ARM-based embedded remote monitoring camera

Abstract

The invention discloses an embedded remote monitoring camera based on an ARM, and relates to the technical field of video monitoring, and the system comprises an ARM processor module, an ARM processor carrying a built-in NEON instruction set, a running Linux embedded system, and an integrated dynamic power consumption management bus; the target detection module is used for outputting a detection and prediction result through a machine learning algorithm optimized by an NEON instruction set; the multi-mode sensor interface interacts with the ARM processor module through a dynamic power consumption management bus and dynamically adjusts hardware-level synchronous sampling of the millimeter-wave radar, the infrared sensor and visual data; and the secure transmission module shares the NEON hardware acceleration unit with the target detection module.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of video monitoring, in particular to an ARM-based embedded remote monitoring camera. BACKGROUND

[0002] With the rapid development of information technology, video monitoring technology has been widely used in security, home, industry and other fields. From the initial local analog signal monitoring system to the PC card-based digital monitoring system, today it is developing towards network digital monitoring system based on embedded technology. Embedded remote monitoring technology has gradually become a research and application hotspot due to its advantages in unattended, remote management and other scenarios. At the same time, people's requirements for the intelligence, portability, low power consumption and other performance of remote monitoring equipment are getting higher and higher, which promotes the development of ARM-based embedded remote monitoring cameras.

[0003] Remote monitoring cameras have many important roles. Users can remotely access the camera using a computer, mobile phone or other terminal device through the Internet in any place with network connection, and view the monitoring area in real time to realize the monitoring and management of remote places such as home security, office monitoring and industrial production monitoring.

[0004] Due to the small size and low power consumption of embedded systems, ARM-based embedded cameras have good portability and can be easily installed in various places such as small warehouses and outdoor temporary monitoring points to meet the monitoring needs in different scenarios. However, the existing scheme does not fully utilize the instruction set characteristics of ARM architecture, and the frame rate fluctuation exceeds ±40% in embedded scenarios; which makes it difficult to fully utilize the performance of ARM architecture in target detection and other tasks, thereby affecting the overall detection efficiency and stability.

[0005] At present, there is no effective solution to the above problems. SUMMARY

[0006] The embodiments of the present application provide an ARM-based embedded remote monitoring camera to solve the above technical problems.

[0007] The present application provides an ARM-based embedded remote monitoring camera, comprising: An ARM processor module, which carries an ARM processor with a built-in NEON instruction set, runs a Linux embedded system, and integrates a dynamic power management bus; A target detection module for outputting detection and prediction results through a machine learning algorithm optimized by the NEON instruction set; A multi-modal sensor interface interacts with the ARM processor module through the dynamic power management bus, dynamically adjusting the hardware-level synchronous sampling of millimeter wave radar, infrared sensors, and visual data; A secure transmission module shares a NEON hardware acceleration unit with the target detection module.

[0008] Further, the target detection module includes: A first-level detection layer: a YOLOv5-Tiny network optimized by channel pruning, using NEON instruction set to calculate 3x3 depth separable convolution of INT8 quantized features in parallel; A second-level tracking layer: combined with the Farneback optical flow algorithm, the ARMMaliGPU's OpenCL kernel analyzes the cross-frame trajectory persistence of moving targets; A third-level prediction layer: a spatio-temporal graph convolution network is deployed to predict the next activity area of the target based on the target position and velocity vector output by the previous two levels, and the predicted coordinates are fed back to the camera holder control unit.

[0009] Further, the channel pruning optimization of the first-level detection layer is achieved by the following steps: Based on the L1 norm of the weight tensor, the importance score of each output channel is dynamically calculated, and the channel whose importance score is lower than the adaptive threshold is marked as redundant; The pruning operation of the redundant channel is performed in parallel using the NEON instruction set, preserving the spatial correlation features between channels; A pruning mapping table is constructed in the ARM processor L2 cache to dynamically skip the convolution calculation of the pruned channel; The pruned network layer is retrained with 8-bit fixed-point through the CMSIS-NN library to compensate for the loss of accuracy.

[0010] Further, the calculation of the INT8 quantized features includes: The sliding average and dynamic standard deviation of the current convolution layer input feature map are calculated in real time; According to the statistical value, a non-linear quantization mapping function is constructed to divide the feature values into non-uniform intervals; Through the parallel processing function of the NEON instruction set, the product accumulation operation is performed on the quantized feature map and the weight matrix; A lightweight calibration network is embedded in the quantization output layer to dynamically compensate for quantization errors.

[0011] Further, the operation of the spatio-temporal graph convolution network includes: A spatio-temporal topology graph is constructed with the target as the center node, and the node attributes include position, speed, and appearance features; The spatio-temporal attention weight between nodes is calculated in parallel using the NEON instruction set to filter key associated nodes; Fuse historical motion trajectory with current observation data through gated recurrent unit; Output confidence distribution map of target position in future 5 frames, and generate gimbal pre-aiming coordinates.

[0012] Further, the multi-modal sensor interface realizes: Intelligent sampling mechanism for increasing sampling rate of the millimeter wave radar from 10 Hz to 30 Hz when the target detection module identifies human body target; Data fusion unit for spatio-temporal alignment of visual data and radar point cloud in L2 cache through DMA controller built in ARM processor, to generate motion trajectory with depth information; Abnormal linkage strategy for starting fire feature recognition sub-model of the target detection module when temperature change amount to time change amount ratio detected by temperature and humidity sensor is greater than threshold value.

[0013] Further, the ARM-based embedded remote monitoring camera further comprises: Dynamic power management module for receiving motion intensity data output by the target detection module and distributing processor core voltage in real time through the dynamic power management bus.

[0014] Further, the dynamic power management module comprises: Load awareness unit for monitoring BBox generation rate of the target detection module and triggering DVFS strategy when the rate is < 5 frames / second; Hierarchical regulation strategy for setting three voltage levels according to number of detected targets, and increasing voltage level by one for each additional 3 tracking targets; The hierarchical regulation strategy comprises: Establishing nonlinear mapping relationship between target number and voltage level, and triggering voltage increase request when target number reaches integer multiple of 3; Broadcasting current voltage state to the multi-modal sensor interface through the dynamic power management bus, and adjusting sensor sampling rate in linkage; Embedding voltage mutation suppression algorithm in the DVFS strategy, to limit single voltage adjustment amplitude to no more than 0.05 V.

[0015] Further, the secure transmission module comprises: Differential encryption unit for encoding normal background frames using H.265 standard, and superimposing SM9 algorithm encryption on feature frames containing human face; Key linkage mechanism for dynamically generating temporary session key through ARM TrustZone when the target detection module identifies sensitive target; A secure storage area for storing encrypted data and algorithm parameters in a physically partitioned manner using the TZC-400 storage controller of the ARM processor.

[0016] Further, the NEON instruction set includes a NEON SIMD instruction set; The ARM-based embedded remote monitoring camera further includes: An adaptive focusing module for controlling the camera holder to perform 3 times digital zoom in the predicted area according to the output coordinates of the third-level prediction layer; A multi-algorithm cooperation mechanism for closing the long-term prediction function of the ST-GCN and switching to the Kalman filter short-term prediction when the optical flow algorithm detects a motion speed > 5 m / s; A resource arbitration unit for dynamically allocating the time slice ratio of the NEON unit between algorithm operation and encryption operation through the CCI-550 interconnection bus of the ARM processor.

[0017] Based on the embodiments provided in the present application, the computing power is significantly enhanced through the NEON instruction set built-in the ARM processor module, especially when dealing with a large amount of data related to target detection. This enables the camera to complete tasks more quickly and efficiently when performing real-time monitoring and target detection, effectively reducing data processing delay and improving overall performance. The dynamic power management bus integrated in the ARM processor module can intelligently adjust the power consumption of the hardware modules according to the working state and task requirements of the camera. It reduces power consumption at low load and provides sufficient energy support at high load, thereby prolonging battery life and reducing energy consumption, especially suitable for remote monitoring scenarios that require long-term continuous operation. The multi-modal sensor interface interacts with the ARM processor module through the dynamic power management bus, achieving hardware-level synchronous sampling of millimeter wave radar, infrared sensors and visual data. This synchronization mechanism ensures the consistency of data from different sensors in time, providing more comprehensive and accurate information for subsequent target detection and analysis, and improving the system's perception ability in complex environments.

[0018] The secure transmission module and the target detection module share the NEON hardware acceleration unit, which not only improves the security of data transmission, but also ensures the efficiency of data processing through hardware-level acceleration. When transmitting sensitive monitoring data, it can effectively prevent data leakage and tampering, ensuring the confidentiality and integrity of monitoring information. The machine learning algorithm optimized by the NEON instruction set makes the target detection module more stable and reliable when dealing with target recognition in complex scenarios. At the same time, through the efficient cooperation and optimization design among the modules, the system reduces the unstable factors caused by hardware or software failures, ensuring the long-term stable operation of the remote monitoring camera in various harsh environments. BRIEF DESCRIPTION OF DRAWINGS

[0019] The accompanying drawings, which are included to provide a further understanding of the embodiments of the application and are incorporated in and constitute a part of this application, illustrate embodiments of the application and together with the description serve to explain the application. In the drawings: Figure 1 An optional structure diagram of an ARM-based embedded remote monitoring camera according to an embodiment of the application; Figure 2 An optional flowchart of channel pruning optimization for realizing a first-level detection layer according to an embodiment of the application; Figure 3 An optional flowchart of realizing a hierarchical adjustment strategy according to an embodiment of the application.

[0020] The implementation, functional features and advantages of the application will be further described with reference to the embodiments and the accompanying drawings. DETAILED DESCRIPTION

[0021] The technical solutions in the embodiments of the application will be described clearly and completely below with reference to the drawings in the embodiments of the application. Obviously, the described embodiments are only part of the embodiments of the application, rather than all the embodiments of the application. Based on the embodiments in the application, all other embodiments obtained by those skilled in the art without creative work fall within the protection scope of the application.

[0022] Optionally, as shown in the figure, Figure 1 The application provides an ARM-based embedded remote monitoring camera, which comprises: An ARM processor module 101, which is loaded with an ARM processor with a built-in NEON instruction set, runs a Linux embedded system, and is integrated with a dynamic power management bus, i.e., DPMB; A target detection module 102, which is used to output detection and prediction results through a machine learning algorithm optimized by the NEON instruction set; In this embodiment, after the detection and prediction results are output, the detection and prediction results are synchronized to a dynamic power management module; A multi-modal sensor interface 103, which interacts with the ARM processor module through the dynamic power management bus, and dynamically adjusts the hardware-level synchronous sampling of millimeter wave radar, infrared sensors and visual data; The multi-modal sensor interface 103 is integrated with a temperature and humidity sensor and a millimeter wave radar connected through an I2C / SPI bus. A secure transmission module 104, which shares a NEON hardware acceleration unit with the target detection module.

[0023] Based on the embodiments provided in this application, the computing power is significantly enhanced through the NEON instruction set built-in the ARM processor module, especially when dealing with a large amount of data related to target detection. This enables the camera to perform real-time monitoring and target detection more quickly and efficiently, effectively reducing data processing delays and improving overall performance. The dynamic power management bus integrated in the ARM processor module can intelligently adjust the power consumption of the hardware modules according to the working state and task requirements of the camera. It reduces power consumption under low load and provides sufficient energy support under high load, thereby prolonging battery life and reducing energy consumption, especially suitable for remote monitoring scenarios that require long-term continuous operation. The multi-modal sensor interface interacts with the ARM processor module through the dynamic power management bus, achieving hardware-level synchronous sampling of millimeter wave radar, infrared sensors, and visual data. This synchronization mechanism ensures the temporal consistency of data from different sensors, providing more comprehensive and accurate information for subsequent target detection and analysis, and improving the system's perception ability in complex environments.

[0024] The secure transmission module shares the NEON hardware acceleration unit with the target detection module, which not only improves the security of data transmission but also ensures the efficiency of data processing through hardware-level acceleration. When transmitting sensitive monitoring data, it can effectively prevent data leakage and tampering, ensuring the confidentiality and integrity of monitoring information. The machine learning algorithm optimized by the NEON instruction set makes the target detection module more stable and reliable when dealing with target recognition in complex scenarios. At the same time, through efficient cooperation and optimized design among modules, the system reduces the instability factors caused by hardware or software failures, ensuring the long-term stable operation of the remote monitoring camera in various harsh environments.

[0025] Further, the target detection module comprises: The first detection layer adopts the YOLOv5-Tiny network optimized by channel pruning, and uses the NEON instruction set to perform parallel calculation on the INT8 quantized features of 3x3 depth separable convolution; The second tracking layer combines the Farneback optical flow algorithm and analyzes the cross-frame trajectory persistence of moving targets through the OpenCL kernel of ARMMaliGPU; The third prediction layer deploys the spatio-temporal graph convolution network (ST-GCN), predicts the next activity area of the target according to the target position and velocity vector output by the previous two levels, and feeds back the predicted coordinates to the camera gimbal control unit.

[0026] Based on the embodiments provided in the present application, efficient target positioning is realized through the detection layer (YOLOv5-Tiny), the cross-frame continuity is ensured through the tracking layer (optical flow algorithm), the spatio-temporal motion model is established through the prediction layer (ST-GCN), and a complete "perception-tracking-prediction" chain is formed. In the existing scheme, the detection and prediction modules are separated, and in the embodiments of the present application, end-to-end modeling is realized through the spatio-temporal graph convolution network, and the problem of trajectory break in the motion mutation scene is solved.

[0027] Further, as shown in Figure 2 , the channel pruning optimization of the first-level detection layer is realized through the following steps: S201, the importance score of each output channel is dynamically calculated based on the L1 norm of the weight tensor, and the channel whose importance score is lower than the adaptive threshold is marked as redundant; S202, the pruning operation of the redundant channel is performed in parallel using the NEON instruction set, and the spatial correlation feature between channels is preserved; S203, a pruning mapping table is constructed in the L2 cache of the ARM processor, and the convolution calculation of the pruned channel is dynamically skipped; S204, the network layer after pruning is retrained by 8-bit fixed-point through the CMSIS-NN library, and the precision loss is compensated.

[0028] In the embodiments of the present application, the channel importance score is calculated based on the following formula: Wherein, is the channel importance score (ChannelSignificanceScore), which dynamically measures the contribution of the cth output channel in the convolution layer; is the L1 norm of the weight matrix of the cth output channel, which represents the activation strength of the channel; is the maximum value of the L1 norm of all output channels of the current convolution layer, which is used for normalization processing; is the model precision change (PrecisionDeviation) after pruning the cth channel, which is calculated through inference testing; is the exponential moving average (ExponentialMovingAverage) of the precision change, which reflects the long-term pruning impact; is the state balance coefficient, which is adjusted in the range of 0.2 to 0.5 according to the current processor load.

[0029] ​​The scoring mechanism dynamically fuses the channel activation strength (L1 norm) and the historical pruning impact (EMA accuracy change) to achieve adaptive pruning in embedded scenarios. Specifically, when the processor load is high (λ tends to 0.5), the formula focuses on short-term activation strength, quickly prunes redundant channels to reduce computational load; when the load is low (λ tends to 0.2), the above formula strengthens long-term accuracy compensation to avoid excessive pruning leading to model failure. Through real-time adjustment of the dynamic balance coefficient λ, the embodiment realizes the optimal trade-off between model compression rate and detection accuracy on the ARM platform.

[0030] Based on the embodiments provided in the present application, the dual-factor evaluation mechanism based on L1 norm and accuracy change avoids over-pruning caused by a single indicator; the EMA accuracy compensation mechanism is introduced to adapt to the dynamic load changes in the embedded environment.

[0031] Further, the calculation of INT8 quantized features includes: Real-time statistics of the sliding average value and dynamic standard deviation of the current convolution layer input feature map; According to the statistical value, a non-linear quantization mapping function is constructed to divide the feature values into non-uniform intervals; Through the parallel processing function of the NEON instruction set, the product accumulation operation is performed on the quantized feature map and the weight matrix; A lightweight calibration network is embedded in the quantization output layer to dynamically compensate for quantization errors.

[0032] In the embodiments of the present application, Wherein, is the input feature value, the original floating point output from the convolution layer; is the sliding average value of the current feature map, calculated in real time through a ring buffer; is the dynamic standard deviation, calculated by block statistics according to the feature map area; is the hyperbolic tangent function, used to compress the quantization range of extreme feature values; is the rounding function, which maps continuous values to discrete INT8 integers.

[0033] Further, the operation of the spatio-temporal graph convolution network includes: A spatio-temporal topology graph is constructed with the target as the center node, and the node attributes include location, speed and appearance features; The spatio-temporal attention weights between nodes are calculated in parallel using the NEON instruction set to filter key associated nodes; The historical motion trajectory and current observation data are fused through a gated recurrent unit; The confidence distribution map of the target position in the next 5 frames is output, and the gimbal pre-aiming coordinates are generated.

[0034] In the embodiments of the present application, the spatio-temporal attention weight is calculated based on the following formula: wherein, is the spatio-temporal attention coefficient, representing the attention degree of node u to node v; is the query vector, generated by position encoding of the target ; is the key vector, generated by motion feature extraction of the target v; is the vector dot product operation accelerated based on ARM NEON instruction set; is the spatio-temporal neighborhood set of node u, containing the associated targets in the current frame and the previous two frames; is the motion feature extraction of the target .

[0035] In some embodiments, the operation is realized in parallel through ARM NEON instruction set: the query vector and the key vector are divided into 128-bit data blocks, and the multiply-accumulate (MAC) operation is performed in parallel using the VMLA instruction of NEON; a single instruction simultaneously calculates the dot product of 4 groups of 32-bit floating-point numbers, so that the operation speed is increased to 3.2 times that of traditional scalar calculation.

[0036] The process of screening key associated nodes includes: calculating the attention scores of all neighborhood nodes ; screening out the nodes with top 20% scores as key associated nodes through parallel comparison of scores by NEON instruction set; and performing subsequent gated recurrent unit calculation on the key nodes to reduce the inference time consumption of the spatio-temporal graph convolution network. Based on the embodiments provided in the present application, the dot product operation accelerated by NEON is used to screen key associated nodes, and the interference of irrelevant targets is suppressed; through hardware-level acceleration and dynamic screening, the computational load of the embedded platform is significantly reduced, while the prediction accuracy in complex scenes is guaranteed; the existing scheme uses a fixed time window for prediction, and the present application models the target interaction relationship through a spatio-temporal topology graph, thereby improving the prediction robustness in complex scenes.

[0037] Further, the multi-modal sensor interface realizes: an intelligent sampling mechanism, which is used to increase the sampling rate of the millimeter wave radar from 10 Hz to 30 Hz when the target detection module identifies a human target; a data fusion unit, which is used to perform spatio-temporal alignment of visual data and radar point cloud in L2 cache through the DMA controller built in the ARM processor, to generate a motion trajectory with depth information; An abnormal linkage strategy is used to start a fire feature recognition sub-model of the target detection module when the temperature change amount to time change amount ratio detected by the temperature and humidity sensor is greater than a threshold value. For example, the threshold value is 5℃ / min.

[0038] Further, the ARM-based embedded remote monitoring camera further comprises: A dynamic power consumption management module is configured to receive the motion intensity data output by the target detection module and to distribute the processor core voltage in real time through a dynamic power consumption management bus.

[0039] Further, the dynamic power consumption management module comprises: A load sensing unit is configured to monitor the BBox generation rate of the target detection module and to trigger the DVFS strategy when the rate is less than 5 frames per second; A hierarchical regulation strategy is used to set three voltage levels according to the number of detected targets, and to increase the voltage level by one level for each additional 3 tracking targets; for example, the three voltage levels are 0.9V, 1.2V and 1.5V, which can avoid abnormal sensor data caused by voltage mutation.

[0040] As shown in Figure 3 the implementation of the hierarchical regulation strategy comprises: S301, a nonlinear mapping relationship between the target number and the voltage level is established, and a voltage increase request is triggered when the target number reaches an integer multiple of 3; S302, the current voltage state is broadcast to the multi-modal sensor interface through the dynamic power consumption management bus to link and adjust the sensor sampling rate; S303, a voltage mutation suppression algorithm is embedded in the DVFS strategy to limit the single voltage adjustment amplitude to not more than 0.05V.

[0041] wherein, VoltageStep is the voltage adjustment step (VoltageStep), the maximum voltage change allowed at a time; The current number of tracking targets is calculated in real time by the target detection module; The number of tracking targets in the previous period is used to calculate the target number change rate; is an exponential decay function that suppresses voltage oscillation caused by target number mutation.

[0042] Further, the secure transmission module comprises: A differential encryption unit is configured to encode the ordinary background frame using the H.265 standard and to superimpose the SM9 algorithm encryption on the feature frame containing the face; A key linkage mechanism is used to dynamically generate a temporary session key through the ARM TrustZone when the target detection module identifies a sensitive target. A secure storage area for storing encrypted data and algorithm parameters in a physically partitioned manner using a TZC-400 storage controller of an ARM processor.

[0043] Further, the NEON instruction set comprises a NEON SIMD instruction set; The ARM-based embedded remote monitoring camera further comprises: An adaptive focusing module configured to control the camera holder to perform 3 times digital zoom in a predicted area according to output coordinates of the third prediction layer; A multi-algorithm cooperation mechanism configured to turn off the long-term prediction function of the ST-GCN and switch to the Kalman filter short-term prediction when the optical flow algorithm detects a motion speed greater than 5 m / s. A resource arbitration unit configured to dynamically allocate the time slice ratio of the NEON unit between algorithm operation and encryption operation through a CCI-550 interconnection bus of the ARM processor.

[0044] The above is only the preferred embodiment of the present application, and does not limit the patent scope of the present application, and any equivalent structure or equivalent process transformation using the content of the specification and drawings, or direct or indirect application in other related technical fields, are also included in the patent protection scope of the present application.

Claims

1. An ARM-based embedded remote monitoring camera, characterized in that, include: The ARM processor module is equipped with an ARM processor with built-in NEON instruction set, runs a Linux embedded system, and integrates a dynamic power management bus. The target detection module is used to output detection and prediction results through machine learning algorithms optimized by the NEON instruction set; The multimodal sensor interface interacts with the ARM processor module through the dynamic power management bus to dynamically adjust the hardware-level synchronous sampling of millimeter-wave radar, infrared sensor and visual data; The secure transmission module shares the NEON hardware acceleration unit with the target detection module.

2. The ARM-based embedded remote monitoring camera according to claim 1, characterized in that, The target detection module includes: The first detection layer employs a channel-pruned optimized YOLOv5-Tiny network, utilizing the NEON instruction set to compute INT8 quantized features of 3×3 depthwise separable convolutions in parallel. Second-level tracking layer: Combining the Farneback optical flow algorithm, the cross-frame trajectory persistence of moving targets is analyzed through the OpenCL kernel of ARM MaliGPU; The third prediction layer deploys a spatiotemporal graph convolutional network, which predicts the target's next activity area based on the target's position and velocity vectors output from the first two levels, and feeds back the predicted coordinates to the camera gimbal control unit.

3. The ARM-based embedded remote monitoring camera according to claim 2, characterized in that, The channel pruning optimization of the first-level detection layer is achieved through the following steps: The importance score of each output channel is dynamically calculated based on the L1 norm of the weight tensor, and channels with an importance score lower than the adaptive threshold are marked as redundant. The NEON instruction set is used to perform pruning operations on redundant channels in parallel, preserving the spatial correlation characteristics between channels; A pruning map is built in the L2 cache of the ARM processor to dynamically skip the convolution calculation of the pruned channels; The CMSIS-NN library is used to perform 8-bit fixed-point retraining on the pruned network layers to compensate for the accuracy loss.

4. The ARM-based embedded remote monitoring camera according to claim 3, characterized in that, The calculation of the INT8 quantization feature includes: Real-time statistics of the moving average and dynamic standard deviation of the input feature map of the current convolutional layer; A nonlinear quantization mapping function is constructed based on statistical values, and the eigenvalues ​​are divided into non-uniform intervals. The parallel processing capabilities of the NEON instruction set are used to perform multiplication and accumulation operations on the quantized feature map and weight matrix. A lightweight calibration network is embedded in the quantization output layer to dynamically compensate for quantization errors.

5. The ARM-based embedded remote monitoring camera according to claim 2, characterized in that, The operation of the spatiotemporal graph convolutional network includes: Construct a spatiotemporal topology graph with the target as the central node. The node attributes include position, velocity, and appearance features. The NEON instruction set is used to compute the spatiotemporal attention weights between nodes in parallel to filter key related nodes. Historical motion trajectories are fused with current observation data through a gated loop unit; Output the confidence distribution map of the target position within the next 5 frames, and generate the gimbal pre-aiming coordinates.

6. The ARM-based embedded remote monitoring camera according to claim 1, characterized in that, The multimodal sensor interface is implemented as follows: An intelligent sampling mechanism is used to increase the sampling rate of the millimeter-wave radar from 10Hz to 30Hz when the target detection module identifies a human target. The data fusion unit is used to perform spatiotemporal alignment of visual data and radar point cloud in L2 cache through the DMA controller built into the ARM processor, and generate motion trajectory with depth information. An abnormal linkage strategy is used to activate the fire feature recognition sub-model of the target detection module when the ratio of temperature change to time change detected by the temperature and humidity sensor is greater than a threshold.

7. The ARM-based embedded remote monitoring camera according to claim 1, characterized in that, The ARM-based embedded remote monitoring camera also includes: The dynamic power management module is used to receive motion intensity data output by the target detection module and allocate processor core voltage in real time through the dynamic power management bus.

8. The ARM-based embedded remote monitoring camera according to claim 7, characterized in that, The dynamic power management module includes: The load-aware unit is used to monitor the BBox generation rate of the target detection module, and triggers the DVFS strategy when the rate is <5 frames / second. A graded adjustment strategy is used to set three voltage levels according to the number of targets being detected, with the voltage increasing by one level for every three additional targets being tracked. The implementation of the hierarchical adjustment strategy includes: Establish a non-linear mapping relationship between the target quantity and the voltage level, and trigger a boost request when the target quantity reaches an integer multiple of 3; The current voltage status is broadcast to the multimodal sensor interface via the dynamic power management bus, and the sensor sampling rate is adjusted accordingly. A voltage surge suppression algorithm is embedded in the DVFS strategy to limit the single voltage adjustment amplitude to no more than 0.05V.

9. The ARM-based embedded remote monitoring camera according to claim 1, characterized in that, The secure transmission module includes: The differential encryption unit is used to encode ordinary background frames using the H.265 standard and to encrypt feature frames containing faces using the national cryptographic SM9 algorithm. A key linkage mechanism is used to dynamically generate a temporary session key through ARMTrustZone when the target detection module identifies a sensitive target; The secure storage area is used to store encrypted data and algorithm parameters in isolation within a physical partition using the TZC-400 storage controller of the ARM processor.

10. The ARM-based embedded remote monitoring camera according to claim 2, characterized in that, The NEON instruction set includes the NEONSIMD instruction set; The ARM-based embedded remote monitoring camera also includes: An adaptive focusing module is used to control the camera gimbal to perform 3x digital zoom in the prediction area based on the output coordinates of the third-level prediction layer. A multi-algorithm collaboration mechanism is used to disable the long-term prediction function of ST-GCN and switch to short-term prediction using Kalman filtering when the optical flow algorithm detects a motion speed > 5 m / s. The resource arbitration unit is used to dynamically allocate the time slice ratio of NEON units between algorithm operations and encryption operations via the CCI-550 interconnect bus of the ARM processor.

Images