achieve the above object, the present invention adopts the following technical solutions.

[0005] A cooperative perception system for driverless vehicle groups with high communication bandwidth in complex real-world scenarios, comprising an optical wireless communication module, a programmable transceiver, a multi-source data fusion module, and a weight learning module;

[0006] The optical wireless communication module and the programmable transceiver are deployed in an edge computing system and belong to optical hardware devices. The multi-source data fusion module and the weight learning module are both neural networks. The multi-source data fusion module and the weight learning module are interconnected. The inputs of the optical wireless communication module and the programmable transceiver are ambient light of different wavelengths around the vehicle, which is converted into an electrical signal as the output through the transceiver. The inputs of the multi-source data fusion module and the weight learning module are respectively the sensing data of the driverless scenario, and the outputs are the corresponding perception decisions and lightweight fusion features.

[0007] The optical wireless communication module is used to achieve seamless connection with 6G infrastructure;

[0008] The programmable transceiver is used for delay signal processing to establish communication and additionally supports free space transmission within 100 meters;

[0009] The edge computing system is used to move the data processing and analysis capabilities from the centralized data center or cloud to the edge devices or network nodes close to the data source or users;

[0010] The multi-source data fusion module is used to fuse the features of the trained collaborative perception data in the unmanned driving scenario and obtain the most representative fused features, and broadcast them to each unmanned vehicle in the collaborative perception system through the optical wireless communication module to make corresponding driving decisions.

[0011] The weight learning module is used to adaptively screen the learned multi-source data weights, obtain the weighted dynamic feature fusion weight parameters, and update them to the multi-source data fusion module;

[0012] The collaborative perception data features include static features, dynamic features, constraint features, original point cloud features and intermediate-level features generated during the driving process of the unmanned vehicle. The fusion weight parameters include static weights, constraint weights, and dynamic weights. The collaborative perception driving decisions include target position and vehicle speed.

[0013] Preferably, the optical wireless communication module includes an optical emission module, an optical reception module, a beamformer, a signal processing module, a control module, a communication interface and a power management module;

[0014] The optical emission module uses a high-efficiency VCSEL light source and a high-speed electro-optic modulator;

[0015] The optical reception module uses a silicon photomultiplier (SiPM) and a low-noise transimpedance amplifier (TIA);

[0016] The beamformer is based on a photonic integrated circuit (PIC) to achieve two-dimensional electronically controlled beam deflection;

[0017] The signal processing module supports high-efficiency forward error correction coding (FEC) and QAM modulation to ensure high data rates;

[0018] The control module integrates AI algorithms for beam dynamic tracking and network resource optimization configuration;

[0019] Furthermore, the emission module includes a light source and a modulator, the optical reception module includes a detector and a signal amplifier, the beamformer includes a fan beam and a pencil beam, and the signal processing module includes an encoder, a decoder and a filter;

[0020] The control module is used for beam tracking and dynamic resource allocation;

[0021] The communication module is used to provide a communication interface with other modules or systems;

[0022] The power management module is used to supply power to the optical wireless communication module and optimize energy consumption.

[0023] The transmitting module, optical receiving module, beamformer, signal processing module, control module, communication interface, and power management module cooperate to complete data transmission, reception, processing, and interaction through optical fibers, free space, and standard interfaces.

[0024] Preferably, the programmable transceiver includes an optical transmitting module and an optical receiving module;

[0025] The optical transmitting module is used to convert the input electrical signal (modulation signal) into an optical signal and transmit it into free space or an optical fiber at an appropriate wavelength;

[0026] The optical receiving module is used to capture the transmitted optical signal and demodulate it into an electrical signal;

[0027] Furthermore, the optical transmitting module includes a tunable laser, an electro-optic modulator, and a power controller, and the optical receiving module includes an automatic gain control module and a noise suppression circuit;

[0028] The tunable laser, electro-optic modulator, and power controller are respectively connected to the optical wireless communication module;

[0029] The tunable laser is used for dynamic adjustment of multiple wavelengths;

[0030] The electro-optic modulator is used for amplitude, frequency, or phase modulation of the optical signal;

[0031] The power controller is used for dynamically adjusting the transmitted optical power to adapt to different communication distances;

[0032] The automatic gain control module is used for dynamically adjusting the amplification factor according to the received signal strength;

[0033] The noise suppression circuit is used to filter out background light interference and improve the receiving sensitivity.

[0034] Preferably, the multi-source data fusion module includes a data acquisition interface, a data synchronization module, and a multi-modal deep learning network;

[0035] The data acquisition interface is connected to the data synchronization module and the multi-modal deep learning network; high-level semantic understanding and environmental modeling are performed on the fused data using deep learning technology;

[0036] The data acquisition interface is used to receive raw data from various sensors (such as lidar, millimeter-wave radar, cameras, IMUs, GPS, etc.) and external systems (such as V2V networks);

[0037] The high-speed sensor interface, wireless communication interface, and data preprocessing module serve as the sensing data receiving ports. The high-speed sensor interface, wireless communication interface, and data preprocessing module are connected to the data synchronization module and the multi-modal deep learning network;

[0038] Furthermore, the data synchronization module includes a time synchronization unit and a spatial calibration unit;

[0039] The time synchronization unit and the spatial calibration unit are connected to the deep learning network, and the time synchronization unit and the spatial calibration unit are connected to the data acquisition interface;

[0040] The time synchronization unit is used for the alignment problem of multi-source data in terms of time;

[0041] The spatial calibration unit is used to ensure that different sensor data can be unified into the same spatio-temporal framework;

[0042] Furthermore, the multi-modal deep learning network includes a modality-specific feature extraction network, a cross-modal alignment and fusion network, and an output and feedback network;

[0043] The modality-specific feature extraction network, the cross-modal alignment and fusion network, and the output and feedback network are respectively multi-layer fully connected neural networks, and the modality-specific feature extraction network, the cross-modal alignment and fusion network, and the output and feedback network are respectively connected to the multi-modal deep learning network;

[0044] The modality-specific feature extraction network is used to independently extract low-level and middle-level features from different data modalities to enhance the global context awareness ability;

[0045] The cross-modal alignment and fusion network is used to align, correlate, and fuse features of different modalities to construct a unified feature representation;

[0046] The output and feedback network is used to transfer the inference results of multi-modal deep learning and make decisions, and improve the model performance through a feedback loop.

[0047] Preferably, the weight learning module includes a static feature weight learning unit, a soft constraint weight learning unit, a dynamic feature weight learning unit, and a weight fusion module;

[0048] The static feature weight learning unit, the soft constraint weight learning unit, the dynamic feature weight learning unit, and the weight fusion module are respectively connected to the modality-specific feature extraction network, and the weight fusion module is connected to the cross-modal alignment and fusion network;

[0049] The static feature weight learning unit is used to learn the features from static data such as road layout, fixed obstacles, traffic signs, etc., and generate static weights for long-term path planning;

[0050] The soft constraint weight learning unit is used to learn soft constraint information such as time constraints, traffic signals, and priority preferences, and generate flexibly adjusted soft constraint weights.

[0051] The dynamic feature weight learning unit is used to analyze the real-time data of sensors and cameras, generate dynamic weights, provide a basis for rapid adjustment for short-term decision-making, and at the same time input the weight parameters into the decision-making module to update the weight parameters of the meta-network;

[0052] The weight fusion module is used to comprehensively process the weights from static features, soft constraint features, and dynamic features to generate the final weights;

[0053] The weight learning module forms a global weight model adapted to the complex scenarios of driverless through the separate learning and comprehensive processing of static features, soft constraint features, and dynamic features. The dynamics and flexibility of the weights ensure the decision-making efficiency and accuracy of the system in a changing environment.

[0054] An efficient and low-bandwidth driverless cooperative perception method, including any one of the foregoing efficient and low-bandwidth driverless cooperative perception systems, and the specific steps are as follows:

[0055] Integrate the optical communication module into the edge computing of driverless vehicle equipment to extend the distance for the communication programmable transceiver to receive and transmit optoelectronic information;

[0056] Input the feature of different periods of the driverless scenario into the trained deep learning module, and the weight learning module learns to obtain the corresponding dynamic weight parameters;

[0057] Distinguish the data fusion methods of isomorphic and homologous driverless vehicles and heterogeneous and heterologous driverless vehicles, and update their respective fusion weights into the multi-modal deep learning network.

[0058] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0059] The method of the present invention facilitates system modularization, realizes independent perception and decision-making of a single vehicle, and further reduces the dependence on real-time high-bandwidth communication. Compared with a single vehicle using limited sensor devices (lidar, radar, and camera) to perceive the surrounding environment, the perception network formed by vehicle-to-vehicle cooperative perception vehicles added to the perception system has excellent performance. By aggregating and receiving sensing detection information from multiple independent individual vehicles to the central vehicle, the blind spot recognition accuracy caused by occlusion and complex traffic environments can be compensated, and the robustness and real-time performance of autonomous driving can be improved in limited spatial resolution, adverse weather conditions, and busy intersections. For the fusion process between heterogeneous sensor-configured driverless vehicles, through this method, new sensors can be seamlessly integrated with the existing perception system without a large amount of labeled data, solving the adaptation problem between different sensors and also improving the recognition accuracy. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] Figure 1 FIG. is a schematic diagram of the overall structural framework of an efficient and low-bandwidth driverless cooperative perception method;

[0061] Figure 2 FIG. is a schematic diagram of the structural framework of an optical wireless communication module;

[0062] Figure 3 FIG. is a schematic diagram of the structural framework of a multi-source data fusion module;

[0063] Figure 4 FIG. is a schematic diagram of the structural framework of a weight learning module;

[0064] Figure 5 FIG. is a schematic diagram of the flow of an efficient and low-bandwidth driverless cooperative perception method. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0065] In order to make the objectives, technical solutions, and advantages of the present invention clearer and more understandable, the following further details the present invention in conjunction with embodiments. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention. An efficient and low-bandwidth driverless cooperative perception method in this embodiment is carried out according to the following steps:

[0066] Step 1: Integrate the optical communication module and the edge computing system

[0067] First, integrate the optical wireless communication module and the programmable transceiver into the edge computing device of the driverless vehicle. The optical wireless communication module is responsible for achieving seamless connection between the vehicle and the 6G infrastructure, and completing signal exchange between the vehicle and other vehicles or infrastructure through high-speed optical emission and reception modules. The optical wireless communication module uses an efficient VCSEL light source and a high-speed electro-optic modulator to ensure the efficiency and low latency of signal transmission. The programmable transceiver is used for signal delay processing, supports free space transmission within 100 meters, and ensures fast and stable information transmission between the vehicle and the environment.

[0068] The relationship between the power and bandwidth requirements for optical signal transmission is as follows:

[0069]

[0070] Among them, P represents the signal power, E is the total energy transmitted, and t is the signal transmission time.

[0071] In the optical wireless communication system, the effective transmission rate of the system is related to the signal modulation method. In particular, the bandwidth efficiency of QAM (Quadrature Amplitude Modulation) can be expressed by the following formula (2):

[0072] Bandwidth efficiency of QAM modulation:

[0073]

[0074] Among them, M is the modulation order of QAM, and T s is the transmission time of each symbol.

[0075] Step 2: Multi-source data acquisition and data synchronization

[0076] The multi-source data fusion module receives the raw data from multiple sensors through the data acquisition interface and synchronizes it through the data synchronization module (including the time synchronization unit and the space calibration unit). The time synchronization unit is used to ensure that the data from different sensors can be accurately aligned, and the space calibration unit unifies the data from different sensors into the same spatio-temporal framework. This process ensures the coordination and fusion of different modality data and provides an accurate basis for subsequent data processing. The processing formula for time synchronization is:

[0077]

[0078] Among them, t1 and t2 are the time points when different sensors collect data, and t sync is the synchronized time.

[0079] Spatial calibration reduces errors by ensuring that different sensor data is unified into the same coordinate system. Assuming there is a translation and rotation transformation between the coordinate systems of two sensors, spatial calibration can be performed through the following formula:

[0080] T cal = R·P + T (4)

[0081] Where P is the data in the original coordinate system of the sensor, T cal is the calibrated data, R is the calibration coefficient, and T is the data before calibration.

[0082] Step Three: Deep Learning Network and Feature Extraction

[0083] In the deep learning network, the feature extraction part is responsible for extracting features from sensor data of different modalities. The modality-specific feature extraction network extracts low-level and mid-level features from data of different sensors, and feature extraction can be performed through a convolutional neural network (CNN). For an input data X, the calculation of its output feature F can be expressed as:

[0084] F = σ(W * X + b) (5)

[0085] Where σ is the activation function, * represents the convolution operation, and b is the bias term.

[0086] Step Four: In an autonomous driving system, different sensors and scenario factors have different importance. How to intelligently adjust the weights of each feature according to the current environment is the key to improving perception accuracy and decision-making performance. The goal of the weight learning module is to automatically learn the importance of each feature in the current driverless scenario through training, and adaptively adjust the decision-making process according to the learned weight information.

[0087] Static features such as road layout, traffic signs, buildings, etc. usually do not change with time, so fixed weights can be learned through long-term historical data. These static feature weights will help with path planning and long-term decision-making. By using unsupervised learning or rule-based algorithms, we can learn static features and optimize their weights using the gradient descent method.

[0088] Equation 6 is the basis for realizing static weight learning. The error is propagated back from the final decision to the static feature weights through backpropagation, thereby continuously updating the weights to ensure their effectiveness in various scenarios.

[0089]

[0090] Where L(W t ) is the loss function, is the gradient of the loss function with respect to the weights, and η is the learning rate.

[0091] Dynamic feature weight learning pays more attention to the processing of real-time data. Considering that the data quality and characteristics obtained by sensors vary in different environments (such as different weather conditions, road conditions, etc.), the weight learning module will adjust the weights of each feature according to the real-time update of sensor data. This process generates more accurate dynamic weights by combining with real-time sensor data (such as lidar, millimeter-wave radar, etc.).

[0092] Soft constraint features include traffic signals, right-of-way priorities, traffic rules, etc. These pieces of information usually have a flexible impact on decision-making but do not directly restrict behavior. The soft constraint weight learning module will perform dynamic learning according to the changes in the traffic environment and vehicle state and adjust the weights of soft constraints. These adjustments can help the system make flexible decisions in complex traffic scenarios. For example, when passing through a busy intersection, the system will assign a greater weight to traffic signals, while on a relatively empty road section, the weight of dynamic path planning may be more important.

[0093] Finally, all the learned weights will be fed into the weight fusion module for weighted processing to generate the final fused feature weights. This process not only involves the weighted average of static weights, soft constraint weights, and dynamic weights but also includes using optimization algorithms to balance the influence of different features in decision-making. This weighted fusion process needs to be adjusted in real time according to the requirements of different scenarios to ensure that the autonomous driving system can cope with complex environmental changes. As shown in Equation 7, the dynamic feature weights will be continuously updated according to the changes in sensor data:

[0094] W dynamic =αW o +(1-α)·ΔW (7)

[0095] where α is the smoothing factor, ΔW is the current dynamic weight update amount, W o is the initial weight parameter, and W dynamic is the weight parameter after real-time update.

[0096] Step Five; The collaborative perception and decision-making module is the core part of the system. It ensures the efficient sharing of collaborative perception data among multiple autonomous vehicles and makes decisions based on the shared information. Through the collaborative perception among multiple vehicles, the entire fleet can make up for the blind spots of single-vehicle sensors, improving the perception accuracy and real-time response ability of the autonomous driving system. Based on the fused features and weights provided by the multi-source data fusion module, the collaborative perception module generates a global decision for each vehicle. These decisions include target position, vehicle speed, path planning, etc., and are adjusted in real time according to different traffic conditions, obstacle positions, road conditions, etc. For example, when passing through a complex intersection, the system will consider the speed, position, and traffic signals of all vehicles to generate a collaborative optimization decision. This decision-making not only considers the needs of individual vehicles but also comprehensively considers the operating efficiency of the entire fleet. Specifically, the collaboration among vehicles enables the entire fleet to optimize the driving route while avoiding collisions, reducing traffic congestion, and improving the overall road utilization rate. Based on the fused features and weights provided by the multi-source data fusion module, the collaborative perception module generates a global decision for each vehicle. These decisions include target position, vehicle speed, path planning, etc., and are adjusted in real time according to different traffic conditions, obstacle positions, road conditions, etc. For example, when passing through a complex intersection, the system will consider the speed, position, and traffic signals of all vehicles to generate a collaborative optimization decision.

[0097] This decision-making not only considers the needs of individual vehicles but also comprehensively considers the operating efficiency of the entire fleet. Specifically, the collaboration among vehicles enables the entire fleet to optimize the driving route while avoiding collisions, reducing traffic congestion, and improving the overall road utilization rate.

[0098] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any other changes, modifications, substitutions, combinations, and simplifications made without departing from the spirit and principle of the present invention shall be equivalent replacement methods and are all included in the protection scope of the present invention.

Claims

1. A 6G-V2V collaborative perception advanced assisted driving method, characterized in that: The method comprises the following steps: Step 1: Build an advanced assisted driving experimental system based on 6G-V2V collaborative perception, which consists of a chip-level optical beamformer, laser / radio detection, lidar / radar, PIC and ASIC tunable transmitters and receivers, optical switches, and SDN controllers: Use optical beamformers to introduce fan-shaped beams and pencil-shaped beams with two-dimensional electrically controlled deflection angles to provide the system with high-precision spatial resolution capabilities. The laser / radio detection module works in conjunction with the beamformer to achieve accurate perception of multi-modal targets. Among them, lidar and radar are responsible for short-range and medium- and long-range target detection and data acquisition. Through the tunable transmitters and receivers built with PIC (photonic integrated circuits) and ASIC (application-specific integrated circuits), the system can achieve high-frequency dynamic tuning between multiple bands and multiple channels. The optical switch is used for high-speed, low-latency switching between multi-source data streams to ensure the flexibility and reliability of communication. The SDN (software-defined network) controller globally coordinates and intelligently manages the entire system, and dynamically allocates resources to adapt to the efficient deployment of 6G-V2V collaborative perception technology in assisted driving scenarios. Step 2: Apply 6G-V2V collaborative perception and advanced assisted unmanned driving edge computing: First, deploy a distributed Beyond 5G Independent Networking (B5G-SA) distributed 6G network to achieve full coverage in two designated indoor and outdoor areas and connect three unmanned vehicles with the same configuration of LiDAR, Radar, IMU and other sensor equipment; then, use LiDAR to quickly scan the driving scene, aggregate the perception information through 100Gbps bandwidth network edge computing, and broadcast it back to the central vehicle in the collaborative driving system using optical fiber adjustable transmitters and receivers. After the central vehicle decoder decodes the information, small objects in the blind spot are fed back to the vehicle's display system in real time; finally, the integration of avalanche photodetectors (APDs) and low-noise transimpedance amplifiers (TIAs) ensures that the sensitivity of silicon-based large-area APDs operating at 800nm ​​wavelengths is better than that of InGaAs operating at 1300nm wavelengths, even under the influence of ambient sunlight (whose irradiation intensity decreases with wavelength). APD can increase the data transmission rate to 0.5Gb / s in V2V optical wireless communication (OWC) links in the short-range (15m) field; Step 3: Apply heterogeneous data fusion based on 6G-V2V collaborative perception and advanced assisted unmanned driving: In order to flexibly integrate the perception data of unmanned vehicles produced by different manufacturers in actual driving scenarios, first, for the multimodal data collected by heterogeneous sensors (such as lidar, millimeter wave radar, camera, etc.) deployed on different unmanned vehicles, a cross-modal feature alignment algorithm is used to map different types of data to a shared feature representation space; secondly, through a dynamic weight allocation mechanism, the fusion strategy is adjusted in real time according to the confidence of each modal data in the perception scene and environmental changes to ensure the accuracy and robustness of data fusion; finally, in a unified feature space, combined with a collaborative spatiotemporal attention mechanism, the efficiency of perception information sharing between unmanned vehicles is optimized, thereby achieving efficient collaborative perception; the above method can significantly improve the perception accuracy and safety in a multi-vehicle collaborative environment, and is suitable for intelligent decision-making and control in complex driving scenarios; Step 4: Based on the relationship between the accuracy of three-dimensional target recognition and network bandwidth in collaborative perception under complex environment background, a lightweight network system is established: when the computing power of vehicle terminals is insufficient, edge computing and cloud computing become important supplements, monitor bandwidth consumption in real time and adjust the fusion data mode of transmission, and dynamically adjust the fusion stage according to actual consumption to observe the detection performance of unmanned vehicles; through experiments, it is found that when the original point cloud data of sensor data is fused in the early stage, the occlusion problem and field of view limitation in single-vehicle perception can be effectively solved, and significant results have been achieved, but sharing the original sensor data requires a lot of communication resources, which in turn affects its wide deployment and stability in actual application scenarios; compared with the early stage Compared with early fusion, late fusion only needs to transmit the processed detection results, allowing each agent to independently process its own perception data and then fuse them. This method facilitates system modularization and realizes independent perception and decision-making of a single vehicle, thereby reducing dependence on real-time high-bandwidth communication. However, the current late collaborative fusion method is highly dependent on the sensor data of a single agent, which limits its wide applicability. Therefore, intermediate-level feature fusion is used to filter and aggregate the sensor feature maps from different unmanned vehicles in the intermediate feature space generated by the prediction model. In this way, representative information is effectively compressed in the intermediate features, which achieves a better balance between transmission efficiency and perception effect.

2. The advanced driver assistance method for 6G-V2V collaborative perception according to claim 1, characterized in that: The entire operation process is carried out in specific outdoor and indoor scenarios respectively. Through a receiver based on SiPM (silicon photon multiplier), the communication capability supporting a data rate of 100Mb / s is achieved within the range of an optical wireless communication (OWC) link of more than 100 meters; wherein, the communication requires medium-power LEDs to have sufficient electro-optical bandwidth, and a high-intensity extinction ratio is achieved by adopting a SPAD (single photon avalanche diode) layout design in SiPM; in addition, the method targets two types of unmanned driving collaboration scenarios, homogeneous and heterogeneous, and adopts a multimodal data fusion strategy combining early perception fusion, mid-term feature fusion, and late decision fusion to ensure perception efficiency and decision accuracy in complex driving environments, characterized in that: in the designed 6G-V2V communication, the relationship between bandwidth and transmission rate is: Where R is the data transmission rate; B is the channel bandwidth; SNR is the signal-to-noise ratio; The power budget in fiber optic communications is: P r =P t -L-M Among them, P r is the received power; P t is the transmission power; L is the link loss, including the fiber loss and device insertion loss; M is the system design margin; The gain and noise factors for APD sensitivity analysis are: Where R is the detector responsivity; P opt is the optical power; M is the APD gain; q is the electron charge; I d is the dark current; F(M) is the gain noise factor; B is the network bandwidth; k B is the Boltzmann constant; T is the absolute temperature; R f is the transimpedance amplifier feedback resistor.

3. The advanced driver assistance method for 6G-V2V collaborative perception according to claim 1, characterized in that: The entire experimental process uses the linear characteristics of 800nm ​​silicon-based large-area APDs to support advanced modulation schemes, which can increase the data transmission rate to 0.5Gb / s in V2V optical wireless communication (OWC) links in the short-range (15m) field. The silicon-based APD increases the gain by an order of magnitude (M≈[100,150]).

Images