Deep learning-based aov micro-power solar black light full-color monitoring processing method and system

Abstract

This invention relates to the fields of video surveillance and low-power power management technology, and particularly to a deep learning-based AOV low-power solar-powered black-light full-color monitoring and processing method and system. The method involves: acquiring the energy storage state to predict available energy; setting operating mode levels and constructing a multi-frame temporal verification configuration coupled with energy state; triggering continuous sampling when a target is detected in a single frame to construct a temporal judgment system for the target bounding box displacement coherence and area gradient trend; calculating a consistency score and waking up the main processor to apply supplementary lighting; and outputting a full-color image after network enhancement. This invention can solve or at least mitigate the problems of frequent false wake-ups, energy imbalance, and undervoltage shutdown caused by fixed loads resulting from conventional single-frame detection, providing a deep learning-based AOV low-power solar-powered black-light full-color monitoring and processing method and system.

Description

Technical Field

[0001] This invention relates to the field of video surveillance and low-power power management technology, and in particular to an AOV low-power solar-powered black light full-color monitoring processing method and system based on deep learning. Background Technology

[0002] Solar-powered low-power monitoring devices are typically equipped with an AOV (Optical Image Viewer) vision front-end for routine low-power target monitoring. In practical applications, AOV detection suffers from a high false wake-up rate, causing the main processor and black light supplementary lighting module to frequently start, consuming limited solar energy reserves. Furthermore, low-power monitoring systems lack the ability to dynamically adjust operating parameters based on solar charging status and battery state of charge, leading to energy imbalances during periods of insufficient charging.

[0003] To address the issues of false wake-ups and energy imbalances caused by AOV detection, existing conventional methods often reduce the trigger frequency by increasing the confidence threshold for target detection in a single frame image, or by cutting off power to the device when the battery voltage reaches a preset lower threshold at the power management level. Simply increasing the confidence threshold can easily lead to missed detections of real moving targets; relying on setting a lower voltage threshold to cut off power fails to establish a linkage between the detection mechanism and the system's internal energy state. In environments with limited charging input, such as continuous rainy weather, the system maintains fixed operating parameters, causing it to prematurely trigger undervoltage protection and shut down.

[0004] The aforementioned defects stem from the fact that conventional single-frame image triggering architectures only extract static spatial features from isolated image frames, lacking analysis of the continuous motion patterns of the target in the temporal dimension. At the algorithmic discrimination level, they cannot distinguish between spatially continuous real-moving targets and transient random physical disturbances such as swaying branches and flickering light and shadow. Management methods that rely solely on real-time voltage cutoff are passive response mechanisms, severing the feedback regulation loop between energy input and power output. Static, fixed operating load models lack the ability to calculate future available energy and fail to establish a dynamic linkage mechanism between power status and system image recognition parameters. The system cannot adaptively adjust the frequency and threshold of image detection based on the current energy storage and charging status.

[0005] This invention introduces a deep learning-based multi-frame timing verification mechanism in the AOV detection link to reduce the false wake-up rate, and introduces a deep learning-based power prediction-driven dynamic switching mechanism for operating modes to address the aforementioned shortcomings. This invention outputs the expected available energy through an energy prediction model, compares the expected available energy with the preset energy consumption models for each operating mode, and then determines the operating mode level. The operating mode level determines the number of verification frames and the consistency judgment threshold configuration for multi-frame timing verification. The consistency score of the timing verification determines whether to wake up the main processor. The operating mode level synchronously determines the black light supplementary lighting power configuration, forming a nested closed loop of energy perception and detection confirmation, enabling the timing verification parameters and supplementary lighting parameters to adaptively adjust according to the energy state. Summary of the Invention

[0006] To achieve the above-mentioned objectives, this invention provides a deep learning-based AOV low-power solar black light full-color monitoring and processing method and system. It aims to solve or at least mitigate the problems of frequent false wake-up deviations caused by transient physical interference due to the lack of continuous motion correlation criteria in the time dimension of traditional single-frame image triggering architecture, the physical defects of energy balance caused by the fixed operating load model under external charging constraints, and the insufficient ability of existing passive power management systems to ensure the continuous operation of equipment under continuous environmental fluctuations.

[0007] To achieve the above objectives, the present invention provides the following technical solution: a deep learning-based method for low-power solar-powered black-light full-color monitoring and processing of AOV (Automated Optical Vehicle), comprising: The auxiliary processor acquires the state of charge data of the lithium battery energy storage module and the output power data of the solar panel. It outputs the expected available energy for a future preset period through a deep learning time-series prediction network. The expected available energy is compared with the preset energy consumption model of each working mode to determine the current working mode level. The multi-frame time-series verification parameters and supplementary light driving parameters corresponding to the working mode are loaded. The multi-frame time-series verification parameters include the number of verification frames and the consistency judgment threshold. The AOV visual front end continuously acquires low-resolution frames, and the lightweight detection unit in the auxiliary processor performs target detection frame by frame. When a target is detected, it enters multi-frame timing verification. The auxiliary processor performs temporal correlation analysis on the target bounding box sequence in the trigger frame and subsequent frames according to the number of verification frames loaded, and outputs a consistency score. When the consistency score reaches the loaded consistency judgment threshold, the main vision processor is woken up. The main vision processor drives the black LED fill light module according to the loaded fill light driving parameters, acquires full-color image frames, and outputs full-color monitoring images after low-light enhancement and color restoration through a deep learning full-color image enhancement network.

[0008] To further realize the present invention, the following technical solutions may be preferred: Preferably, the operating mode levels include full-function mode, balanced mode and power-saving mode, wherein the energy consumption value corresponding to the preset energy consumption model of the full-function mode is higher than that of the balanced mode, and the energy consumption value corresponding to the balanced mode is higher than that of the power-saving mode.

[0009] Preferably, the number of verification frames in the full-function mode is less than the number of verification frames in the power-saving mode, and the consistency determination threshold in the full-function mode is lower than the consistency determination threshold in the power-saving mode.

[0010] Preferably, the multi-frame timing verification parameters further include an inter-frame acquisition interval, and the auxiliary processor acquires subsequent frames after the trigger frame according to the loaded inter-frame acquisition interval.

[0011] Preferably, the temporal correlation analysis includes: extracting the center coordinates and area of ​​the target bounding box in each frame, analyzing the directional coherence of the target displacement and the gradual trend of the target area between frames, and outputting a consistency score.

[0012] Preferably, in the step of waking up the main vision processor, the auxiliary processor sends a wake-up interrupt through the interrupt request line and writes the target position data into the mailbox register. After responding to the interrupt, the main vision processor reads the target position data from the mailbox register.

[0013] Preferably, when the consistency score does not reach the consistency determination threshold, the auxiliary processor clears the frame buffer and returns to the AOV continuous detection state.

[0014] Preferably, the auxiliary processor reacquires state of charge data and output power data at a preset cycle and re-executes the output of expected available energy and determination of the operating mode level. When the operating mode level changes, it synchronously updates multi-frame timing verification parameters and supplementary light drive parameters.

[0015] A deep learning-based AOV low-power solar-powered black light full-color monitoring and processing system for implementing the above method includes: Solar panels, maximum power point tracking circuit, lithium battery energy storage module, power monitoring circuit, AOV vision front end, auxiliary processor, main vision processor, black LED fill light module and network communication interface; The solar panel is connected to the lithium battery energy storage module via the maximum power point tracking circuit. The power monitoring circuit collects the terminal voltage and charging / discharging current of the lithium battery energy storage module and transmits the data to the auxiliary processor. The auxiliary processor is equipped with a lightweight detection unit, a multi-frame timing verification unit, and an energy prediction unit. The main vision processor is equipped with a deep learning full-color image enhancement network.

[0016] Preferably, the auxiliary processor and the main vision processor are connected via an interrupt request line and a mailbox register.

[0017] The beneficial effects of this invention are: This invention adds a multi-frame temporal verification stage to the AOV low-power detection process. It performs temporal correlation analysis on the target bounding box within the trigger frame and subsequent frames, analyzing the continuity of displacement direction and the gradual change trend of area, and outputs a consistency score. By extracting the displacement direction and area change features of the target in consecutive frames, this invention can effectively identify and filter isolated false detection events caused by tree branch swaying or sudden changes in light and shadow. Compared to traditional single-frame static feature extraction, this scheme significantly reduces the false wake-up rate of devices in complex environments. The system relies on the consistency score to intercept invalid event triggers lacking motion continuity, reducing the startup frequency of the main vision processor and the black LED supplementary lighting module, and reducing the additional power consumption of the device.

[0018] This invention constructs a feedforward dynamic adjustment closed loop based on deep learning-based energy timing prediction. It extrapolates the expected available energy for a preset time period based on the output power of the solar panel and the state of charge data of the lithium battery energy storage module, and adaptively updates the operating mode level based on this. When the estimated available energy is limited, the system proactively increases the number of verification frames, raises the consistency judgment threshold, and reduces the driving power of the black light supplementary lighting module to reduce the overall operating load. After charging conditions recover, the verification threshold is simultaneously lowered and the supplementary lighting power during the full-color imaging stage is increased. This mechanism of adaptive adjustment of parameters according to energy status establishes a dynamic coupling match between ensuring image acquisition quality and system energy balance, maintaining the continuous operation of the monitoring equipment under conditions of continuous rain and other charging fluctuations. Attached Figure Description

[0019] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0020] Figure 2 This is a schematic diagram of the method flow of the present invention.

[0021] Figure 3 This is a schematic diagram of the energy prediction and working mode switching principle of the present invention.

[0022] Figure 4 This is a schematic diagram of the multi-frame timing verification principle of the present invention.

[0023] Figure 5 This is a comparison chart of the false wake-up rate of the present invention.

[0024] Figure 6 This is a comparison chart of the average daily power consumption of this invention.

[0025] Figure 7 This is a comparison chart of battery power change curves during continuous rainy days according to the present invention.

[0026] Figure 8 This is a comparison chart of the PSNR of the full-color images of the present invention. Detailed Implementation

[0027] In the description of this invention, it should also be noted that, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," and "link" should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium; and they can refer to the internal connection of two components. Those skilled in the art can understand the specific meaning of the above terms in this invention based on the specific circumstances.

[0028] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. 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 are within the scope of protection of the present invention. Example 1

[0029] This embodiment provides a deep learning-based AOV low-power solar-powered black light full-color monitoring and processing system. For example... Figure 1 As shown, the system mainly consists of a solar panel, a maximum power point tracking circuit, a lithium battery energy storage module, a power monitoring circuit, an AOV vision front end, an auxiliary processor, a main vision processor, a main image sensor, a black LED fill light module, and a network communication interface.

[0030] The solar panel is installed on top of the monitoring equipment, with its tilt angle adjusted according to the deployment latitude. Its output is connected to the input of a maximum power point tracking (MPPT) circuit, which in turn connects to a lithium-ion battery storage module. The MPPT circuit continuously tracks the solar panel's maximum power operating point under varying light intensities, and its output is connected to the charging terminal of the lithium-ion battery storage module. The lithium-ion battery storage module uses lithium iron phosphate batteries, which have a wide operating temperature range and high cycle life, making them suitable for long-term unattended outdoor environments.

[0031] The power monitoring circuit includes voltage sampling resistors and current sampling resistors, which collect terminal voltage and charge / discharge current status data of the lithium battery energy storage module, respectively. The measured data is transmitted to the auxiliary processor after analog-to-digital conversion. The auxiliary processor also receives real-time charging power data output from the maximum power point tracking circuit.

[0032] The AOV vision front-end uses a low-power CMOS image sensor with a resolution of 320×240 pixels. This sensor consumes less than 0.3mW when operating at 2fps. The AOV vision front-end connects to an auxiliary processor, and its data output is connected to the auxiliary processor's data input port via a MIPICSI interface. The auxiliary processor is a low-power microcontroller with milliwatt-level power consumption, integrating an on-chip neural network acceleration unit capable of running lightweight convolutional networks with parameters in the tens of kilobytes range. Figure 1 As shown, the auxiliary processor contains three functional units: a lightweight detection unit, a multi-frame temporal verification unit, and an energy prediction unit. The lightweight detection unit analyzes the images output by the AOV visual front-end frame by frame to determine the presence of targets. The multi-frame temporal verification unit performs temporal correlation analysis on the target detection results across multiple consecutive frames to determine the motion consistency of the detection results. The energy prediction unit estimates the available energy for future periods based on temporal data of battery status and charging power.

[0033] The main vision processor is an embedded processor with deep learning inference capabilities, consuming power in the range of hundreds of milliwatts to watts, and only starts operating upon receiving a wake-up signal. The main vision processor connects to the main image sensor via a MIPI CSI interface. The main image sensor has a resolution of 1920×1080 pixels and supports full-color image acquisition. A black LED fill light module is installed at the front of the device, consisting of a combination of near-infrared LEDs and warm white LEDs; the fill light power can be adjusted via a PWM signal. A deep learning full-color image enhancement network runs within the main vision processor, performing low-light enhancement and color restoration processing on full-color frames acquired under black light conditions. The main vision processor connects to the main image sensor, the black LED fill light module, and the network communication interface.

[0034] The auxiliary processor and the main vision processor communicate via two pathways. The first is an interrupt request line, through which the auxiliary processor sends a wake-up interrupt signal to the main vision processor after confirming a valid target. The second is a mailbox register, where the auxiliary processor writes the target's position coordinates in the AOV frame to the mailbox register, which the main vision processor reads after waking up. The network communication interface supports Wi-Fi or 4G wireless transmission for transmitting the processed full-color monitoring images back to the remote monitoring platform. Example 2

[0035] Based on the processing system provided in Embodiment 1, this embodiment provides a deep learning-based method for AOV low-power solar black light full-color monitoring. For example... Figure 2As shown in the figure, the logical relationships and flow paths of each stage—energy assessment and mode determination, continuous AOV detection, trigger frame determination, multi-frame timing verification, wake-up decision, full-color acquisition and processing, data transmission and hibernation, and periodic energy reassessment—are illustrated. After the system powers on, the auxiliary processor enters the energy assessment phase. The power monitoring circuit transmits the terminal voltage and charging / discharging current data of the lithium battery energy storage module to the auxiliary processor. The terminal voltage data reflects the current state of charge of the battery, while the charging / discharging current data reflects the current load and charging status. Simultaneously, the auxiliary processor reads the output power data of the solar panel after passing through the maximum power point tracking circuit.

[0036] like Figure 3 As shown, the energy prediction unit deploys a deep learning time series prediction network, which is a lightweight recurrent network with a two-layer GRU structure, a hidden dimension of 16, and a parameter size of approximately 15KB. Figure 3 The diagram illustrates that the input to the deep learning time-series prediction network is the current battery terminal voltage, charging / discharging current, and solar power output, along with historical sequences of these three parameters recorded at 1-hour intervals over the past 24 hours. During the offline training phase, the network is trained using six consecutive months of battery charging / discharging records and solar irradiance data from a monitoring station. The network outputs the predicted available energy value for the next four hours, which physically represents the additional electrical energy the lithium-ion battery storage module can provide to the system within the next four hours while maintaining the current baseline power consumption.

[0037] Figure 3 The document also illustrates the logic for determining the operating mode level by comparing the estimated available energy with the energy consumption models of the three operating modes, as well as the mapping relationship between the operating mode level and the timing verification parameters and the supplementary lighting drive parameters. The auxiliary processor compares the estimated available energy with the pre-stored energy consumption models of the three operating modes. The energy consumption model for full-function mode corresponds to an average hourly power consumption of approximately 120mW under full-load operation. This power consumption includes the continuous AOV detection power consumption, timing verification power consumption, the average power consumption of the main processor at the expected wake-up frequency, and the average power consumption of the supplementary lighting module at its rated power. The average hourly power consumption corresponding to balanced mode is approximately 70mW. The average hourly power consumption corresponding to power-saving mode is approximately 35mW. The auxiliary processor converts the estimated available energy into an average power consumption that can be supported over 4 hours, compares it with the energy consumption values ​​of the three modes from high to low, and selects the highest supported operating mode level.

[0038] After determining the operating mode level, the auxiliary processor reads the corresponding multi-frame timing verification parameter table from the on-chip Flash memory. The parameter table for full-function mode is: 3 verification frames, 200ms inter-frame acquisition interval, and a consistency threshold of 0.6. The parameter table for balanced mode is: 5 verification frames, 300ms inter-frame acquisition interval, and a consistency threshold of 0.7. The parameter table for power-saving mode is: 7 verification frames, 500ms inter-frame acquisition interval, and a consistency threshold of 0.8. The logic for setting the number of verification frames is as follows: in power-saving mode, more frames of timing evidence are needed to filter false detections, avoiding wasting power to wake up the main processor when battery power is low; in full-function mode, with ample power, fewer frames are used to achieve a faster response speed. The logic for setting the consistency threshold is similar: in power-saving mode, a higher consistency score is required to allow wake-up, reducing the probability of false wake-ups. The longer inter-frame acquisition interval in power-saving mode reduces the active time of the AOV sensor and auxiliary processor, further reducing basic power consumption.

[0039] The auxiliary processor simultaneously loads the driving parameters of the black LED fill light module corresponding to the current operating mode. In full-function mode, the fill light power is 100% of the rated power, and the corresponding PWM duty cycle is determined by the system during factory calibration. In balanced mode, the fill light power is reduced to 70% of the rated power. In power-saving mode, the fill light power is reduced to 40% of the rated power. The reduction in fill light power will result in a slight decrease in the signal-to-noise ratio of the full-color image, but in power-saving mode, the system prioritizes ensuring continuous operation.

[0040] After energy assessment and parameter loading are completed, the system enters the continuous AOV detection state. The AOV vision front end continuously acquires low-resolution grayscale frames (320×240 resolution) at a frame rate of 2fps and transmits them to the auxiliary processor. The lightweight detection unit in the auxiliary processor performs target detection inference for each frame. The lightweight detection unit is implemented based on a depthwise separable convolutional network quantized to INT8 precision, with a model parameter size of approximately 50KB. The input of this network is a single-frame grayscale image, and the output is a target presence marker, target bounding box coordinates, and detection confidence. The power consumption of inference per frame is approximately 0.5mW. After adding the power consumption of the AOV vision front end, the total power consumption of the continuous AOV detection phase is approximately 0.8mW. When the detection confidence of a frame exceeds a preset threshold of 0.5, the frame is marked as a trigger frame, and the system transitions from the continuous detection state to the multi-frame temporal verification phase. When no target is detected, the auxiliary processor continues to wait for the next frame and repeats the detection process.

[0041] After the trigger frame is generated, the auxiliary processor stores the target bounding box coordinates of that frame into the frame buffer. Subsequently, the auxiliary processor controls the AOV vision front end to continue acquiring subsequent frames according to the inter-frame acquisition interval corresponding to the current operating mode. The number of subsequent frames acquired is determined by the number of verification frames corresponding to the current operating mode. Taking full-function mode as an example, with an inter-frame acquisition interval of 200ms and 3 verification frames, 3 subsequent frames are acquired within 600ms after the trigger frame. After each subsequent frame is acquired, the lightweight detection unit also performs target detection inference and appends the detected target bounding box coordinates into the frame buffer.

[0042] After the target bounding box information of the trigger frame and all subsequent frames is accumulated in the frame buffer, the multi-frame timing verification unit performs timing correlation analysis on the frame sequence. Combined with... Figure 4 The diagram shows the principle of multi-frame temporal verification. It illustrates the frame sequence consisting of the trigger frame and subsequent frames, as well as the extraction process of the center coordinates and area of ​​the target bounding box in each frame. The analysis process is as follows: The multi-frame temporal verification unit extracts the center coordinates of the target bounding box in each frame and calculates the displacement vector of the target center between adjacent frames. The displacement vector sequence is evaluated for directional continuity, i.e., the change in the direction angle of adjacent displacement vectors is calculated and the statistical mean is taken. The evaluation result reflects whether the displacement direction remains continuous. If the target is a real moving object, its displacement direction in consecutive frames shows a certain continuity; if it is a false detection caused by random noise or instantaneous illumination changes, the detection positions in each frame usually lack spatial continuity. Simultaneously, the multi-frame temporal verification unit extracts the area of ​​the target bounding box in each frame and calculates the rate of change of area between adjacent frames. The rate of change of area sequence is evaluated for gradual trend, and the evaluation result reflects whether the area change presents a smooth gradual change rather than an abrupt change. Figure 4 The parallel processing of directional coherence evaluation and area gradient trend evaluation is also shown. The two scores are weighted and combined into a consistency score with a weight ratio of 0.6:0.4, and the score range is between 0 and 1.

[0043] Auxiliary processor execution Figure 4 The determination logic shown compares the consistency score with the consistency determination threshold corresponding to the current working mode. When the consistency score reaches or exceeds the threshold, the auxiliary processor determines that the detection sequence in the frame buffer corresponds to the real target, confirming that the wake-up condition is met. The auxiliary processor sends a wake-up interrupt signal to the main vision processor through the interrupt request line, and simultaneously writes the center coordinates and size information of the target bounding box in the trigger frame into the mailbox register. When the consistency score is lower than the threshold, the auxiliary processor determines that the current detection sequence is a false detection result, then clears all cached data in the frame buffer, and the system returns to the AOV continuous detection state to continue monitoring.

[0044] During the transition period of mode switching, if a mode switch triggered by energy reassessment occurs while the auxiliary processor is executing the timing verification process, the auxiliary processor will update the parameters only after completing the currently ongoing verification process. This is to avoid inconsistencies in the judgment benchmark caused by changes in thresholds during the verification process.

[0045] Upon receiving a wake-up interrupt signal, the main vision processor exits sleep mode, completes internal initialization, and reads the target position data from the mailbox register. Subsequently, based on the illumination drive parameters corresponding to the current operating mode, the main vision processor drives the black LED illumination module to activate via a PWM signal. Under the control of the PWM signal, the illumination module illuminates the near-infrared LED and warm white LED at appropriate power. After the illumination conditions are established, the main vision processor controls the main image sensor to acquire full-color image frames at a resolution of 1920×1080. Due to insufficient ambient light in black-light scenes, the acquired full-color frames exhibit low brightness and color shift.

[0046] A deep learning full-color image enhancement network running in the main vision processor processes the acquired full-color frames. This network comprises two cascaded processing stages. The first stage is low-light enhancement processing, where the network converts the input frame from the sRGB color space to a representation that separates luminance and chrominance, non-linearly boosting the luminance component while suppressing noise introduced during the amplification process. The second stage is color restoration processing, where the network performs color temperature correction on the brightened frames, correcting the color temperature shift caused by near-infrared illumination to a color distribution close to natural lighting conditions. The network parameters for both stages are trained offline using a large paired dataset of black-light scenes. The training data includes low-light full-color frames acquired at different illumination power levels and reference frames under normal lighting conditions. After processing, the output full-color monitoring image is transmitted to a remote monitoring platform via a network communication interface.

[0047] After the full-color image transmission is complete, the main vision processor shuts down the PWM drive signal of the black LED fill light module, and the fill light module is turned off. The main vision processor then enters sleep mode, only retaining the listening function for interrupt request lines. The system as a whole returns to continuous AOV detection mode, and the auxiliary processor continues to execute the low-power detection process.

[0048] During system operation, the auxiliary processor re-executes the energy assessment process every 30 minutes. During each assessment, the auxiliary processor re-acquires the current values ​​of battery terminal voltage, charging / discharging current, and solar output power, and the energy prediction unit re-runs the deep learning time-series prediction network to output updated predicted available energy values. If the new predicted available energy value causes a change in the operating mode level, the auxiliary processor reloads the corresponding time-series verification parameter table and supplementary lighting drive parameters from memory. The updated parameters take effect after the current verification process ends. In the event of recharging recovery, such as after a cloudy day when the weather clears up and solar output power rebounds, the system operating mode level may rise from power-saving mode to balanced mode or full-function mode. In this case, the number of verification frames for time-series verification decreases, the threshold decreases, and the system response speed improves; the supplementary lighting power increases, and the quality of the full-color image improves accordingly.

[0049] Example 3 Verification Test This embodiment was tested and verified in an outdoor solar-powered monitoring scenario. The test location was an open site in North China, and the deployment time was summer. The rated power of the solar panel was 5W, and the capacity of the lithium battery energy storage module was 10000mAh / 3.7V (approximately 37Wh). The test period was 7 consecutive days, with the first 3 days being sunny to partly cloudy, the 4th to 6th days being continuous overcast and rainy, and the 7th day returning to sunny weather.

[0050] The comparative scheme is a traditional AOV binary wake-up scheme. In this scheme, the AOV detection unit directly wakes up the main processor when it detects a target, without going through a timing verification step; the system has no energy sensing mode switching function and always operates with fixed parameters.

[0051] In the false wake-up rate test, various scenarios were set up within the monitoring field of view, including pedestrians, vehicles, swaying tree branches, changes in light and shadow, and small animals passing by. For example... Figure 5 The chart showing the false wake-up rate comparison compares the false wake-up rates of the comparative example and the embodiment in full-function mode, balanced mode, and power-saving mode using a bar chart. During the 7-day test, the comparative example woke up the main processor a total of 4127 times, of which 3153 were valid wake-ups confirmed by manual verification, and 974 were false wake-ups, with an average false wake-up rate of 23.6%. The distribution of false wake-up events was as follows: tree branch swaying accounted for approximately 38%, changes in light and shadow accounted for approximately 29%, small animals passing by accounted for approximately 22%, and other random interference accounted for approximately 11%. In the embodiment, during full-function mode operation, the main processor was woken up a total of 1862 times, with 1709 valid wake-ups and 153 false wake-ups, resulting in a false wake-up rate of 8.2%. During balanced mode operation, the false wake-up rate was 4.7%. During power-saving mode operation, the false wake-up rate was 2.1%. The timing verification showed particularly outstanding filtering effect on transient interference such as tree branch swaying and changes in light and shadow, reducing these two types of false wake-ups by approximately 78% and 83%, respectively, in the embodiment.

[0052] In the system's daily power consumption test, the total battery discharge over 24 hours was recorded by the power monitoring circuit and converted into average power consumption. For example... Figure 6 The chart showing the daily average power consumption comparison uses a bar chart to compare the daily average power consumption performance of the comparative example and the embodiment. The comparative example's 7-day average daily power consumption was 142mW. The embodiment's 7-day average daily power consumption was 78mW, a reduction of approximately 45% compared to the comparative example. The power consumption reduction mainly comes from two aspects: the timing verification mechanism filters out a large number of false wake-ups, reducing the number of invalid startups of the main processor and the supplementary lighting module; and the energy-sensing mode switching reduces the supplementary lighting power and system activity during periods of low power.

[0053] In the continuous rainy day battery life test, such as Figure 7 The graph shows a comparison of battery power changes over consecutive rainy days. It illustrates the remaining battery power over time for the comparative and embodiment under three consecutive days of rain. From day 4 to day 6, continuous rain occurred, and the average daily charge of the solar panels was approximately 18% of that on sunny days (approximately 22Wh on sunny days and approximately 4Wh on rainy days). In the comparative example, the battery power dropped below the protection voltage on the morning of day 5 (the second day of continuous rain), causing the system to shut down. In the embodiment, after detecting a decrease in expected available energy on the afternoon of day 4, the system automatically switched from full-function mode to balanced mode, and further switched to power-saving mode on the evening of day 4. In power-saving mode, the system's average daily power consumption dropped to approximately 35mW, maintaining normal operation during the three consecutive days of rain. After the weather recovered on day 7, the system automatically returned to full-function mode.

[0054] In full-color image quality testing, such as Figure 8 The PSNR comparison chart for full-color images shown uses Peak Signal-to-Noise Ratio (PSNR) to evaluate full-color images acquired in black light scenes and enhanced by deep learning. The chart uses a bar chart to compare the average PSNR of the full-color images in the comparative and embodiment models under various operating modes. The test was based on a reference image acquired manually under normal lighting conditions using the same main image sensor. The average PSNR of the comparative full-color image is 28.3 dB. The average PSNR of the embodiment model is 29.1 dB in full-function mode, 27.8 dB in balanced mode, and 26.2 dB in power-saving mode. The image quality in full-function mode is better than the comparative model because the main processor only starts on valid events that have been verified by timing, and processor resources are concentrated on enhancing valid frames. The image quality in balanced and power-saving modes is lower than the comparative model because the reduced supplementary light power leads to a decrease in the signal-to-noise ratio of the original frame.

[0055] In the effective target detection rate test, all manually labeled effective target events were used as the benchmark. The comparative effective target detection rate was 91.4%. The example showed a detection rate of 90.8% in full-function mode, 89.3% in balanced mode, and 85.6% in power-saving mode. The decrease in detection rate in power-saving mode was due to the increased number of verification frames and the increased threshold, which caused some fast-moving and short-duration effective targets to fail the timing verification.

[0056] The above are merely preferred embodiments of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A deep learning-based method for low-power solar-powered black-light full-color monitoring and processing of AOV (Automated Optical Vehicle) devices, characterized in that... include: The auxiliary processor acquires the state of charge data of the lithium battery energy storage module and the output power data of the solar panel. It outputs the expected available energy for a future preset period through a deep learning time-series prediction network. The expected available energy is compared with the preset energy consumption model of each working mode to determine the current working mode level. The multi-frame time-series verification parameters and supplementary light driving parameters corresponding to the working mode are loaded. The multi-frame time-series verification parameters include the number of verification frames and the consistency judgment threshold. The AOV visual front end continuously acquires low-resolution frames, and the lightweight detection unit in the auxiliary processor performs target detection frame by frame. When a target is detected, it enters multi-frame timing verification. The auxiliary processor performs temporal correlation analysis on the target bounding box sequence in the trigger frame and subsequent frames according to the number of verification frames loaded, and outputs a consistency score. When the consistency score reaches the loaded consistency judgment threshold, the main vision processor is woken up. The main vision processor drives the black LED fill light module according to the loaded fill light driving parameters, acquires full-color image frames, and outputs full-color monitoring images after low-light enhancement and color restoration through a deep learning full-color image enhancement network.

2. The method according to claim 1, characterized in that, The operating mode levels include full-function mode, balanced mode and power-saving mode. The energy consumption value corresponding to the preset energy consumption model of the full-function mode is higher than that of the balanced mode, and the energy consumption value corresponding to the balanced mode is higher than that of the power-saving mode.

3. The method according to claim 2, characterized in that, The number of verification frames in the full-function mode is less than the number of verification frames in the power-saving mode, and the consistency determination threshold in the full-function mode is lower than the consistency determination threshold in the power-saving mode.

4. The method according to claim 1, characterized in that, The multi-frame timing verification parameters also include the inter-frame acquisition interval, and the auxiliary processor acquires subsequent frames after the trigger frame according to the loaded inter-frame acquisition interval.

5. The method according to claim 1, characterized in that, The temporal correlation analysis includes: extracting the center coordinates and area of ​​the target bounding box in each frame, analyzing the directional coherence of target displacement and the gradual trend of target area change between frames, and outputting a consistency score.

6. The method according to claim 1, characterized in that, In the step of waking up the main vision processor, the auxiliary processor sends a wake-up interrupt through the interrupt request line and writes the target position data into the mailbox register. After responding to the interrupt, the main vision processor reads the target position data from the mailbox register.

7. The method according to claim 1, characterized in that, When the consistency score fails to reach the consistency determination threshold, the auxiliary processor clears the frame buffer and returns to the AOV continuous detection state.

8. The method according to claim 1, characterized in that, The auxiliary processor reacquires state of charge data and output power data at a preset cycle and re-executes the output of expected available energy and determination of the operating mode level. When the operating mode level changes, it synchronously updates multi-frame timing verification parameters and supplementary light drive parameters.

9. A deep learning-based AOV low-power solar-powered black light full-color monitoring and processing system, used to implement the method according to any one of claims 1 to 8, characterized in that, include: Solar panels, maximum power point tracking circuit, lithium battery energy storage module, power monitoring circuit, AOV vision front end, auxiliary processor, main vision processor, black LED fill light module and network communication interface; The solar panel is connected to the lithium battery energy storage module via the maximum power point tracking circuit. The power monitoring circuit collects the terminal voltage and charging / discharging current of the lithium battery energy storage module and transmits the data to the auxiliary processor. The auxiliary processor is equipped with a lightweight detection unit, a multi-frame timing verification unit, and an energy prediction unit. The main vision processor is equipped with a deep learning full-color image enhancement network.

10. The system according to claim 9, characterized in that, The auxiliary processor and the main vision processor are connected via an interrupt request line and a mailbox register.

Images