A low-power, lightweight indoor people detection system and method based on infrared imaging

Abstract

This invention relates to a low-power, lightweight indoor personnel detection system and method based on infrared imaging. The invention addresses the problems of existing systems being network-dependent, having high power consumption, and being unable to operate in dark environments. The detection system achieves all-weather sensing by removing the infrared filter and combining it with an infrared supplementary lighting device. An embedded processing module deploys a lightweight YOLOv8 model reconstructed using LDS-Backbone and pruned with a Dual-Conf prediction head. Personnel detection and counting are achieved on the edge through INT8 quantization and a "no intermediate buffer single-loop aggregation" strategy. The system also includes a wireless communication module and a host computer module. The system proposes a "perception wake-up - on-demand inference - aggregation transmission" architecture, significantly reducing power consumption by utilizing a low-power shutdown mode. This invention belongs to the field of intelligent monitoring and image processing technology.

Description

Technical Field

[0001] This invention relates to a low-power, lightweight indoor personnel detection system and method, belonging to the field of intelligent monitoring and image processing technology. Background Technology

[0002] Indoor personnel detection and counting have significant application value in fields such as smart buildings, energy management, and security monitoring. Existing personnel detection solutions mainly fall into two categories: one is based on passive infrared (PIR) sensors or motion sensors. Their advantages are low cost and low power consumption, but PIR sensors can only detect general human movement and cannot accurately count the number of people; motion sensors require personnel to carry the device, limiting application scenarios and preventing seamless detection. The other category is visual detection methods based on cloud computing or high-end processors. These methods upload images to the cloud or high-performance servers for processing. While they can achieve accurate identification and counting, they have several key drawbacks: first, they rely on a network connection and cannot work when the network is down or the signal is weak; second, they require substantial hardware resources, resulting in high hardware costs; and third, continuously uploading video streams incurs significant power consumption and bandwidth costs, and poses a risk of user privacy leaks.

[0003] Furthermore, traditional cameras cannot effectively capture images at night or in low-light environments, causing the system to completely fail in the absence of visible light, making all-weather operation impossible. Therefore, designing a system that can operate independently locally, support offline work, consumes extremely low power, and accurately detect the number and time of people around the clock (including in dark environments) despite limited resources is a problem that urgently needs to be solved in this field. Summary of the Invention

[0004] To address the problems of network dependence, high power consumption, and inability to operate in dark environments in existing technologies, this invention proposes a low-power, lightweight indoor personnel detection system and method based on infrared imaging.

[0005] The technical solution adopted by the present invention to solve the above problems is as follows: The low-power lightweight indoor personnel detection system based on infrared camera of the present invention includes an image acquisition module, an embedded processing module, a wireless communication module and a host computer module; The image acquisition module is used to acquire images of the indoor environment and send the data to the embedded processing module; The embedded processing module is connected to the image acquisition module. The embedded processing module is equipped with a lightweight YOLOv8 model that has been pruned and quantized with INT8, which is used to detect and count people in the acquired images and send the result data to the wireless communication module. The wireless communication module is connected to the embedded processing module and is used to send the result data to the host computer module and send the host computer module's instructions to the embedded processing module. The host computer module establishes communication with the embedded processing module through the wireless communication module, which is used to send control commands and receive detection results and image data. The image acquisition module includes a camera whose built-in infrared filter has been removed to enable it to sense infrared light. The image acquisition module also includes a combination of an infrared fill light and a focusing lens to generate and focus infrared light to illuminate the detection area when the ambient light is insufficient.

[0006] Furthermore, the embedded processing module also runs an automatic exposure control program, which dynamically adjusts the camera's exposure parameters based on the brightness of the currently acquired image. When the exposure parameters reach the preset upper limit and the image brightness is still below the set threshold, the embedded processing module controls the infrared fill light to be turned on, enabling adaptive all-weather personnel detection under visible and infrared light conditions.

[0007] Furthermore, the embedded processing module employs a single-loop aggregation processing strategy without intermediate buffers for the preprocessing of the acquired images. The specific process includes: While traversing the RGB565 image pixels captured by the camera, pixel displacement decoding, floating-point normalization calculation, INT8 asymmetric quantization calculation using model scaling factor, and index remapping from HWC arrangement to CHW arrangement are completed synchronously within the same loop cycle. The quantized INT8 values ​​are directly written into the static input tensor area of ​​the AI ​​model through forced pointer mapping, without allocating any intermediate image cache, thereby significantly reducing the static memory overhead and peak memory usage of the system.

[0008] Furthermore, the lightweight YOLOv8 model consists of an LDS-Backbone network, an SFPN neck network, and a Dual-Conf detector head. Computational and memory overhead are reduced through the reconstruction of the network structure and output mechanism; specifically including: Step 1: Construct the LDS-Backbone network: Replace the original convolutional units of the YOLOv8 backbone network with lightweight LDB modules. Step 2, Aggressive downsampling strategy: In the LDS-Backbone network, by increasing the stride of specific convolutional layers, the resolution of the input image is rapidly reduced, and the spatial size of the feature map is reduced in advance to reduce the SRAM usage of the intermediate feature extraction process. Step 3: Construct the SFPN neck network: Abandon the complex convolutional upsampling of the traditional feature pyramid, and adopt a combination of max pooling and skip connections to directly concatenate the shallow features with the deep features after dimensionality reduction, so as to achieve multi-scale feature fusion with low parameter count. Step 4: Pruning and Reconstructing the Dual-Conf Detection Head: Remove the P5 prediction head used for large target detection in the YOLOv8 model, and retain only the P3 and P4 dual-scale prediction heads adapted to the normal distance of people indoors; in addition, the Dual-Conf detection head targets the same candidate prediction box, and simultaneously decouples the output color image feature confidence and infrared image feature confidence.

[0009] Furthermore, in step one, the LDB module includes a first 1*1 pointwise convolutional layer, a 3*3 depthwise separable convolutional layer, and a second 1*1 pointwise convolutional layer connected in sequence; wherein, the number of groups in the 3*3 depthwise separable convolutional layer is equal to the number of channels in the filter, and it is used for spatial feature extraction; the second 1*1 pointwise convolutional layer uses a linear activation function, and adds residual skip connections when the number of input and output channels is the same, in order to prevent the loss of low-dimensional features; The output mechanism of the Dual-Conf detection head described in step four is as follows: when the input image size is 160×160, the P3 and P4 dual-scale prediction heads generate 400 and 100 anchor points respectively, outputting a total of 500 fixed prediction results; each prediction result contains tensor parameters in six dimensions, namely: x-coordinate of the bounding box center, y-coordinate of the bounding box center, bounding box width w, bounding box height h, color feature confidence Conf1, and infrared feature confidence Conf2; the embedded processing module automatically extracts the maximum value of Conf1 and Conf2 as the final comprehensive confidence of the prediction box in post-processing.

[0010] Furthermore, the control commands generated by the host computer module include at least one of the following: real-time person recognition command, timed person recognition interval setting command, selected person recognition time setting command, real-time environmental photo acquisition command and corresponding person acquisition command, and historical data acquisition command.

[0011] Furthermore, the embedded processing module also includes a local storage unit, which is used to store the results of each detection locally when the system is offline; and after the detection system reconnects with the host computer module, it transmits the stored historical detection results to the host computer module, and the host computer module then calculates the historical data timestamps based on the time interval and the transmission order.

[0012] Furthermore, the embedded processing module and the host computer module use a handshake packet-splitting mechanism to transmit image data; The embedded processing module divides the image data to be transmitted into multiple data packets, each containing a frame header, frame trailer, and data content. After sending a data packet, the embedded processing module waits for a correct reception confirmation signal returned by the host computer module before sending the next data packet. If an error confirmation signal is received, the current data packet is retransmitted.

[0013] Furthermore, the detection system adopts a system-level low-power architecture of "perception wake-up-on-demand inference-aggregation transmission" and also includes a personnel perception module connected to the embedded processing module; In the timed loop detection mode, the embedded processing module enters a low-power shutdown state after completing one detection, and is woken up by the internal timer at the next detection time.

[0014] The steps of the low-power, lightweight indoor people detection method based on infrared imaging described in this invention include: Step 1: Arrange the entire detection system module on a vertical wall and align the image acquisition module with the acquisition area; Step 2: At this time, the embedded processing system is in a waiting state, and the host computer connects to the wireless communication module to send instructions; Step 3: The image acquisition module acquires images of the indoor environment; Step 4: The embedded processing module runs the automatic exposure control program, evaluates the brightness level of the current image, dynamically adjusts the exposure parameters, and then re-acquires the image; if the exposure parameters reach the upper limit but the image brightness is still insufficient, the infrared fill light is turned on. Step 5: The embedded processing module performs image processing and sends the result data to the wireless communication module; Step 5.1: The embedded processing module preprocesses the acquired image and converts it in-situ into a quantized tensor that conforms to the model input format using a zero-copy strategy; Step 5.2: The embedded processing module calls the built-in lightweight YOLOv8 model to perform inference, obtains the black-and-white confidence score and color confidence score of each prediction box, and takes the larger value as the comprehensive confidence score of the corresponding prediction box; the model has removed the prediction head used for small object detection in advance and increased the convolution stride to reduce the intermediate computation. Step 5.3: The embedded processing module performs post-processing on the detection results based on the comprehensive confidence level, including confidence level filtering and non-maximum suppression, and finally calculates the number of people. Step 5.4: The embedded processing module stores the detection results locally or sends them to the wireless communication module for uploading to the host computer module.

[0015] The beneficial effects of this invention are as follows: This invention solves the pain points of existing technologies such as reliance on networks, high power consumption, and failure in low light. It has advantages such as edge intelligence, ultra-low power consumption, and offline storage and retransmission, and is widely applicable to intelligent buildings and security monitoring. This invention gives the device night vision capabilities by removing the infrared filter of traditional cameras and combining infrared fill lights and focusing lenses; and by deploying a lightweight AI model on the embedded processing module, it realizes the completion of all image processing and personnel counting locally without network and cloud support; at the same time, by utilizing the "perception wake-up-on-demand inference-aggregation transmission" architecture and personnel perception wake-up mechanism, the system can be in an extremely low-power shutdown state, significantly extending battery life.

[0016] This invention successfully achieves network-free, latency-free, and privacy-preserving edge-side intelligent detection on resource-constrained microcontrollers by deploying a lightweight model that integrates the LDS-Backbone backbone network and Dual-Conf detection heads, coupled with a memory optimization strategy of 'no intermediate cache, single-loop aggregation'. Combined with a system-level architecture of 'perception wake-up - on-demand inference - aggregation transmission' and an offline storage and retransmission mechanism, it achieves ultra-low power consumption sufficient to support long-term operation on a single battery while ensuring functional integrity. The system has a compact structure and is easy to deploy, making it widely applicable to indoor personnel monitoring scenarios requiring privacy protection, offline statistics, and long-term independent operation. Attached Figure Description

[0017] Figure 1 This is a system block diagram of the present invention; Figure 2 This is a system module flowchart of the present invention; Figure 3 This is the schematic diagram of the system motherboard; Figure 4 This is a diagram of the improved lightweight YOLOv8 network topology. Detailed Implementation

[0018] like Figures 1 to 3 As shown, a low-power, lightweight indoor personnel detection system based on infrared cameras includes an image acquisition module, an embedded processing module, a human body sensing module, a wireless communication module, and a host computer module. The detection system can be installed on the vertical surface of the indoor cavity wall. The detection system is powered by a single lithium battery and has extremely low power consumption and independent operation capability.

[0019] The image acquisition module is used to capture images of the indoor environment. At its core is a camera whose built-in infrared filter has been physically removed, allowing its CMOS sensor to detect infrared light. Simultaneously, the module integrates an infrared fill light and a focusing lens. When ambient light is insufficient, the focusing lens concentrates the infrared light to illuminate the detection area, thus enabling the device to capture clear images in complete darkness.

[0020] The embedded processing module is the core of this system, using the STM32H743 main control chip. This module is directly connected to the image acquisition module and is responsible for executing all computational tasks. Addressing the hardware bottleneck of the microcontroller's extremely limited SRAM, the lightweight YOLOv8 model deployed within it underwent deep topology customization. like Figure 4 As shown: The backbone network adopts an LDS-Backbone architecture, using LDB modules (containing 3×3 depthwise separable convolutions) to replace traditional heavy convolutions, and with a large stride setting, the size of intermediate feature maps is rapidly reduced to prevent memory overflow; the neck fusion adopts an SFPN structure, using simplified max pooling and skip concatenation to replace RAM-intensive convolutional upsampling; the prediction end prunes the computationally intensive P5 large target prediction head and constructs a unique Dual-Conf detection head. This customized model is fully adaptable to the computing power of a single microcontroller, achieving edge intelligence without sending images to the cloud, ensuring extremely low system latency and data privacy.

[0021] The wireless communication module uses the ESP-01s WIFI module, which is connected to the embedded processing module via a serial port. It is responsible for establishing a wireless communication link between the system and the host computer module. On the one hand, it uploads the detection results (number of people) or image data to the host computer interface on the web page. On the other hand, it receives and forwards various control commands from the host computer.

[0022] The host computer module is a web-based interface accessible via Wi-Fi. Users can send commands through this interface, such as: instantly capturing and recognizing people, setting time intervals for cyclical detection, specifying a future point in time for recognition, retrieving historical data stored locally on the device, and obtaining real-time images in developer mode to verify the recognition effect.

[0023] Working principle After the detection system is powered on, the embedded processing module first enters a standby state, while its wireless communication module enters hotspot mode. Users connect to this Wi-Fi hotspot using their terminal devices and send control commands through a web-based host computer interface. The detection system supports multiple operating modes: Mode 1: Real-time photo recognition mode The host computer sends a direct image capture command, and the embedded processing module immediately exits the waiting state and calls the image acquisition module to capture the image. To ensure processing efficiency, the hardware controls the camera to crop the image to 160×160 pixels and store it in RAM. Next, the embedded processing module executes an automatic exposure control program: first, it assesses the brightness of the current image; if the brightness is too low, it dynamically increases the camera's exposure parameters until the upper limit is reached; if the image brightness is still below the set threshold at the upper exposure limit, it determines that the environment is dark, immediately turns on the infrared fill light, and re-acquires the image, ensuring that effective images can be obtained under all weather conditions.

[0024] After image acquisition, the system does not allocate a large floating-point array as an intermediate image cache. Instead, it performs "single-loop aggregation processing without intermediate cache." The underlying program starts a loop that iterates through 160×160 pixels. At the instant each RGB565 pixel is read, it simultaneously performs color channel decomposition, floating-point normalization, and INT8 asymmetric quantization combined with a scaling factor. The calculated INT8 value is directly written into the static input tensor area of ​​the AI ​​engine through address offset. This extreme optimization of "zero-copy in place" completely opens up the data flow from the camera to the AI ​​model, bringing peak memory usage close to the theoretical minimum.

[0025] After inference using the deployed lightweight model, the P3 and P4 scale branches of the Dual-Conf detection head output a fixed six-dimensional tensor (center coordinates x, y, width and height w, h, color confidence score, and infrared confidence score) for 500 candidate prediction boxes. In the post-processing stage, the system analyzes these 500 prediction boxes one by one: first, it adaptively compares the color confidence score and the infrared confidence score, extracting the maximum value as the comprehensive confidence score to eliminate feature fluctuations during day-night transitions; then, it performs initial screening using a preset confidence threshold (e.g., 0.12) to remove background noise; finally, it executes a non-maximum suppression (NMS) algorithm with an intersection-over-union (IOU) threshold of 0.2 on the remaining valid boxes. The number of final prediction boxes retained by the NMS algorithm represents the precise number of people indoors at the current moment. This result is then uploaded to the host computer for display via the WIFI module.

[0026] Mode 2: Timed Loop Photo Taking and Offline Recording Mode The user sets a time interval via the host computer. Upon receiving the instruction and interval parameters, the embedded processing module enters a low-power shutdown mode. This mode is the core element showcasing the system's low-power advantage. In this mode, the CPU stops running, and all peripherals except the internal real-time timer are powered off, reducing power consumption to the microamplitude level. When the internal timer reaches the set time interval, the system is awakened and automatically executes the photo-taking, supplementary lighting judgment, and AI recognition process as described in Mode 1. However, unlike Mode 1, the detected number of people is not immediately uploaded but is stored in the local storage unit along with the current internal counter information. After storage, the system re-enters shutdown mode, repeating this cycle to achieve uninterrupted offline monitoring with extremely low power consumption. When the user needs to retrieve data, approaching the device triggers the human detection module, which emits a high-level signal to wake the system and enter the waiting area. The user connects to Wi-Fi and clicks to retrieve historical data; the embedded processing module then uploads all stored detection records to the host computer at once. The host computer module then uses the preset time interval and retransmission order to deduce the precise timestamp of each record and finally generate a statistical chart of personnel distribution over 24 hours.

[0027] Mode 3: Select Time to Take Photos The user selects a future time on the webpage. The host computer calculates the difference between the current time and the target time and sends this delay information to the system. The system then enters a shutdown mode, only starting a timer. When the designated time arrives, the system wakes up and performs a photo recognition operation, then returns to the waiting area. At this point, by connecting to Wi-Fi, the user can view the number of people at that specific time.

[0028] Mode 4: Developer Mode and Image Transfer Entering this mode requires a password to protect privacy. In this mode, users can choose to acquire a real-time image with a detection frame. After taking a picture and completing AI recognition, the embedded processing module immediately uploads the 160×160 RGB raw image data to the host computer. To ensure the reliability of large amounts of data (approximately 150KB per frame) during wireless transmission, the system employs a handshake packet segmentation mechanism: the entire image data is divided into multiple data packets, each with a payload of 1024 bytes, identified by specific headers and trailers. The sender must wait for the receiver to return an acknowledgment signal "RCEOK" before sending the next packet; if a "NOK" signal is received, the current packet is retransmitted. This mechanism ensures the integrity of the image data and the efficiency of transmission. The host computer ultimately displays the image along with the number of detected people for developers to evaluate the recognition effect.

[0029] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention in any way. Although the present invention has been disclosed above with reference to preferred embodiments, it is not intended to limit the present invention. Any person skilled in the art can make some modifications or alterations to the above-disclosed technical content to create equivalent embodiments without departing from the scope of the present invention. Any simple modifications, equivalent substitutions, and improvements made to the above embodiments without departing from the scope of the present invention, based on the technical essence of the present invention and within the spirit and principles of the present invention, shall still fall within the protection scope of the present invention.

Claims

1. A low-power, lightweight indoor personnel detection system based on infrared imaging, characterized in that, It includes an image acquisition module, an embedded processing module, a wireless communication module, and a host computer module; The image acquisition module is used to acquire images of the indoor environment and send the data to the embedded processing module; The embedded processing module is connected to the image acquisition module. The embedded processing module is equipped with a lightweight YOLOv8 model that has been pruned and quantized with INT8, which is used to detect and count people in the acquired images and send the result data to the wireless communication module. The wireless communication module is connected to the embedded processing module and is used to send the result data to the host computer module and send the host computer module's instructions to the embedded processing module. The host computer module establishes communication with the embedded processing module through the wireless communication module, which is used to send control commands and receive detection results and image data. The image acquisition module includes a camera whose built-in infrared filter has been removed to enable it to sense infrared light. The image acquisition module also includes a combination of an infrared fill light and a focusing lens to generate and focus infrared light to illuminate the detection area when the ambient light is insufficient.

2. The low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The embedded processing module also runs an automatic exposure control program, which dynamically adjusts the camera's exposure parameters based on the brightness of the currently acquired image. When the exposure parameters reach the preset upper limit and the image brightness is still below the set threshold, the embedded processing module controls the infrared fill light to be turned on, enabling adaptive all-weather personnel detection under visible and infrared light conditions.

3. The low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The embedded processing module performs preprocessing on the acquired images using a single-loop aggregation strategy without intermediate buffers. The specific process includes: While traversing the RGB565 image pixels captured by the camera, pixel displacement decoding, floating-point normalization calculation, INT8 asymmetric quantization calculation using model scaling factor, and index remapping from HWC arrangement to CHW arrangement are completed synchronously within the same loop cycle. The quantized INT8 values ​​are directly written into the static input tensor area of ​​the AI ​​model through forced pointer mapping, without allocating any intermediate image cache, thereby significantly reducing the static memory overhead and peak memory usage of the system.

4. The low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The lightweight YOLOv8 model consists of an LDS-Backbone network, an SFPN neck network, and a Dual-Conf detector head. Computational and memory overhead is reduced through reconstruction of the network structure and output mechanism. Specifically, it includes: Step 1: Construct the LDS-Backbone network: Replace the original convolutional units of the YOLOv8 backbone network with lightweight LDB modules. Step 2, Aggressive downsampling strategy: In the LDS-Backbone network, by increasing the stride of specific convolutional layers, the resolution of the input image is rapidly reduced, and the spatial size of the feature map is reduced in advance to reduce the SRAM usage of the intermediate feature extraction process. Step 3: Construct the SFPN neck network: Abandon the complex convolutional upsampling of the traditional feature pyramid, and adopt a combination of max pooling and skip connections to directly concatenate the shallow features with the deep features after dimensionality reduction, so as to achieve multi-scale feature fusion with low parameter count. Step 4: Pruning and Reconstructing the Dual-Conf Detection Head: Remove the P5 prediction head used for large target detection in the YOLOv8 model, and retain only the P3 and P4 dual-scale prediction heads adapted to the normal distance of people indoors; in addition, the Dual-Conf detection head targets the same candidate prediction box, and simultaneously decouples the output color image feature confidence and infrared image feature confidence.

5. A low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 4, characterized in that, In step one, the LDB module includes a first 1*1 pointwise convolutional layer, a 3*3 depthwise separable convolutional layer, and a second 1*1 pointwise convolutional layer connected in sequence. The number of groups in the 3*3 depthwise separable convolutional layer is equal to the number of channels in the filter, and it is used for spatial feature extraction. The second 1*1 pointwise convolutional layer uses a linear activation function and adds residual skip connections when the number of input and output channels is the same to prevent the loss of low-dimensional features. The output mechanism of the Dual-Conf detection head described in step four is as follows: when the input image size is 160×160, the P3 and P4 dual-scale prediction heads generate 400 and 100 anchor points respectively, outputting a total of 500 fixed prediction results; each prediction result contains tensor parameters in six dimensions, namely: x-coordinate of the bounding box center, y-coordinate of the bounding box center, bounding box width w, bounding box height h, color feature confidence Conf1, and infrared feature confidence Conf2; the embedded processing module automatically extracts the maximum value of Conf1 and Conf2 as the final comprehensive confidence of the prediction box in post-processing.

6. The low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The control commands generated by the host computer module include at least one of the following: real-time person recognition command, timed person recognition interval setting command, selected person recognition time setting command, real-time environmental photo acquisition command and corresponding person acquisition command, and historical data acquisition command.

7. A low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The embedded processing module also includes a local storage unit, which is used to store the results of each detection locally when the system is offline; and after the detection system reconnects with the host computer module, it will retransmit the stored historical detection results to the host computer module, and the host computer module will then calculate the historical data timestamps based on the time interval and the retransmission order.

8. A low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The embedded processing module and the host computer module use a handshake packet-splitting mechanism to transmit image data; The embedded processing module divides the image data to be transmitted into multiple data packets, each containing a frame header, frame trailer, and data content; After sending a data packet, the embedded processing module must wait for a correct reception confirmation signal returned by the host computer module before sending the next data packet; if an error confirmation signal is received, the current data packet will be retransmitted.

9. A low-power, lightweight indoor personnel detection system based on infrared imaging according to claim 1, characterized in that, The detection system adopts a system-level low-power architecture of "perception wake-up-on-demand inference-aggregation transmission" and also includes a personnel perception module connected to the embedded processing module; In the timed loop detection mode, the embedded processing module enters a low-power shutdown state after completing one detection, and is woken up by the internal timer at the next detection time.

10. A detection method based on the detection system according to any one of claims 1 to 9, characterized in that, The specific steps include: Step 1: Arrange the entire detection system module on a vertical wall and align the image acquisition module with the acquisition area; Step 2: At this time, the embedded processing system is in a waiting state, and the host computer connects to the wireless communication module to send instructions; Step 3: The image acquisition module acquires images of the indoor environment; Step 4: The embedded processing module runs the automatic exposure control program, evaluates the brightness level of the current image, dynamically adjusts the exposure parameters, and then re-acquires the image; if the exposure parameters reach the upper limit but the image brightness is still insufficient, the infrared fill light is turned on. Step 5: The embedded processing module performs image processing and sends the result data to the wireless communication module; Step 5.1: The embedded processing module preprocesses the acquired image and converts it in-situ into a quantized tensor that conforms to the model input format using a zero-copy strategy; Step 5.2: The embedded processing module calls the built-in lightweight YOLOv8 model to perform inference, obtains the black-and-white confidence score and color confidence score of each prediction box, and takes the larger value as the comprehensive confidence score of the corresponding prediction box; the model has removed the prediction head used for small object detection in advance and increased the convolution stride to reduce the intermediate computation. Step 5.3: The embedded processing module performs post-processing on the detection results based on the comprehensive confidence level, including confidence level filtering and non-maximum suppression, and finally calculates the number of people. Step 5.4: The embedded processing module stores the detection results locally or sends them to the wireless communication module for uploading to the host computer module.

Images