An unmanned full-automatic unmanned aerial vehicle observation system

Abstract

The application discloses an unattended full-automatic unmanned plane observation system and relates to the technical field of unmanned planes, which comprises the following parts: a multi-sensor group for collecting real-time multi-sensor data flow; a memory storing an observation blind area self-discovery filling algorithm; a processor executing the observation blind area self-discovery filling algorithm, which comprises a blind area detection module, a conditional generative adversarial network variant module, a Bayesian uncertainty quantification submodule and a reinforcement learning agent module; the blind area detection module extracts a blind area boundary mask through a convolution layer and quantifies geometric shapes; the conditional generative adversarial network variant module generates candidate filling data; the Bayesian uncertainty quantification submodule infers and calculates the lower bound of evidence of each generated pixel or point; and the reinforcement learning agent module decides to fuse the candidate filling data or triggers a reflight verification path. Through the multi-stage hybrid generation framework of the observation blind area self-discovery filling algorithm, the self-discovery and filling of the blind area of the unmanned plane are realized.

Description

TECHNICAL FIELD

[0001] The application relates to the technical field of unmanned aerial vehicles, in particular to an unattended full-automatic unmanned aerial vehicle observation system. BACKGROUND

[0002] In the prior art, the unattended full-automatic unmanned aerial vehicle observation system is widely used in the fields of environmental monitoring, geological exploration and ecological tracking, and the system usually relies on a multi-sensor group to collect real-time data streams, such as camera images, LiDAR point clouds and sound wave sensor reflections, to realize autonomous observation. However, in a long-time observation process, the dynamic blind area problem has become a key bottleneck restricting the performance of the system. These dynamic blind areas are often caused by factors such as fog obstruction, vegetation obstruction or topographic depression, which result in the inability of the sensor to directly obtain complete data. Traditional methods mainly adopt passive avoidance strategies, such as path detouring or post-data interpolation, but these methods have significant shortcomings: on the one hand, path detouring increases energy consumption and flight time, reducing observation efficiency; on the other hand, simple interpolation ignores the multi-modal weak signals at the edge of the blind area, cannot accurately reconstruct complex data distribution, and is easy to introduce errors and illusions, resulting in incomplete overall observation data, and further affecting the reliability and decision accuracy of subsequent analysis. In addition, the existing system lacks quantitative evaluation and adaptive decision-making mechanism for blind area uncertainty, cannot realize end-to-end real-time processing, and further limits the autonomy and robustness of the system in complex environments. SUMMARY

[0003] The technical problem to be solved by the present application is to provide an unattended full-automatic unmanned aerial vehicle observation system to solve the problems mentioned in the background.

[0004] In order to achieve the above-mentioned purpose, the technical scheme adopted by the present application is as follows:

[0005] An unattended full-automatic unmanned aerial vehicle observation system, comprising a unmanned aerial vehicle body, a multi-sensor group, a processor and a memory.

[0006] The multi-sensor group is used to collect real-time multi-sensor data streams.

[0007] The memory stores an observation blind area self-discovery filling algorithm, and the processor executes the observation blind area self-discovery filling algorithm to process the real-time multi-sensor data streams, realizes blind area self-discovery and filling.

[0008] The observation blind area self-discovery filling algorithm is a multi-stage hybrid generation framework, comprising a blind area detection module, a conditional generative adversarial network variant module, a Bayesian uncertainty quantification submodule and a reinforcement learning agent module.

[0009] The blind area detection module extracts a blind area boundary mask and quantifies a geometric shape through a convolution layer; the conditional generative adversarial network variant module incorporates a self-attention mechanism, extracts a multi-modal feature vector from a blind area edge, and inputs the feature vector as a condition to generate candidate filling data; the Bayesian uncertainty quantification submodule calculates a lower bound of evidence for each generated pixel or point based on variational inference, and dynamically adjusts a confidence threshold;

[0010] The reinforcement learning agent module evaluates the consistency reward of the filled data based on a deep Q network, decides to fuse the candidate filling data or trigger a reflight verification path, and realizes an end-to-end self-discovery, generation, and verification cycle.

[0011] More specifically, the multi-sensor group includes a camera, a LiDAR sensor, and a sound wave sensor for collecting real-time multi-sensor data streams, including camera images, LiDAR point clouds, and sound wave sensor reflections.

[0012] More specifically, the blind area detection module analyzes the real-time multi-sensor data streams, extracts a blind area boundary mask from the camera images through the convolution layer, extracts a point cloud boundary mask from the LiDAR point clouds, extracts a reflection boundary mask from the sound wave sensor reflections, and fuses the blind area boundary mask, the point cloud boundary mask, and the reflection boundary mask to quantify the geometric shape, which includes a convex hull boundary.

[0013] More specifically, the multi-modal feature vector includes a Sobel gradient of an image edge, a point density gradient of a LiDAR, and a frequency spectrum feature of a sound wave.

[0014] More specifically, the process of processing the multi-modal feature vector includes calculating an attention weight of the Sobel gradient of the image edge and the point density gradient of the LiDAR, and applying the attention weight to generate pixel values or point cloud values in the candidate filling data.

[0015] More specifically, the dynamic adjustment of the confidence threshold is applied to a historical filling accuracy rate through gradient descent, and the initial value of the confidence threshold is 0.75.

[0016] More specifically, the historical filling accuracy rate is obtained from a comparison of the candidate filling data and actual verification data previously fused, and the comparison includes calculating a mean square error at a pixel level or a point cloud level.

[0017] More specifically, the trigger reflight verification path includes generating a backup flight path to bypass the blind area or reapproach the blind area edge to collect additional data.

[0018] More specifically, the blind area detection module is based on a U-Net architecture.

[0019] More specifically, the observation blind area self-discovery filling algorithm application includes sensor blind spots caused by fog obstruction, vegetation obstruction or topographic depression.

[0020] The advantage of the present application over the prior art is that the present application realizes blind area self-discovery and filling through a multi-stage hybrid generation framework of observation blind area self-discovery filling algorithm. The framework includes a blind area detection module, a conditional generative adversarial network variant module, a Bayesian uncertainty quantification submodule and a reinforcement learning agent module, which can extract multi-modal feature vectors from the edge of the blind area to generate candidate filling data, dynamically adjust the confidence threshold through evidence lower bound calculation and consistency reward evaluation, and decide to fuse the filling data or trigger the reflight verification path, thereby ensuring the end-to-end self-discovery, generation and verification cycle. The present application system improves the integrity and accuracy of observation data, reduces data missing and analysis bias caused by blind area; reduces the frequency of reflight verification, optimizes energy consumption and flight efficiency; enhances the robustness and autonomous adaptation ability of the system in dynamic environment, and is suitable for long-time observation process, such as fog obstruction or topographic depression scene, to realize data integrity recovery. BRIEF DESCRIPTION OF DRAWINGS

[0021] Figure 1 is a flowchart of the present application;

[0022] Figure 2 is a system structure diagram of the present application;

[0023] Figure 3 is a blind area detection module flowchart of the present application;

[0024] Figure 4 is a filling data generation flowchart of the present application. DETAILED DESCRIPTION

[0025] The specific embodiments of the present application will be described below in conjunction with the accompanying drawings.

[0026] As shown in Figure 1 and Figure 2 , an unmanned full-automatic unmanned aerial vehicle observation system of the present application is composed of an unmanned aerial vehicle body, a multi-sensor group, a processor and a memory.

[0027] The multi-sensor group specifically includes a camera, a LiDAR sensor and a sound wave sensor, which work cooperatively to collect real-time multi-sensor data stream;

[0028] The camera is responsible for acquiring image data in the visible light range;

[0029] The LiDAR sensor generates three-dimensional point cloud data by measuring distance with laser pulses, and the sound wave sensor emits sound waves and receives reflected signals to detect object surface or medium changes.

[0030] This data stream is generated in real-time during observation, for example, in a foggy environment, the camera image may show a blurred area, the LiDAR point cloud may have a sparse point area, and the sound wave sensor reflection may have a signal attenuation interruption, thereby forming a dynamic blind area.

[0031] The observation blind area self-discovery filling algorithm is pre-stored in the memory, and the processor processes the collected real-time multi-sensor data stream by executing the algorithm, thereby realizing automatic discovery and data filling of the blind area.

[0032] The algorithm of the application adopts a multi-stage hybrid generation framework design, organically integrates multiple modules, and ensures full-process automation from blind area detection to data verification.

[0033] Specifically, the framework of the application includes a blind area detection module, a conditional generative adversarial network variant module, a Bayesian uncertainty quantification submodule, and a reinforcement learning agent module. These modules cooperate in sequence to first detect the blind area position, then generate the filling content, quantify the uncertainty, and finally decide whether to adopt the filling result, forming a closed-loop processing mechanism.

[0034] As shown in Figure 3 The blind area detection module is constructed based on the U-Net architecture, which is a common semantic segmentation network composed of an encoder and a decoder. The encoder gradually extracts features through multiple convolutional layers, and the decoder restores the resolution through upsampling and fuses low-level and high-level features in the skip connection to accurately locate the blind area boundary.

[0035] In actual implementation, this module analyzes the real-time multi-sensor data stream, extracts the blind area boundary mask from the camera image using the convolutional layer, such as identifying the area of pixel value mutation in the image through an edge detection operator; extracts the point cloud boundary mask from the LiDAR point cloud, such as calculating the threshold area of point density drop; extracts the reflection boundary mask from the sound wave sensor reflection, such as analyzing the starting point of waveform amplitude attenuation. Subsequently, these masks are fused, integrated into a unified boundary representation through weighted averaging or logical operation, and the geometric shape is quantified, wherein the geometric shape specifically includes a convex hull boundary, which is obtained by calculating the minimum convex polygon of the mask point set, used to describe the size and shape of the outer contour of the blind area. For example, in a river monitoring scene, if the fog covers the middle section of the water surface, the module will extract the fog edge mask from the image, the sparse point mask of the water bank from the point cloud, and the reflection interruption mask from the sound wave, and after fusion, an elliptical convex hull boundary with an area of about 100 square meters is generated, helping the subsequent module to focus on processing the range.

[0036] As shown in Figure 4As shown, the conditional generative adversarial network variant module is responsible for generating candidate filling data in the framework, which is an improved form of the standard generative adversarial network, containing a generator and a discriminator, where the generator learns to generate data from noise or conditional input, and the discriminator evaluates authenticity.

[0037] To adapt to multi-modal data, the module incorporates a self-attention mechanism, which allows the model to dynamically allocate attention weights when processing features, emphasizing relevant parts and ignoring noise. In implementation, the self-attention mechanism processes multi-modal feature vectors extracted from the edge of the blind area, including Sobel gradients of image edges, i.e. pixel gradient vectors calculated by Sobel operators, for capturing texture changes; point density gradients of LiDAR, i.e. local density differences in point clouds, for representing three-dimensional structure transitions; spectral features of sound waves, i.e. frequency components obtained by Fourier transform, for reflecting medium properties. The processing process first calculates the attention weights of Sobel gradients of image edges and point density gradients of LiDAR, for example, normalizing the dot product results using the softmax function, and then applies these weights to generate pixel values or point cloud values in candidate filling data.

[0038] Specifically, the generator takes these weighted vectors as conditional input and outputs reconstructed blind area data, such as filling image pixels or supplementary point cloud points. During training, this module uses joint optimization of adversarial loss and reconstruction loss, and the generator tries to deceive the discriminator while minimizing the L1 distance from the real data. The training data comes from a labeled observation dataset, such as synthetic samples simulating fog and blind areas, and the number of iterations is usually set to 10,000, with a learning rate starting from 0.001 and gradually decaying. Through this design, the module can generate filling content that is more consistent with the context, such as generating ground point clouds hidden under leaves in forest observations with vegetation obstruction after extracting features from edge gradients, avoiding distortion of simple interpolation.

[0039] The Bayesian uncertainty quantification sub-module runs in parallel with the conditional generative adversarial network variant module, and this parallelism is achieved through multi-threading or GPU acceleration, ensuring that it does not affect the generation speed. This sub-module calculates the lower bound of evidence for each generated pixel or point based on the variational inference method, which is a technique for approximating Bayesian posteriors by introducing a variational distribution q to minimize the KL divergence from the true posterior, thereby estimating uncertainty.

[0040] In the calculation, samples are generated from the candidate filling data to form pixels or points, for example, 100 samples are randomly selected, and then the evidence lower bound value is calculated for each sample, which is calculated based on the expectation under the variational distribution, log-likelihood and KL divergence. The specific formula is that the evidence lower bound is equal to the log-likelihood under the variational expectation minus the KL divergence of the variational distribution and the prior distribution. This setting is because it provides a lower bound approximation, which can efficiently quantify the model's confidence in generating data without precise integration of high-dimensional posterior.

[0041] The higher the evidence lower bound value, the lower the uncertainty, and the value range can be set between -10 and 0, with negative values representing approximation errors. After quantifying the uncertainty, the sub-module dynamically adjusts the confidence threshold by applying gradient descent to the historical filling accuracy, where the initial value of the confidence threshold is 0.75, and the value range is 0.5 to 0.95, which is adjusted according to the actual environment to balance accuracy and efficiency. The historical filling accuracy is obtained from the comparison of the previously fused candidate filling data and the actual verification data, for example, by calculating the pixel-level mean square error, if the error is less than 0.1, it is considered accurate, and the accuracy is accumulated. Gradient descent uses the Adam optimizer with a step size of 0.01 and iterates 5 to 10 times. This mechanism ensures that only reliable filling is used, for example, in mountainous areas where the terrain is concave, if the evidence lower bound of the generated point cloud is lower than -5, the sub-module will lower the confidence threshold to 0.6, preventing the introduction of high-uncertainty data.

[0042] The reinforcement learning agent module is based on a Deep Q-Network, which is a reinforcement learning model for value function approximation, composed of multiple fully connected networks, with the input state being the filled data and the output action being the Q value for decision-making. The architecture of the Deep Q-Network includes an input layer that receives consistency reward features, a hidden layer that uses a three-layer structure with 128 neurons, and an output layer that corresponds to actions such as fusion or reflight. The training process uses an experience replay buffer to store state transitions, and a target network is updated regularly to stabilize learning, with a discount factor set to 0.99 and an exploration rate decaying from 1 to 0.1. Training data comes from simulated observation episodes, with each episode simulating a blind area event, and the reward is designed as negative KL divergence plus positive confidence score.

[0043] This module receives the consistency reward of the filled data as input, which is based on the KL divergence between the distribution of the filled data and the surrounding data, the confidence weight factor and the confidence score. The KL divergence measures the distribution difference, with a value range of 0 to infinity, the smaller the better; the confidence weight factor is set to 0.4 to 0.6, adjusted according to importance; and the confidence score is directly obtained from the Bayesian sub-module.

[0044] After reward calculation, used to evaluate consistency, if the reward is higher than 0, the decision fusion candidate filling data, that is, the generated content is inserted into the real-time multi-sensor data stream to complete the blind area, this insertion is performed smoothly through linear interpolation, and only when the confidence exceeds a preset value such as 0.7. Otherwise, trigger the re-flight verification path, for example, generate a backup flight path to bypass the blind area or re-approach the edge to collect additional data, the path is integrated in the path planning module through the A* algorithm, seamlessly combined with the observation algorithm, the planning module uses a grid map with a resolution of 0.1 meters, considering the energy cost. This decision-making process realizes an end-to-end self-discovery, generation, and verification cycle, which is executed in real time on the processor, continuously inputting real-time multi-sensor data streams, outputting filled observation data, and feeding back to the system path control to adjust the flight attitude.

[0045] The observation blind area self-discovery filling algorithm is particularly suitable for long-time observation processes, such as ecological monitoring that lasts for several hours, in which dynamic blind areas such as image blur under fog obstruction, point cloud missing under vegetation obstruction, or sensor blind spots caused by topographic depressions are handled. Through the cooperation of the multi-stage hybrid generation framework, these blind areas are detected, filled, quantified uncertainty, and verified, ultimately achieving data integrity recovery. For example, in a long-time river pollution tracking task, the UAV flies for 8 hours and encounters a fog area blind area, the detection module quantifies the convex hull boundary as a rectangular region; the generation module generates a pollutant concentration map from the Sobel gradient and spectral features; the uncertainty submodule calculates the lower bound of evidence -1.5, adjusts the threshold to 0.8; the agent module calculates the reward 0.3, fuses the data to complete the distribution map, avoiding the omission of pollution peaks caused by missing data, and the entire cycle is executed every second to ensure data continuity.

[0046] In terms of system hardware configuration, the UAV body adopts a quadrotor structure, the processor uses NVIDIA Jetson series chips supporting parallel computing, and the memory capacity is not less than 64GB to accommodate algorithm models and data buffering. The sampling rate of the multi-sensor group is set to 30Hz to ensure data stream resolution, the camera resolution is 1920x1080, the LiDAR range is 500 meters with centimeter-level accuracy, and the sound wave sensor frequency is 50-200kHz.

[0047] In software implementation, the algorithm framework is developed with PyTorch framework, the U-Net model parameters are about 5 million, the training is performed on GPU with a batch size of 16, and the loss function is binary cross-entropy; the conditional generative adversarial network variant is trained with an adversarial loss weight of 0.5 and a reconstruction loss of 0.5, and the training set includes 10,000 pairs of real observations and GAN enhanced samples. The variational distribution of the Bayesian submodule adopts a Gaussian distribution with a mean of 0 and a variance of 1, and the prior is a standard normal; the Deep Q-network is trained for 5,000 episodes, with a buffer size of 20,000 and a target update every 100 steps. This detailed implementation ensures the robustness of the algorithm, for example, in simulation tests, the filling accuracy reaches 88%, which is 20% higher than the baseline interpolation.

[0048] Further refine the blind area fusion process, in the detection module, the mask fusion uses a pixel-level voting mechanism, if at least two of the three masks overlap, the boundary point is confirmed, and the Graham scan algorithm is used when quantizing the convex hull, with a time complexity of O(nlog n), n is the number of points, usually less than 1000.

[0049] In the self-attention calculation of the generation module, the weight matrix dimension is set to 64, the feature vector length, and the softmax temperature parameter is 1.0 to avoid gradient explosion. In the uncertainty calculation, the sampling number can be dynamically set to 10%-50% of the number of generated pixels, the evidence lower bound optimization uses ELBO maximization, and the gradient updates the variational parameters through backpropagation. The learning rate of the gradient descent learning rate is 0.005, and the stopping condition is that the accuracy changes less than 0.01. The input state of the reinforcement agent includes a reward vector dimension of 5, the action space is a binary selection, and the Q value update uses Huber loss to handle outliers.

[0050] In one specific embodiment, consider a fully autonomous unmanned aerial vehicle observation system deployed in a dense forest area for long-term monitoring of wild animal migration paths. The system is equipped with a UAV body, a multi-sensor group, a processor and a memory, wherein the multi-sensor group integrates a camera to collect high-resolution images, a LiDAR sensor to generate detailed point cloud mapping of forest structure, and a sound wave sensor to capture reflection waveforms caused by animal activity.

[0051] In a typical vegetation-shielded observation task, the UAV flies along the preset path of the forest path, and real-time multi-sensor data streams are collected: camera images show blurred ground traces in dense leaf areas, LiDAR point clouds have point blanking in the leaf layer, and sound wave reflections are interrupted due to vegetation absorption, forming a dynamic blind area of about 30 meters in diameter, which blocks potential animal paths.

[0052] The multi-stage hybrid generation framework of the observation blind zone self-discovery filling algorithm is activated immediately. First, the blind zone detection module is launched based on the U-Net architecture, which processes the data stream through convolutional layers: the blind zone boundary mask is extracted from the camera image to identify the edge pixel interruption area of the leaf; the point cloud boundary mask is extracted from the LiDAR point cloud to mark the boundary of the sharp decrease in branch point density; the reflection boundary mask is extracted from the sound wave sensor reflection to detect the starting segment of the waveform amplitude attenuation. These masks are fused to quantify the geometric shape, such as calculating the convex hull boundary, forming a polygon to enclose the blind zone, with a total edge length of about 100 meters, helping the framework to locate the focus of processing.

[0053] Subsequently, the conditional generative adversarial network variant module incorporates a self-attention mechanism to extract multi-modal feature vectors from the blind zone edge: the Sobel gradient of the image edge captures the gradual change of leaf veins, the point density gradient of the LiDAR analyzes the transition of branches and leaves, and the spectral features of the sound wave reflect the ground echo shift. The self-attention mechanism calculates the attention weights of the Sobel gradient and the point density gradient, such as weighting the structural correlation, and then applies these weights to generate candidate filling data, such as reconstructing animal footprint pixels within the blind zone or supplementing point cloud path points, to ensure that the generated content is coordinated with the surrounding forest data.

[0054] At the same time, the Bayesian uncertainty quantification submodule is based on variational inference and runs in parallel to sample pixels or points from the candidate filling data, such as selecting 200 samples, to calculate the lower bound of the evidence for each (which is a calculation belonging to the prior art), which is obtained by subtracting the log-likelihood and KL divergence from the expectation under the variational distribution, representing the reliability of generation. The initial confidence threshold is 0.75, which is dynamically adjusted through gradient descent based on historical filling accuracy, such as obtaining an accuracy of 0.85 from the pixel-level mean square error of 0.08 in the previous animal path filling, and adjusting the threshold to 0.77.

[0055] Finally, the reinforcement learning agent module evaluates the consistency reward of the filled data based on the deep Q network, with a reward value of about 0.55 calculated by the KL divergence 0.15 between the filling data distribution and the surrounding distribution, the confidence weight factor 0.5, and the confidence score 0.8. Based on this, the agent decides to fuse the candidate filling data, inserts the reconstructed path into the data stream to complete the blind zone, and triggers the re-flight verification path to save energy when the confidence exceeds the preset value of 0.7. The entire end-to-end self-discovery, generation, and verification cycle is completed on the processor, outputting complete migration path data for base station analysis of animal behavior patterns.

[0056] With this process, the system successfully reconstructs the migration traces under the vegetation blind zone, avoids path breaks caused by incomplete data, and ultimately tracks the hidden migration route of a group of deer, optimizing forest protection measures.

[0057] The above merely describes preferred specific embodiments of the present application, but the protection scope of the present application is not limited thereto, and any person skilled in the art, according to the technical solution and inventive concept of the present application, makes equivalent replacement or change within the technical range disclosed by the present application, should be covered within the protection scope of the present application.

Claims

1. An unmanned full-automatic unmanned aerial vehicle observation system, characterized in that, The unmanned aerial vehicle comprises a body, a multi-sensor group, a processor and a memory. The multi-sensor group is used to collect a real-time multi-sensor data stream. The memory stores an observation blind area self-discovery filling algorithm, and the processor executes the observation blind area self-discovery filling algorithm to process the real-time multi-sensor data stream, realize blind area self-discovery and filling. The observation blind area self-discovery filling algorithm is a multi-stage hybrid generation framework, comprising a blind area detection module, a conditional generative adversarial network variant module, a Bayesian uncertainty quantification submodule and a reinforcement learning agent module. The blind area detection module extracts a blind area boundary mask through a convolution layer and quantifies a geometric shape. The conditional generative adversarial network variant module incorporates a self-attention mechanism, and a multi-modal feature vector extracted from a blind area edge is input as a condition to generate candidate filling data. The Bayesian uncertainty quantification submodule calculates a lower bound of evidence of each generated pixel or point based on variational inference, and dynamically adjusts a confidence threshold. The reinforcement learning agent module evaluates the consistency reward of the filled data based on a deep Q network, decides to fuse the candidate filling data or trigger a reflight verification path, realizes an end-to-end self-discovery, generation and verification cycle. The multi-sensor group comprises a camera, a LiDAR sensor and a sound wave sensor, and is used to collect a real-time multi-sensor data stream, including camera images, LiDAR point clouds and sound wave sensor reflections.

2. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The multi-modal feature vector comprises a Sobel gradient of an image edge, a point density gradient of a LiDAR and a frequency spectrum feature of a sound wave.

3. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The blind area detection module analyzes the real-time multi-sensor data stream, extracts a blind area boundary mask from the camera images through the convolution layer, extracts a point cloud boundary mask from the LiDAR point clouds, extracts a reflection boundary mask from the sound wave sensor reflections, and fuses the blind area boundary mask, the point cloud boundary mask and the reflection boundary mask to quantify the geometric shape, which comprises a convex hull boundary.

4. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The process of processing the multi-modal feature vector comprises calculating an attention weight of the Sobel gradient of the image edge and the point density gradient of the LiDAR, and applying the attention weight to generate pixel values or point cloud values in the candidate filling data.

5. The unattended fully automatic unmanned vehicle observation system according to claim 4, wherein, The dynamic adjustment of the confidence threshold is applied to a historical filling accuracy rate through gradient descent, and the initial value of the confidence threshold is 0.

75.

6. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The historical filling accuracy rate is obtained from a comparison of the candidate filling data previously fused with actual verification data, and the comparison comprises calculating a mean square error at a pixel level or a point cloud level.

7. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The trigger reflight verification path comprises generating a backup flight path to bypass the blind area or reapproach the blind area edge to collect additional data.

8. The unattended fully automatic unmanned vehicle observation system according to claim 1, wherein, The blind area detection module is based on a U-Net architecture. The observation blind area self-discovery filling algorithm is applied to sensor blind spots caused by fog obstruction, vegetation obstruction or topographic depressions.

Images