Micro-nano unmanned aerial vehicle edge end data preprocessing system

Abstract

The invention relates to the field of edge calculation, and discloses a micro-nano unmanned aerial vehicle edge end data preprocessing system, which comprises a state space discretization mapping module used for collecting three-axis attitude data output by an inertial measurement unit, and quantitatively mapping the three-axis attitude data to a three-dimensional discrete coordinate system according to a preset angular resolution step length to generate an attitude grid index; the visual fingerprint extraction module is used for acquiring original image data and calculating a current global feature hash value; and the index arbitration logic control module is used for querying a preset background feature lookup table based on the attitude grid index, and generating a wake-up instruction to activate a rear-end depth reasoning unit when the Hamming distance between the current global feature hash value and the reference value is greater than a sudden change judgment threshold value. The coupling relation between the fuselage movement and the environment change is stripped, and low-power-consumption accurate awakening of the edge end under the dynamic non-stable working condition is achieved.

Description

Technical Field

[0001] This invention relates to a micro / nano-type unmanned aerial vehicle (UAV) edge data preprocessing system, belonging to the field of edge computing technology. Background Technology

[0002] Currently, micro-nano drones are widely used for inspection and fixed-point monitoring tasks in confined spaces. They are typically equipped with visual sensors and deep learning inference units to capture and identify environmental targets. To meet the requirements of long-term operation and endurance, the system design often adopts a hierarchical wake-up strategy. In standby mode, the high-power inference unit is turned off, while the low-power front-end perception module is retained to monitor environmental dynamics. When a pixel-level change is detected in the field of view, the back-end processor is woken up for recognition. Due to the extremely small physical size and rotational inertia, the drone generates high-frequency random attitude jitter when performing hovering or monitoring tasks. This is due to the non-stationary motion of the carrier itself. In the continuous image frames captured by the visual sensor, it appears as a displacement of background pixels across the entire screen. Using traditional inter-frame difference or background modeling algorithms, the front-end monitoring module has difficulty distinguishing between the background motion caused by the change in the drone's attitude and the local motion of the real target. Introducing electronic image stabilization or optical flow calculation image processing algorithms to suppress jitter interference requires continuous computing power consumption during the data preprocessing stage, which conflicts with the limited battery energy budget of the micro-nano platform.

[0003] Against this backdrop, although existing edge computing technologies have attempted to address detection accuracy issues by enhancing edge computing power, limitations still exist under extremely low power consumption constraints. For example, Chinese invention patent CN117274843B discloses a method and system for front-end defect recognition of drones based on lightweight edge computing. This scheme performs preprocessing such as Gaussian filtering and contrast enhancement on the acquired image data through an edge computing unit, and further utilizes the YOLOv4 model for defect feature extraction and comparison. Although such schemes can achieve high accuracy when image quality is good and computing power is sufficient, they still have limitations. While accurate identification is currently achieved, it still relies on real-time full image enhancement and deep learning inference of continuous video streams. For micro- and nano-sized drones with extremely tight energy budgets, the lack of a pre-emptive low-power filtering mechanism for carrier attitude disturbances not only fails to eliminate the false triggering of detection logic caused by full-frame pixel migration due to body vibration, but also results in the precious power of the edge processor being mainly consumed in processing invalid jitter data due to the continuous operation of high-load image processing algorithms. This cannot meet the stringent requirements of achieving milliwatt-level long-term operation of micro- and nano-sized platforms under dynamic and non-stationary conditions.

[0004] Therefore, the technical problem to be solved by this invention is to construct an image preprocessing mechanism that does not rely on high computing power and is triggered to wake up low-power data processing in response to changes in the real environment. Summary of the Invention

[0005] To address the problems mentioned in the background art, the technical solution of the present invention is as follows: A micro / nano-type UAV edge data preprocessing system, the system being configured on an airborne processing platform integrating an inertial measurement unit and an image sensor, the system comprising: The state space discretization mapping module is used to collect the three-axis attitude data output by the inertial measurement unit in real time, and quantize and map the three-axis attitude data to a three-dimensional discrete coordinate system according to the preset angular resolution step size to generate the attitude mesh index at the current moment. The visual fingerprint extraction module is used to acquire the raw image data output by the image sensor and calculate the current global feature hash value of the raw image data based on the perceptual hash algorithm. The index arbitration logic control module, connected to the state space discretization mapping module and the visual fingerprint extraction module, is used to access a preset background feature lookup table and execute the following logic control: When no record matching the attitude grid index is found in the background feature lookup table, the current global feature hash value is written as the baseline value to the storage address pointed to by the attitude grid index in the background feature lookup table, keeping the backend deep inference unit in a dormant state; when a baseline value matching the attitude grid index is found in the background feature lookup table, the Hamming distance between the current global feature hash value and the baseline value is calculated; only when the Hamming distance is greater than the preset mutation judgment threshold, a wake-up command is generated to activate the backend deep inference unit to perform target recognition operations on the original image data.

[0006] Preferably, the index arbitration logic control module further includes a topological neighborhood elastic retrieval unit, used to perform the following steps when querying the background feature lookup table: generating a candidate index set containing the attitude grid index and its spatially adjacent grid indexes based on preset topological adjacency rules; reading multiple benchmark values ​​corresponding to each index in the background feature lookup table and the candidate index set in parallel; calculating the Hamming distance between the current global feature hash value and each of the multiple benchmark values, and selecting the smallest Hamming distance as the final decision distance; comparing the final decision distance as the Hamming distance with the mutation decision threshold.

[0007] Preferably, the system further includes a dynamic validity gating module, which is used to monitor the three-axis angular velocity vector output by the inertial measurement unit in real time and calculate the magnitude of the three-axis angular velocity vector; when the magnitude is greater than a preset optical validity threshold, a processing lockout signal is generated; in response to the processing lockout signal, the index arbitration logic control module blocks the generation of the wake-up command and suspends the read and write operations of the background feature lookup table until the magnitude falls back below the optical validity threshold.

[0008] Preferably, the background feature lookup table also stores a suppression weight counter associated with each attitude grid index. The system also includes a negative feedback adaptive adjustment loop connected between the back-end deep inference unit and the index arbitration logic control module. This loop receives the inference results from the back-end deep inference unit. If the inference result indicates that the preset target was not identified, an incremental operation is performed on the suppression weight counter associated with the attitude grid index at the current moment. The index arbitration logic control module dynamically adjusts the mutation judgment threshold corresponding to the attitude grid index according to the current value of the suppression weight counter, based on the following formula: ,in, The adjusted mutation determination threshold. The initial threshold preset by the system. These are the preset positive weighting coefficients. To suppress the current value of the weight counter.

[0009] Preferably, the system further includes a semantically driven baseline reset module, which generates a baseline overwrite signal in response to the feedback from the backend deep inference unit to the wake-up command, characterized as an inference result without a target of interest; the index arbitration logic control module responds to the baseline overwrite signal by taking the current global feature hash value at the current moment as the new baseline value and overwriting it into the storage address pointed to by the attitude grid index in the background feature lookup table.

[0010] Preferably, the topological adjacency rules include directional neighborhood rules based on the angular velocity direction vector output by the inertial measurement unit, and 26-neighborhood omnidirectional rules defined in the three-dimensional mesh space; the topological neighborhood elastic retrieval unit is used to preferentially read the reference value corresponding to the adjacent mesh index on the side pointed to by the angular velocity direction vector.

[0011] Preferably, the angular resolution step size configured in the state space discretization mapping module is linearly related to the instantaneous field of view of the image sensor, which is used to ensure that the spatial coverage of the attitude grid index is less than or equal to the coverage of the instantaneous field of view; the background feature lookup table is stored in the on-chip static random access memory of the airborne processing platform.

[0012] Preferably, the visual fingerprint extraction module is used to downsample the original image data to a preset size, convert it into a grayscale image, calculate the discrete cosine transform coefficients of the grayscale image, and extract the low-frequency region coefficients to generate the current global feature hash value; the mutation determination threshold ranges from 10% to 20% of the total number of bits in the current global feature hash value.

[0013] Preferably, the dynamic validity gating module is also used to: send an update signal to the index arbitration logic control module in the first processing cycle after the lockout signal is released; the index arbitration logic control module responds to the update signal, skips the Hamming distance calculation step, and directly executes the operation of writing the current global feature hash value into the background feature lookup table.

[0014] Preferably, the negative feedback adaptive adjustment loop is also used to: generate a count reset signal when the inference result fed back by the back-end deep inference unit is characterized as the recognition of a preset target; the index arbitration logic control module responds to the count reset signal by clearing the suppression weight counter associated with the current attitude grid index and restoring the corresponding mutation judgment threshold to the initial threshold.

[0015] Compared with the prior art, the beneficial effects of the present invention are: 1. A discrete attitude grid coordinate system is constructed using inertial measurement data, and the global feature hash values ​​of the visual sensor are mapped to the coordinate system to establish a spatial state revisit logic triggering mechanism. The high-frequency attitude jitter of the UAV fixed-point monitoring mission is used as a passive environmental scanning source. The temporal difference operation of the continuous video stream is transformed into a hash feature comparison operation in the discrete attitude space. When the Hamming distance between the current attitude and the background fingerprint corresponding to the historical revisit attitude changes abruptly, the deep inference unit is activated, and the coupling relationship between the aircraft's own motion and the changes in the external environment is decoupled at the logical level. Based on the attitude index data processing method, the system does not need to perform high-computing-power electronic image stabilization or optical flow calculations, filters out global pixel displacement interference caused by aircraft vibration, reduces the frequency of invalid wake-up under dynamic non-stationary conditions of edge computing nodes, and solves the technical problem of long-term all-time monitoring under the limited energy constraints of micro-nano UAVs.

[0016] 2. By utilizing the neighborhood elastic retrieval mechanism of the index arbitration logic module, the problem of index jumps caused by discretized grid boundaries due to inertial sensor measurement noise is solved. When the UAV attitude is in the critical region of the grid, the current grid index and the storage reference hash value of adjacent grid indices are read concurrently based on topological rules, and the minimum Hamming distance is used as the final decision criterion. Based on the breadth-tolerance logic of data retrieval, a logical buffer is constructed in the attitude space, which transforms the small drift and measurement error at the sensor physical level into topological redundancy at the data retrieval level. The mechanism ensures that the system maintains the stability of background feature comparison under the condition of using low-cost and low-precision inertial measurement units, avoids false wake-ups caused by uncertainty of a single index boundary, and enhances the engineering effectiveness of the data preprocessing system under non-ideal sensor conditions without increasing hardware costs and signal processing complexity.

[0017] 3. Establish a dual closed-loop circuit of semantic feedback and kinetic gating between the deep inference unit and the index arbitration logic to achieve adaptive dynamic maintenance of the background feature lookup table and purification of input data; use the angular velocity magnitude as a physical circuit breaker threshold to block motion-blurred images generated during high-dynamic maneuvers from participating in hash calculations or benchmark updates, ensuring that the data source stored in the lookup table is an optically valid sample; use the targetless decision result output by the deep inference unit as a safety signal to trigger the background hash benchmark overwrite under the current illumination conditions; the cross-level feedback update mechanism improves the system's ability to adapt to gradual changes in ambient illumination and scene structural reorganization evolution, enabling the preprocessing logic to automatically correct deviations caused by environmental drift over time, ensuring that the wake-up accuracy of the system does not decay over time during long-term operation, and maintaining a high energy efficiency ratio in complex dynamic environments of edge computing resources. Attached Figure Description

[0018] Figure 1 This is a schematic diagram of the overall logical architecture and data interaction path of the present invention; Figure 2 This is a comparison chart of the Hamming distance response of a single index and a neighborhood elastic retrieval strategy under dynamic operating conditions of the present invention. Figure 3 This is a flowchart illustrating the time-series interaction of multiple modules working collaboratively during the initialization phase of the background model of this invention. Detailed Implementation

[0019] To enable those skilled in the art to better understand the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below in conjunction with the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the scope of protection of the present invention.

[0020] This invention proposes a micro / nano-type UAV edge data preprocessing system, integrated into an onboard processing platform including an inertial measurement unit and an image sensor. It comprises a state-space discretization mapping module, a visual fingerprint extraction module, and an index arbitration logic control module. These modules interact via an on-chip bus. Located between the image sensor and the backend deep inference unit in the data link, this system performs gated filtering on the raw video stream based on the correlation between the UAV's attitude and visual features. When a substantial change in environmental features not caused by the UAV's movement is detected, a wake-up command is generated to activate the dormant backend deep inference unit. Addressing the high-frequency attitude jitter problem inherent in micro / nano-type UAVs during fixed-point monitoring tasks due to the lack of mechanical stabilization devices, the state-space discretization mapping module constructs a discretization verification mechanism based on attitude indexes. This module acquires the current three-dimensional data output from the inertial measurement unit in real time. The three-axis attitude data, including pitch, roll, and yaw angles, is quantized and segmented into a continuous attitude space according to a preset angular resolution step size. This angular resolution step size is set to be less than or equal to the instantaneous field of view of the image sensor, ensuring that the discrete grid coordinates can represent relatively independent observation perspectives. This module maps the quantized three-axis attitude data to a preset three-dimensional discrete coordinate system to generate a unique attitude grid index for the current moment. This index serves as the logical address encoding of the UAV's current spatial orientation, transforming the temporal continuity analysis of the video stream into spatial revisit detection in the attitude domain. The visual fingerprint extraction module performs dimensionality reduction and feature solidification on the original image data. Its input is connected to the signal output interface of the image sensor. After acquiring the original image data, this module executes a fingerprint generation procedure based on a perceptual hash algorithm, specifically including downsampling the original image data to a preset low-resolution size, for example... The pixels are converted into a single-channel grayscale image to remove color redundancy. Then, the discrete cosine transform coefficients of this grayscale image are calculated, and the low-frequency coefficients in the upper left corner of the transform matrix are extracted, for example... The submatrix generates a fixed-length binary string by comparing the magnitude of each coefficient value with the mean coefficient value within the region. This string is the current global feature hash value, which is insensitive to high-frequency random noise but sensitive to structural environmental changes.

[0021] The index arbitration logic control module performs differential arbitration based on historical data. It accesses a background feature lookup table stored in the on-chip static random access memory of the airborne processing platform. This table is constructed as a key-value database with attitude grid indices as keys and baseline background hash values ​​as values. When the state space discretization mapping module outputs a new attitude grid index, the index arbitration logic control module queries the table to see if a corresponding record exists. If the query result is empty, the system performs a baseline entry operation, writing the current global feature hash value as the baseline value to the storage address pointed to by the attitude grid index and maintaining the dormant state of the backend deep inference unit. If the query result indicates that the index already contains a corresponding baseline value, the system calculates the Hamming distance between the current global feature hash value and the stored baseline value. This Hamming distance represents the number of different bits in two binary strings. The system compares this Hamming distance with a preset mutation detection threshold, which is set to a certain percentage of the total number of bits in the hash value. to When the Hamming distance exceeds the threshold, it is determined that an environmental change has occurred within the field of view that is not caused by the aircraft's shaking, and a wake-up command is generated to activate the backend depth inference unit. To solve the problem of high-frequency index jumps at the boundaries of adjacent attitude grids caused by measurement noise from inexpensive inertial measurement units, the index arbitration logic control module integrates a topological neighborhood elastic retrieval unit. When executing a query, this unit generates a candidate index set containing the current attitude grid index and spatially adjacent grid indices based on preset topological adjacency rules. The rules are set as either 26-neighborhood omnidirectional rules defined in the three-dimensional grid space or directional neighborhood rules selected based on the angular velocity direction vector output by the inertial measurement unit. The unit concurrently reads the reference value corresponding to each index in the background feature lookup table and calculates the Hamming distance between the current global feature hash value and the above multiple reference values, and selects the minimum value as the final judgment distance to participate in the threshold comparison. This logic builds a fault-tolerant buffer at the data retrieval level, transforming sensor physical drift into topological redundancy in the retrieval space to eliminate false alarms caused by boundary jitter effects.

[0022] The system also includes a dynamic validity gating module, used to block invalid calculations when the UAV encounters gusts or performs large maneuvers that cause motion blur or shutter distortion in the image sensor. This module monitors the three-axis angular velocity vector output by the inertial measurement unit in real time and calculates its magnitude. When the magnitude exceeds the optically effective threshold determined by the image sensor's exposure time and readout speed, the module generates a processing lockout signal. In response to this signal, the index arbitration logic control module suspends read and write operations on the background feature lookup table and blocks the generation of wake-up commands until the magnitude of the three-axis angular velocity vector falls back below the optically effective threshold. This module is also used to send an update signal in the first cycle after the lockout is released to instruct the system to directly update the reference using the current frame. To address dynamic background interference in nature, the system stores a suppression weight counter associated with each attitude grid index in the background feature lookup table and constructs a negative feedback adaptive adjustment loop connecting the backend deep inference unit and the index arbitration logic control module. When the inference result fed back by the backend deep inference unit indicates that a preset target has not been identified, the system performs an incremental operation on the suppression weight counter associated with the current attitude grid index. The index arbitration logic control module then adjusts the counter according to the formula... The mutation detection threshold of the grid is dynamically increased, whereby... The adjusted mutation determination threshold. As the initial threshold, These are positive weighting coefficients. To suppress the current value of the weight counter, and conversely, if the inference result identifies a valid target, the system generates a count reset signal to clear the corresponding suppression weight counter and restore the mutation judgment threshold to the initial threshold, thereby achieving adaptive desensitization to dynamic noise in a specific area. The system also includes a semantically driven benchmark reset module to solve the problem of historical benchmark failure caused by gradual changes in ambient lighting or permanent changes in scene structure during long-term monitoring. When the inference result of the backend deep inference unit in response to the wake-up command is characterized as no target of interest, this module generates a benchmark overwrite signal. The index arbitration logic control module responds to this signal by obtaining the current global feature hash value output by the visual fingerprint extraction module at the current moment, and overwrites it as the new benchmark value into the storage address pointed to by the current pose grid index in the background feature lookup table. By using the semantic feedback of the backend deep inference unit to calibrate the front-end hash benchmark, the background model is ensured to be automatically updated with changes in the environment.

[0023] Example 1: When a micro / nano-sized UAV performs structural crack monitoring inside a semi-enclosed pipe with a diameter of less than 1 meter, the extremely small physical size and rotational inertia of the UAV cause continuous high-frequency random attitude jitter due to airflow disturbances within the pipe. This manifests as drastic displacement of background pixels across the entire image frame in the continuous image frames acquired by the visual sensor. This causes target detection algorithms based on traditional inter-frame difference or optical flow methods to misidentify the background as a moving target, leading to continuous false wake-ups of the high-power deep inference unit and depleting the onboard battery within minutes. When the system is deployed under the above conditions, the state space discretization mapping module collects the three-axis attitude data output by the inertial measurement unit in real time. Based on an angular resolution step size of 0.5 degrees, the continuous attitude space is quantized into a discrete grid. This step size is smaller than the instantaneous field of view of the image sensor, ensuring that each grid corresponds to an independent observation angle. The system generates the attitude grid index for the current moment. As the sole logical address of the current observation perspective, the visual fingerprint extraction module synchronously acquires the original image data, downsamples it, converts it into a grayscale image, calculates the discrete cosine transform coefficients, extracts low-frequency region coefficients, and generates the current global feature hash value.

[0024] The index arbitration logic control module utilizes The system queries the background feature lookup table in the on-chip static random access memory. If the memory address corresponding to the index is empty, the system writes the current global feature hash value as the base value to that address and keeps the backend deep inference unit in sleep mode. If the base value already exists at the index, the system calculates the Hamming distance between the current global feature hash value and the base value. When this distance is greater than a preset mutation detection threshold, such as a certain percentage of the total number of bits in the hash value, the system will proceed with the query. At that time, a wake-up command is generated, and by converting temporal difference into spatial revisit comparison, optical flow interference caused by fuselage motion is eliminated; for cases where the UAV's attitude changes frequently between adjacent grids due to minor airflow disturbances in the grid boundary region, the topological neighborhood elastic retrieval unit reads concurrently based on directional neighborhood rules. The system calculates and selects the minimum Hamming distance as the final criterion for determining the reference value corresponding to the adjacent grid index on one side of the angular velocity vector pointing to, thereby eliminating false wake-ups caused by discretization boundary effects. At the same time, the dynamic validity gating module monitors the angular velocity magnitude in real time. When a sudden strong airflow causes the angular velocity magnitude to exceed the optical validity threshold, the system blocks the wake-up command and suspends lookup table read and write operations to prevent motion blur images from contaminating the reference library. Under the above logic coordination, the system filters out invalid visual changes caused by body shaking and only activates backend inference when a real crack appears or expands in the pipe wall, causing a sudden change in the hash value. Under dynamic non-stationary operating conditions, the system maintains the low duty cycle operation of the backend deep inference unit to meet the continuous sensing needs of the micro-nano platform under limited energy conditions.

[0025] Example 2: To verify the effectiveness of the edge data preprocessing system proposed in this invention, a verification platform based on physical simulation was built. This platform includes a six-DOF robotic arm, an inertial measurement unit (ICM-20689), a CMOS image sensor (OV5640), and an embedded processing unit (Jetson Nano). Sensor components are fixed to the end effector of the robotic arm to simulate the attitude motion of a micro / nano drone in a complex airflow environment. The test scenario is set as an outdoor environment including trees, grass, and moving pedestrians. Unsteady airflow is generated using a programmable fan array to induce attitude jitter of varying intensities. To simulate noise under real-world conditions, Gaussian white noise with a signal-to-noise ratio of 20 dB is superimposed on the output data of the inertial measurement unit. The experiment was divided into a prototype group and a control group. The prototype group employed a complete inertial-visual hash anchoring mechanism, while the control group used the traditional inter-frame difference method as the triggering mechanism. The experiment verified the system's ability to suppress aircraft jitter. The robotic arm was controlled to simulate the attitude jitter of the drone under different wind speeds, with the jitter amplitude increasing from 0.5 degrees to 5.0 degrees. The number of wake-ups and false wake-up rates of the two systems were recorded. Under light wind conditions (jitter amplitude of 0.5 degrees), the false wake-up rates of both groups were low. However, under strong wind conditions (jitter amplitude of 4.0 degrees), the false wake-up rate of the control group soared to 95.3%, showing extreme sensitivity to the background displacement of the entire screen. In contrast, the false wake-up rate of the prototype group was only 5.2%, proving that the discretization mechanism based on attitude index effectively isolates the interference of aircraft motion.

[0026] Table 1: Comparison of Wake-up Performance under Different Jitter Amplitudes

[0027] In summary, the experimental data objectively confirms the stability of the present invention under complex working conditions such as fuselage vibration, boundary effects, dynamic background, and environmental drift.

[0028] Example 3: This example combines Figures 1 to 3 This document describes a data preprocessing system for the edge of a micro / nano-type unmanned aerial vehicle (UAV). Figure 1As shown, this system includes an inertial measurement unit (IMU), an image sensor, a state space discretization mapping module, a visual fingerprint extraction module, an index arbitration logic control module, a background feature lookup table, a back-end deep inference unit, a dynamic validity gating module, and a negative feedback adaptive adjustment loop. The IMU outputs three-axis attitude data and angular velocity vectors. The state space discretization mapping module receives the three-axis attitude data and maps the quantization to a three-dimensional discrete coordinate system based on the angular resolution step size to generate an attitude mesh index. The visual fingerprint extraction module performs perceptual hashing algorithm dimensionality reduction on the raw image data acquired by the image sensor to calculate the current global feature hash value. The index arbitration logic control module receives the index and hash value respectively and interacts with the background feature lookup table stored in the on-chip static random access memory. It determines whether to generate a wake-up command to activate the back-end deep inference unit, which is in a normal dormant state, through topological neighborhood elastic retrieval and Hamming distance comparison logic. The dynamic validity gating module monitors the angular velocity magnitude in real time and sends a processing lockout signal to the arbitration module to perform circuit breaker protection when the magnitude exceeds the limit. The negative feedback adaptive adjustment loop dynamically adjusts the mutation judgment threshold or triggers semantically driven benchmark reset based on the inference results fed back by the back-end deep inference unit.

[0029] like Figure 2 As shown, the horizontal axis represents time in seconds, and the vertical axis represents the Hamming distance in percentage. The dashed line represents the Hamming distance for a single index query, the solid line represents the Hamming distance for neighborhood elastic retrieval, and the dotted line represents a mutation threshold set at 15%. In the time intervals of 13-16 seconds and 19-21 seconds, the single index query curve fluctuates significantly due to attitude grid boundary effects and exceeds the mutation threshold. In contrast, the neighborhood elastic retrieval curve, by using a minimum distance strategy based on adjacent grid reference values, consistently keeps its calculated value below the mutation threshold. Figure 3 As shown, the pitch, roll, and yaw angle data output by the inertial measurement unit are processed by the state space discretization mapping module to perform zero-bias compensation, coordinate transformation, and angular resolution step-size quantization mapping to generate an attitude mesh index. The visual fingerprint extraction module downsamples the raw image data output by the image sensor to 32×32 pixels and converts it into a grayscale image. After calculating the discrete cosine transform coefficients, it outputs the current global feature hash value. The index arbitration logic control module uses the index... When querying the background feature lookup table, if the query result is empty, it is determined that this is the first time the attitude mesh has been accessed. The system then initiates an M-frame verification sampling sequence, continuously calculates the hash value of the M frames, performs timing consistency checks, and finally synthesizes the baseline hash value by bit majority voting and writes it. The address it points to completes the initial construction of the background model.

[0030] Example 4: Before the micro / nano UAV performs a fixed-point monitoring task, the system performs sensor co-registration and static zero-bias compensation operations to avoid the impact of physical installation errors between the inertial measurement unit and the image sensor, as well as gyroscope temperature drift on attitude index accuracy. During the pre-takeoff self-check phase, the UAV maintains a stationary and horizontal state for a preset calibration time window. During this period, the state space discretization mapping module collects raw triaxial angular velocity data at the maximum sampling rate, calculates the statistical mean and variance of the data to generate a real-time zero-bias compensation vector, and calls the pre-stored external parameter rotation matrix. A rigid body coordinate transformation is performed on the corrected attitude data to map the body coordinate system defined by the inertial measurement unit to the optical coordinate system of the image sensor, thus generating an attitude mesh index. The spatial orientation is geometrically aligned with the actual optical axis orientation.

[0031] The system is configured with baseline confidence establishment logic based on temporal consistency verification, used to fill blank grid records in the background feature lookup table during the initial stage of the mission. When the state space discretization mapping module determines that the UAV has first stably stopped at a certain attitude grid without records, the index arbitration logic control module starts a verification sampling sequence of length M, continuously calculating the current global feature hash value of the M frames in the sequence. When the maximum Hamming distance between any two frame hash values ​​in the sequence is less than the preset consistency stability threshold, the system is considered stable. When the current field of view is determined to be in a stable state, a bitwise majority voting operation is performed on the M binary strings to synthesize a base hash value, which is then written into the corresponding address unit in the on-chip static random access memory to complete the background model initialization for this observation view.

[0032] Example 5: During the system's engineering deployment and parameter initialization phase, to determine the optimal value of the angular resolution step size in the state-space discretization mapping module, ensuring that the attitude mesh division can cover the instantaneous field of view of the image sensor without generating excessive index redundancy, technicians performed a geometric calibration procedure based on the field of view overlap rate, and measured the minimum field of view angle of the image sensor using an optical collimator. The overlap rate coefficient is set according to the minimum requirements of the image overlap region based on the hash algorithm. This coefficient typically ranges from 0.6 to 0.8, according to the formula. Calculate the angular resolution step size The system rounds the calculated result down to an integer multiple of the angular resolution of the inertial measurement unit to determine the final mesh granularity. This ensures that when the UAV transitions from the attitude mesh to adjacent meshes, the images acquired by the visual sensor retain sufficient common feature regions for the hash algorithm to maintain the continuity of comparison. During the parameter calibration of the state space discretization mapping module, a disparity tolerance calibration procedure is performed to determine the angular resolution step size. The lower limit, as specified in the procedure, is based on the far-field approximation principle of the pinhole imaging model. Specifically, when a micro / nano UAV performs a fixed-point monitoring task, such as hovering or micro-motion, if the depth of the observed target is greater than the translational displacement of the UAV within the sampling interval... The visual parallax caused by translation is considered a secondary component and is covered by the tolerance range of the hash algorithm. To quantify this physical constraint, the system uses the formula... Set the grid step size, where The maximum statistical drift amplitude (in meters) of the UAV when hovering in a GPS-free environment is determined by the positioning accuracy of the flight control system. Minimum observation distance (in meters) allowed for the task scenario, for example when Miqie When the meter is used, the calculation is as follows: The system sets the actual angular resolution step size to the upper limit of the calculated value, rounded down. This setting ensures at the physical level that within the space covered by a single attitude grid, the amount of parallax change caused by random translation of the fuselage always falls within the same decision domain of the perceptual hash algorithm.

[0033] To prevent irreversible motion blur caused by excessively high angular velocities from contaminating the background feature lookup table, the system employs a dynamic boundary calculation method based on the physical limits of exposure to solidify this parameter when setting the optical effective threshold in the dynamic effectiveness gating module. During this process, the exposure time of the image sensor is obtained. and the instantaneous field of view corresponding to a single pixel And set the maximum pixel blur displacement that the hash algorithm can tolerate. This displacement is typically set to 2 to 3 pixel units, and then according to the formula... Derive the physical upper limit of the angular velocity magnitude. The calculated value is written into the register of the dynamic validity gating module as a hard circuit breaker threshold. Once the real-time angular velocity magnitude output by the inertial measurement unit exceeds this limit, the system determines that the current image is physically unusable and suspends subsequent processing. For the initial calibration of the mutation judgment threshold in the index arbitration logic control module, to avoid system hypersensitivity or sluggishness caused by empirical settings, a zero-state calibration process based on the statistical characteristics of static environmental noise is executed. The UAV is placed in a static environment with representative texture features. The system continuously acquires K frames of images and calculates the Hamming distance between adjacent frames to construct a distance set characterizing the sensor's inherent noise and environmental perturbations. Next, calculate the mean of the set. with standard deviation And according to the formula Determine the initial mutation threshold ,in As a safety margin coefficient, the value range is set from 4 to 6 to ensure that the threshold setting is based on objective signal-to-noise ratio analysis, providing a reliable logical basis for the dynamic adjustment of the system in subsequent operation.

[0034] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0035] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. A micro / nano-type UAV edge data preprocessing system, characterized in that, The system is configured on an airborne processing platform integrating an inertial measurement unit and an image sensor, and the system includes: The state space discretization mapping module is used to collect the three-axis attitude data output by the inertial measurement unit in real time, and quantize and map the three-axis attitude data to a three-dimensional discrete coordinate system according to the preset angular resolution step size to generate the attitude mesh index at the current moment. The visual fingerprint extraction module is used to acquire the raw image data output by the image sensor and calculate the current global feature hash value of the raw image data based on the perceptual hash algorithm. The index arbitration logic control module, connected to the state space discretization mapping module and the visual fingerprint extraction module, is used to access a preset background feature lookup table and execute the following logic control: When no record matching the attitude grid index is found in the background feature lookup table, the current global feature hash value is written as the baseline value to the storage address pointed to by the attitude grid index in the background feature lookup table, keeping the backend deep inference unit in a dormant state; when a baseline value matching the attitude grid index is found in the background feature lookup table, the Hamming distance between the current global feature hash value and the baseline value is calculated; only when the Hamming distance is greater than the preset mutation judgment threshold, a wake-up command is generated to activate the backend deep inference unit to perform target recognition operations on the original image data.

2. The micro / nano-type UAV edge data preprocessing system as described in claim 1, characterized in that, The index arbitration logic control module also includes a topological neighborhood elastic retrieval unit, which performs the following steps when querying the background feature lookup table: generating a candidate index set containing the attitude grid index and its spatially adjacent grid indices based on preset topological adjacency rules; reading multiple benchmark values ​​corresponding to each index in the background feature lookup table and the candidate index set in parallel; calculating the Hamming distance between the current global feature hash value and each of the multiple benchmark values, and selecting the smallest Hamming distance as the final decision distance; comparing the final decision distance as the Hamming distance with the mutation decision threshold.

3. The micro / nano-type UAV edge data preprocessing system as described in claim 1, characterized in that, The system also includes a dynamic validity gating module, which is used to monitor the three-axis angular velocity vector output by the inertial measurement unit in real time and calculate the magnitude of the three-axis angular velocity vector; when the magnitude is greater than the preset optical validity threshold, a processing lockout signal is generated. The index arbitration logic control module responds to the processing of the latch signal, blocks the generation of the wake-up command and suspends read and write operations on the background feature lookup table until the modulus falls below the optical effective threshold.

4. The micro / nano-type UAV edge data preprocessing system as described in claim 1, characterized in that, The background feature lookup table also stores a suppression weight counter associated with each attitude grid index. The system also includes a negative feedback adaptive adjustment loop, which is connected between the back-end deep inference unit and the index arbitration logic control module. It is used to receive the inference results fed back by the back-end deep inference unit. If the inference result indicates that the preset target has not been identified, an incremental operation is performed on the suppression weight counter associated with the attitude grid index at the current moment. The index arbitration logic control module is used to dynamically adjust the mutation judgment threshold corresponding to the attitude mesh index according to the current value of the suppression weight counter, based on the following formula: ,in, The adjusted mutation determination threshold. The initial threshold preset by the system. These are the preset positive weighting coefficients. To suppress the current value of the weight counter.

5. The micro / nano-type UAV edge data preprocessing system as described in claim 1, characterized in that, The system also includes a semantically driven baseline reset module, which generates a baseline overwrite signal in response to the backend deep inference unit’s feedback on the wake-up command, characterized as an inference result without a target of interest. In response to the baseline overwrite signal, the index arbitration logic control module takes the current global feature hash value at the current moment as the new baseline value and overwrites it in the storage address pointed to by the attitude grid index in the background feature lookup table.

6. The micro / nano-type UAV edge data preprocessing system as described in claim 2, characterized in that, The topological adjacency rules include directional neighborhood rules based on the angular velocity direction vector output by the inertial measurement unit, and 26-neighborhood omnidirectional rules defined in the three-dimensional mesh space; the topological neighborhood elastic retrieval unit is used to preferentially read the reference value corresponding to the adjacent mesh index on the side pointed to by the angular velocity direction vector.

7. The micro / nano-type UAV edge data preprocessing system as described in claim 1, characterized in that, The angular resolution step size configured in the state space discretization mapping module is linearly related to the instantaneous field of view of the image sensor, which is used to ensure that the spatial coverage of the attitude grid index is less than or equal to the coverage of the instantaneous field of view. The background feature lookup table is stored in the on-chip static random access memory of the airborne processing platform.

8. The edge data preprocessing system for micro / nano-type unmanned aerial vehicles as described in claim 1, characterized in that, The visual fingerprint extraction module is used to downsample the original image data to a preset size, convert it into a grayscale image, calculate the discrete cosine transform coefficients of the grayscale image, and extract the low-frequency region coefficients to generate the current global feature hash value. The mutation determination threshold ranges from 10% to 20% of the total number of bits in the current global feature hash value.

9. The micro / nano-type UAV edge data preprocessing system as described in claim 3, characterized in that, The dynamic validity gating module is also used to: send an update signal to the index arbitration logic control module in the first processing cycle after the lockout signal is released; the index arbitration logic control module responds to the update signal, skips the Hamming distance calculation step, and directly executes the operation of writing the current global feature hash value into the background feature lookup table.

10. The micro / nano-type UAV edge data preprocessing system as described in claim 4, characterized in that, The negative feedback adaptive adjustment loop is also used to generate a count reset signal when the inference result fed back by the back-end deep inference unit is characterized as the recognition of a preset target; in response to the count reset signal, the index arbitration logic control module clears the suppression weight counter associated with the current attitude grid index and restores the corresponding mutation judgment threshold to the initial threshold.

Images