Solar power supply edge computing monitoring device for highway equipment

Abstract

The application belongs to the technical field of monitoring, and particularly relates to a highway equipment solar power supply edge computing monitoring device, which comprises a solar cell panel, a power management module and a storage battery, the solar cell panel supplies power to the storage battery and an edge computing load through the power management module; a high-definition camera is used for collecting a highway construction area monitoring picture; an edge computing unit is configured to perform image preprocessing on the monitoring picture and generate a border crossing event element and a parking violation event element based on a multi-outlet Anchor-free target detection network; a 4G or 5G wireless communication module is used for sending comprehensive alarm data obtained by encapsulating the border crossing event element and the parking violation event element to a management center platform; and a local alarm unit is connected with the edge computing unit and is used for performing sound and light alarm based on the comprehensive alarm data. Through the edge computing mechanism capable of adaptively selecting a reasoning outlet, the detection speed and the recognition stability under a complex environment can be improved.

Description

Technical Field

[0001] This invention belongs to the field of monitoring technology, specifically relating to a solar-powered edge computing monitoring device for highway equipment. Background Technology

[0002] Safety monitoring of highway construction areas has long been a crucial issue in traffic management. With increasing traffic volume and frequent construction activities, effectively supervising construction areas and promptly detecting abnormal behavior without disrupting normal traffic flow has become a key technological focus within the industry. Existing publicly available solutions primarily employ manual patrols, fixed camera monitoring, and cloud-based video analytics for construction area safety management. However, these methods have revealed several shortcomings in practice, failing to meet the comprehensive demands of highway construction scenarios for real-time performance, accuracy, and energy self-sufficiency.

[0003] In traditional manual patrol methods, patrol personnel typically inspect construction areas along fixed routes and at fixed intervals. While this method is intuitive, it relies heavily on manpower, and the frequency of patrols is limited by the number of personnel and the working environment, making it impossible to achieve continuous monitoring of the construction area and easily creating blind spots. Furthermore, manual patrols are even less efficient at night, in inclement weather, or during periods of heavy traffic, and their responsiveness and continuity are difficult to guarantee, failing to meet the immediate needs for handling rapidly evolving incidents such as boundary violations and illegal parking.

[0004] With the widespread adoption of video surveillance equipment, fixed cameras have become a primary supplementary means of monitoring highway construction. Typically, these cameras transmit images to a backend server for processing via wired power and network transmission. In existing publicly available technologies, video streams are generally transmitted to a central server via broadband networks or dedicated lines, where the server performs calculations such as behavior recognition and anomaly detection. However, highway construction areas are often located in remote areas, making the deployment of power and network lines costly, difficult, and equipment maintenance complex. Furthermore, the continuous transmission of high-resolution video places a significant burden on bandwidth, and network fluctuations can lead to increased transmission latency or even interruptions, preventing the recognition system from processing events in real time. Summary of the Invention

[0005] Therefore, the main objective of this invention is to provide a solar-powered edge computing monitoring device for highway equipment. This device integrates a high-definition camera, a solar power module, an edge computing unit, a wireless communication module, and a local alarm unit. It deploys a multi-exit network and an anchor-free target detection algorithm within the device to achieve real-time identification of vehicles, personnel, cones, and construction equipment within the construction area. Based on trajectory analysis and area masking, it accurately determines illegal parking, boundary crossings, and abnormal traffic events. This invention improves detection speed and identification stability in complex environments through an edge computing mechanism that can adaptively select inference exits. It also utilizes solar energy for long-term self-powered operation and employs event-driven communication to significantly reduce bandwidth usage. Thus, it possesses comprehensive advantages including strong real-time performance, high identification accuracy, easy deployment, low energy consumption, strong adaptability, and low maintenance costs.

[0006] The technical solution adopted in this invention is as follows:

[0007] A solar-powered edge computing monitoring device for highway equipment includes: a solar panel, a power management module, and a battery; the solar panel supplies power to the battery and edge computing load via the power management module; a high-definition camera for capturing monitoring images of the highway construction area; an edge computing unit electrically connected to the power management module and the high-definition camera, configured to preprocess the monitoring images and generate boundary crossing event elements and illegal parking event elements based on a multi-exit anchor-free target detection network; a 4G or 5G wireless communication module connected to the edge computing unit for sending comprehensive alarm data encapsulated based on boundary crossing event elements and illegal parking event elements to a management center platform; and a local alarm unit connected to the edge computing unit for generating comprehensive alarm data based on the comprehensive alarm data. The system uses alarm data for audio-visual warnings. A multi-exit anchor-free target detection network is integrated into the edge computing unit, including a shared backbone structure, a central candidate exit, a size estimation exit, a category determination exit, and a region segmentation exit. The shared backbone structure takes a top-view equivalent image as input, sequentially sets convolutional and pooling layers to extract texture, structural, and semantic features, and outputs a first feature map, a second feature map, and a third feature map at three different spatial scales. During design, the shared backbone structure establishes the lane's main direction axis based on road markings. In convolution operations, it expands the sensing area along the lane's main direction axis, enabling vehicles, cones, and protective facilities arranged along the lane direction to form continuous responses on the first, second, and third feature maps, providing a unified spatial feature basis for each exit.

[0008] Furthermore, the power management module is electrically connected to the solar panel, battery, and edge computing unit respectively. The power management module collects the output voltage and current of the solar panel and the remaining power of the battery, divides the collected results into multiple discrete energy state levels according to segmentation rules, and sends them to the edge computing unit. The edge computing unit selects between a high-performance working mode and an energy-saving working mode according to the energy state level. In the high-performance working mode, the monitoring screen is collected at the preset highest resolution and highest frame rate and all configured sensors are enabled. In the energy-saving working mode, the acquisition resolution and frame rate of the monitoring screen are reduced and some non-critical sensors are turned off, thereby realizing adaptive control of the overall power consumption of the solar-powered edge computing monitoring device for highway equipment under different energy state levels.

[0009] Furthermore, the system includes millimeter-wave radar and environmental sensors, both of which are electrically connected to the edge computing unit. The edge computing unit controls a high-definition camera to capture monitoring images and performs lens distortion correction, perspective transformation based on road markings, block brightness equalization, and color deviation compensation to obtain the original viewpoint image. Based on the perspective transformation relationship, the original viewpoint image is mapped to a top-view equivalent image with the road surface as a reference. When configuring the millimeter-wave radar, a series of distance points corresponding to the current monitoring image are collected. The distance points are mapped to the top-view equivalent image coordinate system through a calibration matrix, and each distance point is assigned a top-view coordinate. The environmental sensors output information on illumination, fog, precipitation, and wind speed. The edge computing unit associates the environmental information with the original viewpoint image and the top-view equivalent image at the corresponding time, which is used as the environmental state when generating event elements.

[0010] Furthermore, the candidate center exits are connected to the first feature map, the second feature map, and the third feature map respectively, generating a center response map on each scale feature map. Each pixel position in the center response map corresponds to a position in the top-view equivalent image. The edge computing unit selects pixel positions with response values ​​greater than a preset response threshold as candidate center points in the center response map and assigns them top-view coordinates to form a single-scale candidate center set. Each single-scale candidate center set is sorted in descending order of response value, and only the candidate center with the largest response value is retained in each preset neighborhood window to obtain a single-scale center set. The single-scale center sets are merged into a full-scale candidate set, and candidate points are selected as priority centers in descending order of response value. Other candidate points with a distance less than a preset distance threshold from the priority centers are found in the full-scale candidate set. These candidate points are grouped into the same target center group, and only the candidate point with the largest response value in each target center group is retained as the target center to form a full-scale target center set.

[0011] Furthermore, the size estimation output is connected to the first feature map, the second feature map, and the third feature map, respectively, and outputs a size response map containing four channels on each scale feature map. The four channels represent the estimated pixel distance values ​​from the target center to the forward, backward, left, and right edges in the top-view equivalent image, respectively. For each target center in the full-scale target center set, the edge computing unit reads the four pixel distance estimates at the target center position in the size response map of the corresponding scale, and expands the rectangular circumscribed range along the forward, backward, left, and right directions on the top-view equivalent image based on the pixel distance estimates, with the target center as the geometric center. The rectangular circumscribed range covers the outline of vehicles, personnel, cones, construction equipment, or protective facilities. When the rectangular circumscribed range exceeds the boundary of the top-view equivalent image, the edge computing unit clips the rectangular circumscribed range according to the boundary of the top-view equivalent image, so that the rectangular circumscribed range is completely located inside the top-view equivalent image.

[0012] Furthermore, the category determination output is connected to the first feature map, the second feature map, and the third feature map respectively, and outputs a multi-channel category response map on each scale feature map. Each channel corresponds to the categories of vehicle, personnel, cone, construction equipment, and protective facilities, respectively. For each target center in the full-scale target center set, the edge computing unit reads the response value of each channel at the target center location in the category response map of the corresponding scale, and determines the category corresponding to the channel with the largest response value as the target category. When the difference between the largest channel response value and the second largest channel response value is less than a preset difference threshold, the edge computing unit reads the aspect ratio of the rectangle bounding area corresponding to the target center. When the aspect ratio is greater than the first aspect ratio threshold, the target category is adjusted to vehicle. When the aspect ratio is within the range of the second aspect ratio threshold and the area of ​​the rectangle bounding area is within the range of the cone area threshold, the target category is adjusted to cone. The edge computing unit combines the target center, the rectangle bounding area, and the target category to form a structured target set for boundary construction and event element generation.

[0013] Furthermore, the region segmentation exit is connected to the highest-scale feature map of the shared backbone structure to generate a region label map with the same size as the top-view equivalent image. Each pixel in the region label map is labeled as a construction area, lane area, or buffer area. The edge computing unit performs connected component analysis on the region label map, merging connected components of regions with areas smaller than a first area threshold into adjacent regions of the same type with areas not smaller than a second area threshold. Dilation and erosion operations are sequentially performed at the boundaries of the construction area and the buffer area to form continuous closed contour lines, generating a construction area mask and a buffer area mask. The edge calculation unit selects targets categorized as cones and protective facilities from the structured target set, sorts the target center coordinates according to the main direction of the lane, and divides targets on one side of the lane centerline into a left-side marker chain and targets on the other side into a right-side marker chain, using the lane centerline as a reference. In each marker chain, the centers of two adjacent targets are connected to form a broken line sequence. Curve fitting is performed on the broken line sequence to obtain a continuous boundary curve. A strip warning zone is formed on both sides of the boundary curve by buffering and expanding according to a fixed pixel width. The strip warning zone is aligned with the outer boundary of the construction area mask to obtain a dynamic virtual warning boundary covering the edge of the construction area.

[0014] Furthermore, the edge computing unit generates boundary crossing event elements based on the dynamic virtual warning boundary: it selects targets categorized as personnel and vehicles from the structured target set, reads the coordinates of the target center in the top-view equivalent image, and queries the region type corresponding to the coordinates in the region label map; when the region type is a construction area or a buffer area, it calculates the minimum pixel distance from the target center to the inner and outer edge lines of the dynamic virtual warning boundary, and sets a safe distance threshold for determining boundary crossing based on the illumination and fog values ​​given by the environmental sensors; when the illumination value is lower than the illumination threshold or the fog value is higher than the fog threshold, the safe distance threshold is increased; when the minimum pixel distance is less than the safe distance threshold and the target center is located on one side of the construction area or within the buffer area, the target is determined to be a boundary crossing target, and boundary crossing event elements are generated. The boundary crossing event elements include at least the target category, target center coordinates, rectangular bounding range, dynamic virtual warning boundary index, and current environmental state.

[0015] Furthermore, the process by which the edge computing unit generates illegal parking event elements based on evidence of stationary vehicles and encapsulates comprehensive alarm data specifically includes: for targets located within the lane area mask and whose target category is vehicles, constructing wheel texture evidence, ground shadow evidence, and contour clarity evidence within the corresponding rectangular bounding area; dividing the lower part of the vehicle's rectangular bounding area into four wheel candidate sub-regions according to proportions, performing edge detection in each wheel candidate sub-region and counting the number of edge pixels and fine texture lines, and establishing wheel texture evidence when the number of edge pixels and fine texture lines in at least two wheel candidate sub-regions are not less than the corresponding thresholds; within the narrow band area extending downwards from the bottom edge of the vehicle's rectangular bounding area, converting the narrow band area into a grayscale image and dividing it into dark and bright regions according to a grayscale threshold, performing connected component detection, and when a connection is formed near the bottom of the vehicle... When vehicles are continuously distributed laterally and their width falls within a preset width range of dark connected bands, grounding shadow evidence is established. Edge detection is performed and the edge direction distribution is statistically analyzed within the outer band area extending around the vehicle's rectangular outer perimeter. When the edge direction is within a preset dispersion range and the vehicle's outline is continuous within the outer band area, outline clarity evidence is established. When the number of established evidence reaches the evidence quantity threshold, the target vehicle is identified as an illegally parked vehicle, and illegal parking event elements are generated. Illegal parking event elements include at least the vehicle target center coordinates, the rectangular outer perimeter, a list of evidence types, and the lane number. When millimeter-wave radar is configured, the edge computing unit searches for distance point columns within the top-view equivalent image coordinate range near the rectangular outer perimeter of the illegally parked vehicle. When a group of distance points with a number of points not less than the point quantity threshold is detected in multiple consecutive sampling periods, a millimeter-wave radar consistency marker is added to the illegal parking event elements.

[0016] By adopting the above technical solutions, this invention achieves the following beneficial effects: First, by deploying the complex target detection and event recognition process within an edge computing unit, and combining a multi-exit network with an anchor-free detection algorithm, the device can select inference exits of different depths based on scene complexity and power supply status. This effectively reduces unnecessary computation while maintaining recognition accuracy, achieving fast, stable, and adaptive intelligent monitoring. This mechanism avoids the inherent delay caused by a single inference path in traditional detection processes, enabling events such as boundary crossings, illegal parking, and abnormal passage to be identified in a very short time, improving on-site response speed and reducing safety hazards. Second, this invention uses solar panels for energy replenishment, coupled with the power management module's adaptive power adjustment capability, allowing the device to operate for extended periods in highway environments without external power supplies, reducing wiring costs and maintenance workload while ensuring the continuity of monitoring functions. Third, this invention employs an event-driven communication method, uploading alarm data to the management center only when abnormal events are identified. This significantly reduces the bandwidth consumption of the wireless network, improves availability in areas with unstable network coverage, and thus enhances the overall reliability of the system. Furthermore, this invention achieves accurate differentiation of different types of events through rules such as construction area masking, non-parking area masking, trajectory analysis, and pixel-level position determination, effectively reducing false alarms and missed alarms. It also enhances the robustness of detection results in complex boundary areas through a cross-category interference elimination mechanism. Finally, the invention features a highly integrated overall structure and flexible installation methods, enabling rapid deployment at guardrails, streetlights, or temporary supports. It is suitable for various construction scenarios, reducing implementation difficulty and increasing monitoring coverage, providing continuous, green, real-time, and intelligent technical support for highway construction safety. Attached Figure Description

[0017] Figure 1 A comparative schematic diagram illustrating the principle of asymmetric convolution kernel receptive field expansion provided in an embodiment of the present invention;

[0018] Figure 2 This is a schematic diagram illustrating the multi-scale feature extraction and center response generation process of the multi-exit frameless target detection network provided in an embodiment of the present invention.

[0019] Figure 3 A schematic diagram illustrating the system power consumption distribution of the solar-powered edge computing monitoring device for highway equipment under different operating modes, provided in an embodiment of the present invention. Detailed Implementation

[0020] All features disclosed in this specification, or all steps in all disclosed methods or processes, may be combined in any way, except for mutually exclusive features and / or steps.

[0021] Any feature disclosed in this specification (including any appended claims and abstract) may be replaced by other equivalent or similar features, unless specifically stated otherwise. That is, unless specifically stated otherwise, each feature is merely one example of a series of equivalent or similar features.

[0022] Example 1: In one embodiment, a solar-powered edge computing monitoring device for highway equipment is installed on a streetlight pole near the entrance of the highway construction area, at a height of 6 meters. The angle between the top view and the road surface normal is approximately 35 degrees, covering a road surface area of ​​80 meters in front of the construction area and two lanes on each side. The entire device includes a solar panel, a power management module, a battery, a high-definition camera, a millimeter-wave radar, an environmental sensor, an edge computing unit, a 4G or 5G wireless communication module, a local alarm unit, and a management center platform for centralized management. All components are electrically and signal-connected according to the connection relationships described in the preceding claims.

[0023] The solar panels utilize monocrystalline silicon modules with a peak power of 200 watts. They are installed facing due south at an angle of approximately 30 degrees to accommodate varying levels of sunlight in both summer and winter. The solar panel output, after passing through the power management module's buck-boost and charging management circuits, charges a 24V, 100Ah battery, while simultaneously powering the edge computing unit and other electrical components. The power management module continuously monitors the solar panel's output voltage and current, as well as the battery's terminal voltage and charging / discharging current. Based on preset thresholds, it categorizes the current state into four energy status levels: Level 3 indicates a battery charge above 80% and a solar panel output power above 150W; Level 2 indicates a battery charge between 50% and 80%; Level 1 indicates a battery charge between 30% and 50%; and Level 0 indicates a battery charge below 30%. Every 60 seconds, the power management module transmits the current energy status level to the edge computing unit via wired serial communication. Simultaneously, it disconnects non-critical loads when the battery voltage approaches the over-discharge protection threshold, ensuring the device can maintain critical functions even during prolonged periods of cloudy or rainy weather.

[0024] The edge computing unit employs a low-power AI processor platform, configured with a multi-core general-purpose processor and a dedicated neural network inference unit, with a typical power consumption of 15 watts. An embedded operating system runs on the edge computing unit. Upon system startup, it receives energy status level information and selects the operating mode based on the level. When the energy status level is 2 or 3, it enters a high-performance operating mode. The high-definition camera continuously outputs monitoring images at a resolution of 1920×1080 and a frequency of 10 frames per second. The millimeter-wave radar and environmental sensors are all operational, and the 4G or 5G wireless communication module is always available. When the energy status level is 0 or 1, it enters an energy-saving operating mode. The high-definition camera reduces its resolution to 1280×720 and operates at 5 frames per second. The millimeter-wave radar is only temporarily activated when a suspected event occurs. The sampling interval of the environmental sensor is adjusted from 60 seconds to 300 seconds. The 4G or 5G wireless communication module adopts a method of merging multiple comprehensive alarm data points before centralized reporting to reduce power consumption. This method of switching operating modes based on energy status level enables the device to have high monitoring accuracy and response speed in sunny weather, and can still maintain monitoring of critical events without battery over-discharge even during several consecutive days of cloudy and rainy weather.

[0025] The high-definition camera is an industrial-grade wide dynamic range network camera, employing a global shutter with an adjustable single-frame exposure time ranging from 1 millisecond to 10 milliseconds and a dynamic range of no less than 120 dB. The camera lens has a 6mm focal length and an approximately 80-degree field of view, ensuring coverage of the construction area and adjacent lanes at an installation height of 6 meters. The camera connects to the edge computing unit via Ethernet, using a low-latency encoding format for transmission. The millimeter-wave radar operates in the 77 GHz band, with a detection range of up to 120 meters, primarily used to supplement vehicle distance information in dense fog or at night. Environmental sensors include an illuminance sensor, a scattering-based fog sensor, a tipping bucket rain gauge, and an anemometer, providing current light intensity, visibility estimates, precipitation data, and wind speed data.

[0026] In high-performance mode, the edge computing unit receives a monitoring frame from the high-definition camera at fixed time intervals, using this frame as the current monitoring frame. The edge computing unit first performs lens distortion correction on the current monitoring frame. After factory installation, a lookup table is obtained through a calibration process. This lookup table assigns the original pixel position to each output pixel. During runtime, resampling is performed directly according to the lookup table, completing distortion correction in less than 10 milliseconds per frame. Next, the edge computing unit uses key points of road markings obtained from on-site calibration before construction to perform perspective transformation on the current monitoring frame, converting the original view image into a top-down equivalent image. Geometrically, the top-down equivalent image transforms the lane lines that converge under perspective projection into approximately parallel straight lines. Simultaneously, it maintains a relatively consistent ratio between pixel spacing and actual distance along the lane direction throughout the entire image. This allows for direct estimation of the distance difference on the actual road surface based on pixel distance during feature extraction and boundary judgment, avoiding repeated complex spatial reconstruction and improving operational efficiency.

[0027] To adapt to various environments, including strong daylight, streetlights at night, and rainy weather, the edge computing unit divides the corrected original viewpoint image into 8×8 sub-blocks. A brightness histogram is calculated for each sub-block, and the brightness of the sub-block is linearly stretched based on its brightness distribution to ensure sufficient contrast in both dark and bright areas of the monitoring frame. Color deviation compensation is then performed. Gain coefficients for the red, green, and blue color channels are obtained from a pre-acquired standard grayscale image. During actual operation, these gains are applied to each frame to ensure that the colors of images acquired under different weather conditions are close to a uniform standard. After these processes, a stable original viewpoint image and a corresponding top-down equivalent image are obtained. These two images are linked by timestamps and frame numbers to facilitate subsequent analysis of results in the top-down equivalent image and extraction of image segments from the original viewpoint image.

[0028] In scenarios where millimeter-wave radar is installed, the edge computing unit reads a range point column from the millimeter-wave radar while receiving the current monitoring frame. Each point in the range point column contains an azimuth angle and a distance value. Through spatial calibration uniformly completed during the equipment installation phase, the range points in the radar coordinate system can be projected onto the camera coordinate system, and then onto the top-view equivalent image, giving each range point a precise location on the top-view equivalent image. The radar range points can provide stable target distance information in rainy or foggy weather, and even if the image contrast is reduced, they can still help determine whether vehicle targets in a structured target set actually exist, thus being used for millimeter-wave radar consistency marking in subsequent embodiments.

[0029] In this embodiment, the multi-exit anchor-free target detection network is deployed within the edge computing unit, with the input being the top-view equivalent image of the current frame. In high-performance mode, the edge computing unit performs inference once for each top-view equivalent image. The shared backbone structure outputs a first feature map, a second feature map, and a third feature map at three scales through successive convolution and pooling operations. The resolution of the second feature map is typically one-quarter of that of the top-view equivalent image, the first feature map has a higher resolution, and the third feature map has a lower resolution. This arrangement aims to make the first feature map more likely to retain detailed information about small targets such as cones, the third feature map more suitable for representing large targets such as construction equipment and large vehicles, and the second feature map used to accommodate ordinary vehicles and personnel targets.

[0030] In the shared backbone structure, the convolutional kernel is designed to have a longer dimension along the main lane direction; for example, a kernel of length 7 is used in the lane direction, while a kernel of length 3 is used in the lateral direction. This asymmetric kernel allows for a larger receptive area in the lane direction, which is beneficial for extracting texture features of continuous traffic flow and continuous cone arrangements, thus forming continuous response stripes along the lane direction on the first, second, and third feature maps. Since vehicles, cones, and protective facilities are typically arranged along the lane direction, extending the receptive area in the lane direction can improve the detection stability of these targets without significantly increasing the computational cost.

[0031] refer to Figure 1 , Figure 1 A comparative schematic diagram is shown illustrating the principle of receptive field expansion of asymmetric convolution kernels. Figure 1The diagram comprises a left and a right section. The left section demonstrates the receptive field characteristics of a traditional symmetric convolution kernel, while the right section shows the expansion effect of the asymmetric convolution kernel used in this invention in the lane direction. In the left section, the feature map is presented in a grid format, with the horizontal axis representing the lateral coordinates and the vertical axis representing the lane direction coordinates, ranging from 0 to 10 pixels. At the center of the feature map, at coordinates 5,5, a center pixel C is marked, represented by a black square. Around the center pixel, within a 3×3 neighborhood, are distributed eight gray squares, which together constitute the receptive field of the traditional symmetric convolution kernel. This receptive field is 3 pixels in both the lateral and lane directions, exhibiting a symmetrical distribution. Six circular markers are arranged vertically along the lane direction on the feature map, located at coordinates 5,3 to 5,8. These circular markers represent vehicles continuously distributed along the lane direction. The vehicle at the center is marked with a solid black circle, while the other vehicles are marked with hollow white circles. An upward-pointing arrow indicates the lane direction, with the text "Lane Direction" labeled next to the arrow. Below the receptive field, horizontal lines and endpoint markers indicate that the receptive field spans 3 pixels laterally. In the right-hand section, the feature map layout is the same as on the left, but the shape of the convolutional kernel changes significantly. The center pixel C remains at coordinates 5,5, but its receptive field expands to 7 pixels in the lane direction, while remaining 3 pixels horizontally. Specifically, the receptive field extends from coordinates 5,2 to 5,8 in the lane direction, covering a significantly larger area than the symmetrical convolutional kernel on the left. Within this expanded receptive field, more gray squares are distributed, arranged in elongated strips. The number of vehicle targets along the lane direction increases to 9, extending from coordinates 5,1 to 5,9, with the center vehicle located at 5,5 and marked as a solid black circle. The figure also includes upward-pointing arrows and the text "Lane Direction." To the right of the receptive field, vertical lines and endpoint markers indicate that the receptive field spans 7 pixels in the lane direction, while horizontal lines below indicate that the horizontal span remains 3 pixels. The upper right area of ​​the right-hand image is marked with a double-headed arrow connecting the top and bottom of the vehicle sequence. The text box next to the arrow is labeled "continuous response stripes," indicating that the expanded convolutional kernel can form continuous feature responses along the lane direction. A comparison of the left and right sides clearly shows that the asymmetric convolutional kernel, by increasing the receptive field length along the main lane direction, can capture more target information arranged along the lane in a single convolution operation, thus generating response strips continuously distributed along the lane direction on the feature map. This design significantly improves the detection stability for scenarios such as continuous traffic flow and continuous cone arrays, while avoiding excessive expansion of the lateral receptive field and maintaining the ability to distinguish between lanes.

[0032] The center candidate exit, size estimation exit, and category determination exit simultaneously receive the first feature map, the second feature map, and the third feature map, while the region segmentation exit receives the third feature map. In this embodiment, the edge computing unit generates a center response map on each scale feature map through the center candidate exit, obtains several center candidate points through simple thresholding and local maximum search, and forms a full-scale target center set through merging and distance clustering at the three scales. The size estimation exit provides the forward, backward, left, and right range lengths of the rectangular bounding area on the top-view equivalent image at each center location, enabling the edge computing unit to quickly construct the rectangular bounding area. The category determination exit provides response values ​​for five categories—vehicles, personnel, cones, construction equipment, and protective facilities—for each target center through multi-channel output. The edge computing unit corrects the category based on the maximum response value and the aspect ratio of the rectangular bounding area, thereby forming a structured target set. The region segmentation exit generates a region label map, marking each pixel as a construction area, lane area, or buffer area. The edge computing unit performs connected component analysis and morphological operations on the region label map to remove isolated regions with an area of ​​less than 0.5 square meters, resulting in a construction area mask and a buffer area mask.

[0033] In this embodiment, the dynamic virtual warning boundary is automatically constructed based on a structured target set and a construction area mask. The edge computing unit selects targets categorized as cones and protective facilities from the structured target set, projects the corresponding target center coordinates onto the lane's main direction axis, and sorts them according to the lane's main direction coordinates. The target sequences near the left side of the lane and those near the right side of the lane are connected by polylines and fitted with curves to generate left and right boundary curves, respectively. Subsequently, a strip-shaped warning zone is formed on both sides of each boundary curve by a fixed pixel width, shifting towards the construction area so that the outer boundary of the warning zone coincides with the outer boundary of the construction area mask. This dynamic virtual warning boundary can automatically adjust as the construction area layout changes. When the construction unit adjusts the cone positions, the device can update the boundary crossing judgment boundary without manual recalibration.

[0034] The generation of boundary violation event elements is based on dynamic virtual warning boundaries and personnel and vehicle targets in a structured target set. The edge computing unit queries the region label map for the region type corresponding to the target center position of each personnel and vehicle target. When the region type is a construction area or a buffer area, it further calculates the minimum pixel distance from the target center to the inner and outer boundaries of the dynamic virtual warning boundary. Illumination and fog values ​​are provided by environmental sensors. When the illumination value is higher than a preset illumination threshold and the fog value is lower than a preset fog threshold, the safety distance threshold can be set to 5 pixels; when the illumination value is lower than the illumination threshold or the fog value is higher than the fog threshold, the safety distance threshold is increased to 8 pixels. This setting is because the perception ability of drivers and construction workers is weakened under low light or high fog conditions, and reserving a larger safety distance is conducive to issuing warnings in advance and improving the overall safety margin. When the minimum pixel distance is less than the safe distance threshold and the target center is located on one side of the construction area or within the buffer area, the edge computing unit marks the target as an outbound target and generates outbound event elements. The outbound event elements include the target category, target center coordinates, rectangular circumscribed range, dynamic virtual warning boundary index, and current environmental status.

[0035] In this embodiment, the generation of illegal parking event elements adopts a judgment method based on static evidence. The edge computing unit only analyzes vehicle targets in the structured target set that are classified as vehicles and located within the lane area mask. For each vehicle target, the edge computing unit divides the lower part of the vehicle's rectangular outer boundary into four wheel candidate sub-regions. Edge detection is performed on each sub-region, and the number of edge pixels and fine texture lines are counted. When the number of edge pixels and fine texture lines in at least two wheel candidate sub-regions are not less than a preset threshold, the wheel texture features are considered obvious, and the vehicle tire outline is clear. Subsequently, grayscale threshold segmentation and connected component detection are performed on the narrow band area extending downward from the bottom edge of the vehicle's rectangular outer boundary. When a dark connected band with a width within a preset range is detected that spans at least two-thirds of the vehicle width laterally, a stable ground shadow is considered to have formed between the vehicle and the road surface. The edge direction distribution is then counted in the outer band area around the rectangular outer boundary. When the edge direction is within a preset dispersion range and the outline is continuous, the outline clarity is considered to meet the requirements. When at least two of the above three pieces of evidence are valid, the edge computing unit determines the vehicle target as a suspected illegally parked vehicle and generates illegal parking event elements. Unlike time-accumulated parking determination, this method analyzes tire texture, ground shadow, and contour stability in a single frame to determine whether the vehicle is in a stable stopped state. This method is suitable for real-time edge computing scenarios that do not rely on historical frame data.

[0036] After generating boundary crossing and illegal parking event elements, the edge computing unit extracts image segments from the original viewpoint image according to the corresponding rectangular bounding range, using these segments as alarm screenshots. These screenshots are then appended with timestamps, geographic location, target category, event type, dynamic virtual warning boundary index, and environmental status, packaging them into comprehensive alarm data. This comprehensive alarm data is sent to the local alarm unit. Upon detecting a boundary crossing event, the local alarm unit plays a preset voice prompt to the construction area via a directional speaker; upon detecting an illegal parking event, it alerts passing vehicles with a high-frequency short sound and flashing lights. Simultaneously, it is reported to the management center platform via a 4G or 5G wireless communication module. The management center platform displays, statistically analyzes, and stores the comprehensive alarm data.

[0037] In an optional embodiment, millimeter-wave radar can be omitted, retaining only high-definition cameras and environmental sensors. In this configuration, the multi-exit anchor-free target detection network still generates structured target sets, region label maps, and dynamic virtual warning boundaries on the top-view equivalent image. Boundary crossing and illegal parking event elements are determined solely based on image information and environmental conditions. In another optional embodiment, two high-definition cameras can be installed on the same lamppost with some overlap in their fields of view. The edge computing unit performs the aforementioned processing flow on the monitoring frames from both cameras respectively, and cross-validates events from both perspectives when generating comprehensive alarm data, thereby further improving the reliability of boundary crossing and illegal parking events. Through the above implementation methods, the solar-powered edge computing monitoring device for highway equipment can achieve all-weather intelligent monitoring of construction areas and adjacent lanes without human intervention.

[0038] Example 2: In another embodiment, the hardware layout of the solar-powered edge computing monitoring device for highway equipment is consistent with the aforementioned implementation method. It is also installed on streetlights near the highway construction area, and the top-down equivalent image output by a high-definition camera serves as the main input to the multi-exit anchor-free target detection network. This embodiment focuses on illustrating the multi-exit anchor-free target detection network and its specific implementation in the edge computing unit, enabling those skilled in the art to directly deploy and optimize it.

[0039] In this embodiment, the resolution of the top-view equivalent image is set to 1280 pixels wide and 720 pixels high. The shared backbone structure of the multi-exit anchor-free object detection network running in the edge computing unit receives this top-view equivalent image as a single frame input and generates three feature maps with different spatial scales through continuous convolution and pooling operations, denoted as the first feature map, the second feature map, and the third feature map, respectively. Taking a typical configuration as an example, the resolution of the first feature map is 320 pixels wide and 180 pixels high, the resolution of the second feature map is 160 pixels wide and 90 pixels high, and the resolution of the third feature map is 80 pixels wide and 45 pixels high. The convolution kernels in the shared backbone structure use a longer one-dimensional dimension in the lane direction, for example, a length of 7 in the lane direction and a horizontal length of 3, so that each convolution operation can cover more adjacent pixels in the lane direction. This design allows continuous vehicles, continuous cones, and continuous protective facilities to form continuous response strips along the lane direction on the feature map, which is beneficial for stably finding the target center position in subsequent steps, while maintaining the horizontal resolution and ensuring the distinguishability between lanes.

[0040] In the shared backbone structure, after each pooling operation, the width and height of the feature map are reduced to half that of the previous layer, while the number of channels increases. This allows the first feature map to focus more on preserving the details of small targets, the second feature map to take into account both small and medium-sized targets, and the third feature map to mainly reflect the overall outline of large targets such as construction equipment and large vehicles. Through this multi-scale structure, this embodiment can simultaneously process targets with significant size differences, such as cones, personnel, ordinary vehicles, and large construction machinery, in the same network without needing to design separate networks for different targets, thus reducing the overall system complexity.

[0041] The candidate center exit is connected to the first, second, and third feature maps, respectively. At each scale, the candidate center exit outputs a center response map through a set of convolution operations. The size of the center response map is related to the feature map at the corresponding scale. Figure 1 Each pixel location corresponds to a position on the top-view equivalent image. The edge computing unit first performs threshold filtering on the center response map for each scale, marking pixel locations with response values ​​greater than a preset response threshold as candidate center points. The response threshold can be configured according to the environment; for example, it can be set to 0.5 in well-lit road sections and reduced to 0.3 in foggy or backlit conditions, thus ensuring that a sufficient number of candidate center points are retained even in challenging scenarios.

[0042] After obtaining the candidate center points, the edge computing unit performs local maximum detection on the center response map at each scale, pixel by pixel. Specifically, within a neighborhood of 3×3 or 5×5 windows, it checks if the current pixel's response value is the maximum value in that neighborhood. If so, the candidate center point is retained; otherwise, it is deleted. This avoids generating too many adjacent candidate center points on the same target, reducing redundancy in subsequent processing. After local maximum filtering, three single-scale center sets are obtained, which are then merged into a full-scale candidate set. The edge computing unit sorts the full-scale candidate set according to response values ​​from highest to lowest, selecting the candidate point with the largest response value as the priority center. It then searches the full-scale candidate set for other candidate points whose distance to the priority center is less than a preset distance threshold, grouping these candidate points into the same target center group. Only the candidate point with the largest response value in each target center group is retained as the final target center. The distance threshold can be adjusted according to the feature map scale; for example, it can be set to 6 pixels at the scale corresponding to the first feature map, 4 pixels at the scale corresponding to the second feature map, and 3 pixels at the scale corresponding to the third feature map. This multi-scale aggregation process can merge multiple high-response points of the same physical target generated on feature maps of different scales into a full-scale target center, thereby forming a full-scale target center set and avoiding repeated detection of the same vehicle or the same cone.

[0043] refer to Figure 2 ,

[0044] The size estimation output is connected to the first, second, and third feature maps, respectively, outputting a size response map with four channels at each scale. The four channels represent the pixel distances required to extend from the target center position forward, backward, left, and right to the target edge in the top-view equivalent image. To ensure the comparability of these distance estimation results across different target types, these distance estimates are specified as non-negative numbers during network training, and during operation, estimates less than 1 pixel are uniformly boosted to 1 pixel to avoid rectangular bounding margins with zero width or zero height. The edge calculation unit, for each target center in the full-scale target center set, reads the values ​​of the four channels at that location from the size response map at the corresponding scale, and expands along the four directions in the top-view equivalent image based on these values ​​to obtain the rectangular bounding margin. The rectangular bounding margin geometrically roughly conforms to the outline of a vehicle body, a person's body, or a cone-shaped entity. For targets close to the boundary of the top-view equivalent image, if the bounding area of ​​the rectangle exceeds the image boundary, the edge calculation unit performs cropping according to the image boundary, so that the bounding area of ​​the rectangle always falls within the valid image area, avoiding invalid areas when capturing alarm screenshots in the future.

[0045] The category determination output is also connected to the first, second, and third feature maps, respectively, outputting a category response map with five channels at each scale. These five channels correspond to vehicles, personnel, cones, construction equipment, and protective facilities, respectively. For each target center in the full-scale target center set, the edge computing unit reads the response values ​​of the five channels in the category response map at the corresponding scale, and takes the category corresponding to the channel with the largest response value as the initial target category. In some cases, the difference between the largest and second-largest channel response values ​​is small, for example, less than 0.1, indicating uncertainty in the network's category judgment. In this case, the edge computing unit further performs category correction by combining the aspect ratio and area of ​​the rectangular bounding area. For example, when the aspect ratio of the rectangular bounding area is greater than 2.5 and the area is within a medium range, it better matches the appearance characteristics of a vehicle in a top view, and the target category can be adjusted to vehicle; when the aspect ratio of the rectangular bounding area is between 0.8 and 1.2 and the area is small, it better matches the appearance characteristics of a cone in a top view, and the target category can be adjusted to cone. By combining category response values ​​with geometric features, more stable distinctions can be made between categories such as vehicles and construction equipment, cones and protective facilities, thereby improving actual detection accuracy.

[0046] Figure 2 The dataset uses a 2x3 layout, containing six sub-images. The first row of three sub-images, from left to right, shows the center response distributions of the first, second, and third feature maps. The second row of three sub-images, from left to right, shows the threshold selection results for candidate centers, the single-scale center set after local maximum suppression, and the full-scale target center set after distance clustering. The first sub-image in the first row shows the first feature map with a resolution of 320×180 pixels. This feature map is presented as a heatmap, with colors ranging from dark to light to indicate response values ​​from low to high. The horizontal axis represents the width in pixels, with a scale range of 0 to 320. The vertical axis represents the height in pixels, with a scale range of 0 to 180. Multiple high-response regions are distributed on the feature map, appearing as bright patches. There are obvious response peaks near coordinates 80,45, 120,60, 200,100 and 250,130. The center of each peak is marked with a green cross to indicate the detected candidate center point.

[0047] The first feature map is primarily used for detecting small targets, such as cones and people, thus preserving high spatial resolution. The second sub-image in the first row shows a second feature map with a resolution of 160×90 pixels. This feature map is presented in the same way as the first feature map, using a heatmap format. The horizontal axis ranges from 0 to 160 pixels, and the vertical axis ranges from 0 to 90 pixels. High-response regions exist near coordinates 40,23, 60,30, 100,50, and 125,65, and the center candidate points are also marked with green crosshairs. The color mapping of the response values ​​is consistent with the first feature map. The second feature map has half the resolution of the first feature map and is used to detect both small and medium-sized targets, such as ordinary vehicles and people. The third sub-image in the first row shows a third feature map with a resolution of 80×45 pixels. The horizontal axis ranges from 0 to 80 pixels, and the vertical axis ranges from 0 to 45 pixels. High-response regions exist near coordinates 20,12, 30,15, 50,25, and 62,33, marked with a green cross. The third feature map has the lowest resolution and is primarily used for detecting large targets, such as large vehicles and construction equipment. The peak response locations of the three feature maps correspond to the target locations in the top-view equivalent image at their respective scales. However, due to the different resolutions, the pixel coordinates of the same physical location on different feature maps will be scaled proportionally. The first sub-figure in the second row shows the distribution of the center candidate points after thresholding. The horizontal axis is labeled with the top-view coordinate X, in pixels, ranging from 0 to 100. The vertical axis is labeled with the top-view coordinate Y, in pixels, ranging from 0 to 80. Approximately 50 circular markers are scattered throughout the figure; the size and color intensity of these points indicate their response values. Points with higher response values ​​have larger and darker circular markers. The edges of all points are outlined in black. The color bar on the right indicates the response value range.

[0048] These candidate center points are obtained by extracting pixel locations with response values ​​greater than a preset threshold from feature maps at three scales and mapping them back to the top-view equivalent image coordinate system. Due to the lack of further filtering, some targets have multiple densely distributed candidate points around them. The second sub-figure in the second row shows the single-scale center set after local maximum suppression. The coordinate axis settings are the same as the first sub-figure in the second row. The number of marker points in the figure is significantly reduced, approximately 20 to 25. Each retained candidate center point has the largest response value within its neighborhood window, resulting in relatively uniform spacing between points and eliminating dense clustering. The size and color intensity of the marker points still represent the magnitude of the response value, and the edges of the marker points are outlined in bold black to emphasize that these points are important candidates after filtering. Local maximum suppression effectively reduces redundant detection by retaining only the candidate points with the largest response values ​​within the preset neighborhood window. The third sub-figure in the second row shows the full-scale target center set after distance clustering. The coordinate axis settings are consistent with the first two sub-figures. The figure first uses light-colored circular areas of different colors to label each cluster group, and each cluster group contains several candidate center points that are relatively close to each other.

[0049] The boundaries of the clusters are marked with dashed circles, the color of which matches the fill color of the clusters. Within each cluster, only the candidate points with the highest response values ​​are retained as the final target centers. These target centers are represented by large pentagrams, with the color intensity of the star indicating the magnitude of the response value, and the edges are highlighted with a thick blue outline. The final number of target centers is approximately 10 to 15, and these center points are spatially distributed reasonably, each corresponding to an independent target in the equivalent top-view image. Through multi-scale aggregation and distance clustering, multiple high-response points for the same physical target generated on different feature maps are successfully merged into a single target center, avoiding duplicate detection. Figure 2 The entire process chain, from multi-scale feature map generation, center candidate point extraction, local maximum suppression to distance clustering, is demonstrated, reflecting the hierarchical processing strategy of multi-exit networks in target center localization.

[0050] After obtaining the target center, the bounding rectangle, and the target category, the edge computing unit combines these three elements to form a structured target set. Bounding rectangles with areas that are too small, too large, or whose aspect ratios significantly exceed the normal range can be directly removed from the structured target set. For example, if the bounding rectangle's area is less than 4 pixels and the target category is a vehicle, it can be considered noise and deleted. This geometric constraint further reduces false detections and alleviates the pressure on subsequent boundary crossing and illegal parking detection.

[0051] The region segmentation output is connected to the third feature map of the shared backbone structure. A top-down deconvolution and skip connections are used to progressively restore spatial resolution, ultimately outputting a region label map with the same size as the top-view equivalent image. Each pixel in the region label map is assigned a region type label to distinguish between construction areas, lane areas, and buffer areas. Since construction areas are often composed of cones, guardrails, and painted markings, their boundaries often exhibit jagged edges and gaps in the image. To obtain a smoother construction area mask, the edge computing unit performs connected component analysis on the region label map, merging small connected regions with areas less than a certain threshold into the nearest larger connected region to avoid a large number of isolated pixels. Subsequently, dilation and erosion operations are performed sequentially at the boundaries of the construction area and the buffer area, making the boundary lines more continuous and smooth, ultimately forming the construction area mask and the buffer area mask. This method of first subdividing and then smoothing preserves regional details while reducing boundary instability caused by noise.

[0052] Based on the basic configuration of this embodiment, the inference time of the multi-exit anchor-free target detection network is controlled to within 50 milliseconds per frame on the edge computing unit. Combined with the camera's acquisition frequency of 10 frames per second, near real-time target detection and region segmentation can be achieved. The structured target set and region label map generated in this embodiment will be combined with dynamic virtual warning boundaries, boundary crossing event elements, and illegal parking event elements in other embodiments to complete the identification of boundary crossing behavior in construction areas and illegal parking behavior in lanes.

[0053] In an optional implementation, the top-view equivalent image and the original perspective image can be simultaneously input into the shared backbone structure. For example, the top-view equivalent image and the original perspective image can be concatenated along the channel dimension and provided as a multi-channel input to the shared backbone structure. This allows the network to utilize both top-view geometric relationships and original perspective texture information simultaneously when extracting features. For certain scenarios, such as complex overpasses or multi-level ramps, the height variation information in the original perspective image helps distinguish traffic flow at different levels, further improving the localization accuracy of the structured target set. In another optional approach, some ordinary convolutional operations in the shared backbone structure can be replaced by depthwise separable convolutions to reduce inference computation. This reduces the power consumption of the edge computing units while maintaining relatively constant detection accuracy, enabling the solar-powered edge computing monitoring device for highway equipment to maintain normal operation of the multi-exit anchor-free target detection network even during continuous rainy weather.

[0054] Example 3: In an implementation combining Examples 3 and 4, the hardware configuration of the solar-powered edge computing monitoring device for highway equipment remains consistent with the aforementioned examples: a high-definition camera outputs a top-down equivalent image after distortion correction and perspective transformation; a multi-exit anchor-free target detection network outputs a structured target set and region label map; environmental sensors provide illumination and fog values; and a millimeter-wave radar outputs a distance point array when needed. This embodiment focuses on illustrating the specific generation process of dynamic virtual warning boundaries, boundary crossing event elements, and illegal parking event elements consistent with millimeter-wave radar markers.

[0055] During on-site deployment, construction units typically use cones and protective barriers to demarcate the outer edge of the construction area. A multi-exit anchor-free target detection network identifies targets categorized as cones and protective barriers in the equivalent top-view image and incorporates their center points into a structured target set. The edge computing unit first reads the coordinates of the center points of all targets categorized as cones and protective barriers in the structured target set, assigning each center point to the left and right candidate sets based on its position relative to the lane centerline. Within each candidate set, the coordinates are then sorted from smallest to largest along the main lane direction, aligning the cones and protective barriers along the vehicle's direction of travel.

[0056] After sorting, the edge computing unit sequentially connects the centers of two adjacent targets in the left candidate set on the top-view equivalent image to obtain the left-side polyline sequence, and similarly obtains the right-side polyline sequence. Since construction sites may have cones temporarily obscured by vehicles or with inconsistent spacing, simply connecting them with straight lines can easily result in abrupt angles. Therefore, the edge computing unit performs curve fitting on the polyline sequence, transforming the original polylines composed of several line segments into a smooth curve using a small number of control points. This ensures the curve passes through the target center and provides reasonable boundary completion in sparse cone areas. The curve fitting uses a method that limits the maximum degree of curvature to avoid sharp bends at the boundary over short distances, which is beneficial for subsequent calculations of the distance from the target center to the boundary.

[0057] After obtaining the left and right boundary curves, the edge computing unit performs a buffer expansion on both sides of each curve by a fixed pixel width. For example, it can choose to expand by 10 pixels on the construction area side and 5 pixels on the non-construction area side, forming a strip-shaped warning zone with a certain width. The design of the strip-shaped warning zone ensures that boundary crossing judgment is not based on a line without width, but on a transition area. Retaining a wider expansion width on the construction area side can accommodate slight offset of the cones and also provide some room for movement for construction personnel. Subsequently, the edge computing unit uses the position of the construction area label in the area label map to align the strip-shaped warning zone with the outer boundary of the construction area mask. If there is a deviation, the position of the strip-shaped warning zone is locally adjusted based on the outer boundary of the construction area mask to obtain a dynamic virtual warning boundary covering the edge of the construction area. In this way, when the construction unit moves the cones or adds or removes protective facilities, as long as the structured target set can correctly identify the cones and protective facilities, the dynamic virtual warning boundary can be automatically updated without the need for manual configuration of the boundary line again.

[0058] Based on a dynamic virtual warning boundary, the edge computing unit performs boundary crossing determination for personnel and vehicles in the structured target set. For each target categorized as either a person or a vehicle, the edge computing unit first reads the region type corresponding to the target center in the region label map. If the region type is a construction area or a buffer area, it indicates that the target has approached the edge of the construction area or entered its interior, requiring further determination to determine whether it constitutes a boundary crossing. To this end, the edge computing unit calculates the minimum pixel distance from the target center to the inner and outer edge lines of the dynamic virtual warning boundary in the top-view equivalent image, and uses the smaller of the two values ​​as the minimum distance between the target and the boundary. The environmental sensor provides the current illumination and fog values, and the edge computing unit sets a safe distance threshold based on these values. For example, when the illumination value is higher than a preset illumination threshold and the fog value is lower than a preset fog threshold, the safe distance threshold is set to 5 pixels; when the illumination value is lower than the illumination threshold or the fog value is higher than the fog threshold, the safe distance threshold is increased to 8 pixels. The rationale behind this is that when there is insufficient light or heavy fog, the ability of drivers and construction workers to react to changes in the environment is reduced. By increasing the safety distance threshold, an alarm can be triggered in advance before the target is close to the warning line, thereby increasing the safety margin for driving and operation.

[0059] Based on this determination, when the target center is located on one side of the construction area or within the buffer zone, and the minimum distance between it and the dynamic virtual warning boundary is less than the safe distance threshold, the edge computing unit marks the target as an out-of-bounds target. For each out-of-bounds target, the edge computing unit generates out-of-bounds event elements, which include the target category, target center coordinates, rectangular bounding area, dynamic virtual warning boundary index, and current environmental state. Simultaneously, an image segment corresponding to the rectangular bounding area is extracted from the original view image as an out-of-bounds screenshot, providing intuitive evidence for the management center platform and subsequent evidence collection.

[0060] In the same implementation, the generation of illegal parking event elements depends on a structured target set and evidence of stationary vehicles in the current frame image. The edge computing unit selects targets from the structured target set whose target category is vehicle and whose corresponding target center is located within the lane area mask, and considers these targets as candidate vehicles. For each candidate vehicle, the edge computing unit constructs three pieces of evidence of stationary vehicles around its rectangular circumscribed area: wheel texture evidence, ground shadow evidence, and contour sharpness evidence.

[0061] In the wheel texture evidence, the edge computing unit divides the lower region of the vehicle's rectangular outer boundary into four candidate wheel sub-regions based on left and right, and front and rear, corresponding to the left front wheel, right front wheel, left rear wheel, and right rear wheel, respectively. Edge detection is performed in each candidate wheel sub-region, and the number of edge pixels and fine texture lines are counted. When there are obvious tire treads on the wheel surface, their shape boundaries and patterns will form dense and directionally varied light and dark boundary lines in the image, which will make the number of edge pixels and fine texture lines significantly higher than the surrounding road surface or vehicle body area. Therefore, when the number of edge pixels and fine texture lines in at least two candidate wheel sub-regions are simultaneously not less than a preset threshold, the edge computing unit considers the wheel texture evidence to be valid, that is, the vehicle's tires are clearly visible, and the image does not show obvious motion blur due to high-speed movement.

[0062] In grounding shadow evidence, the edge computing unit extends a narrow band downwards towards the road surface, based on the bottom edge of the vehicle's rectangular circumscribed area. The width of this narrow band can be set to 5 to 15 pixels. After converting this narrow band to a grayscale image, it is divided into dark and bright regions using a fixed grayscale threshold. The edge computing unit performs connected component detection on the dark regions, searching for dark connected bands extending laterally near the bottom of the vehicle. When a dark connected band with a width within a preset range and roughly traversing the bottom of the vehicle laterally is detected, it indicates that a relatively continuous grounding shadow has formed under the current exposure conditions. For vehicles traveling at high speeds, the relative position between the vehicle body and the road surface changes significantly. Even if shadows exist, they often appear discontinuous rather than horizontally extending dark connected bands. Therefore, the presence of stable dark connected bands can be considered valid grounding shadow evidence.

[0063] In the evidence of sharpness of the vehicle outline, the edge computing unit extends a fixed-width outer band region around the perimeter of the vehicle's rectangular outline, for example, with a width of 5 to 10 pixels. Edge detection is performed within these outer band regions, and the distribution of edge directions is analyzed. If the vehicle outline is sharp and clear in the image, the edges surrounding the vehicle's outer outline are usually concentrated along the vehicle's boundary direction, and the outline is continuous and uninterrupted. However, when the vehicle is moving at high speed and the shutter speed is long, the vehicle outline is prone to blurring, resulting in scattered edge directions and many broken outline lines in the outer band region. Based on this phenomenon, the edge computing unit calculates the dispersion of edge directions in the outer band region. When the dispersion falls within a preset range, and the vehicle outline lines exhibit a large-span continuous distribution in all four directions, the evidence of sharpness of the outline is considered valid.

[0064] All three types of evidence are not required to be valid. In this embodiment, the edge computing unit sets the evidence quantity threshold to 2. When at least two of the following three pieces of evidence are valid, the vehicle is identified as a stationary vehicle: wheel texture evidence, ground shadow evidence, and outline clarity evidence. If the target center of the stationary vehicle is located within the lane area mask and outside the dynamic virtual warning boundary, it is considered an illegally parked vehicle, and an illegal parking event element is generated. The illegal parking event element includes at least the vehicle target center coordinates, the bounding area of ​​the rectangle, the evidence type list, and the lane number.

[0065] When millimeter-wave radar is involved, the edge computing unit also adds millimeter-wave radar consistency markers to the illegal parking event elements. The millimeter-wave radar outputs a series of distance points at fixed intervals, each containing the distance and azimuth information of the reflecting object. The edge computing unit maps the position of the vehicle's rectangular circumscribed area in the top-view equivalent image to a sector in the millimeter-wave radar coordinate system, and then counts the number of distance points falling within this sector over several consecutive sampling periods. If the number of distance points in each of three consecutive sampling periods is not less than a point count threshold (e.g., not less than 10), it is considered that a stable reflector exists at that spatial location, and the distance and azimuth of this reflector are consistent with the structured target position of the illegally parked vehicle. Therefore, a millimeter-wave radar consistency marker is added to the illegal parking event element. This multi-period, multi-point judgment method effectively filters out sporadic reflections and noise points, improving the reliability of illegal parking vehicle determination.

[0066] After boundary crossing and illegal parking event elements are generated, the edge computing unit extracts corresponding image segments from the original viewpoint image according to the rectangular circumscribed range. These image segments, along with the target category, target center coordinates, event type, dynamic virtual warning boundary index, current environmental status, and optional millimeter-wave radar consistency markers, are encapsulated into comprehensive alarm data. The local alarm unit selects different prompting strategies based on the event type of the comprehensive alarm data. For boundary crossing events involving construction workers, a clear directional voice message can be played to remind the workers to return to a safe area. For illegal parking events, strong audio and visual signals are used to remind passing vehicles to slow down and avoid the area, and the illegally parked vehicle is prompted to move away as soon as possible. The comprehensive alarm data is simultaneously uploaded to the management center platform via a 4G or 5G wireless communication module for remote monitoring and post-event traceability.

[0067] In one alternative implementation, the dynamic virtual warning boundary does not necessarily have to consist of two marker chains on the left and right. If cones are only placed on one side of the construction site, while the other side is adjacent to the central median, the edge computing unit can construct a strip-shaped warning zone using only one side's marker chain, and directly use the outer boundary of the construction area mask on the other side. This reduces the requirements for on-site deployment and maintains a reasonable safety boundary near the central median. In another alternative implementation, the evidence quantity threshold can be adjusted according to the lane speed limit. The evidence quantity threshold can be appropriately increased in sections with higher speed limits, and a vehicle is only judged to be illegally parked when all three types of evidence are valid, thereby reducing false alarms for brief stop-to-avoidance behaviors. Through the above implementation methods, the solar-powered edge computing monitoring device for highway equipment can not only accurately identify boundary crossing behaviors in the construction area based on the dynamic virtual warning boundary, but also identify illegally parked vehicles in a single frame without relying on historical video streams, and improve the overall reliability of the judgment by utilizing millimeter-wave radar consistency marking.

[0068] refer to Figure 3 , Figure 3The horizontal axis represents four different power status levels and their corresponding operating modes, from left to right: Level 0 (Energy Saving Mode), Level 1 (Energy Saving Mode), Level 2 (High-Performance Mode), and Level 3 (High-Performance Mode). The vertical axis represents system power consumption in watts, ranging from 0 to 32 watts. The graph uses a stacked bar chart, with each bar consisting of six differently colored stacked segments, representing the power consumption contributions of the edge computing unit, HD camera, millimeter-wave radar, environmental sensor, communication module, and other loads from bottom to top. In Level 0 Energy Saving Mode, the edge computing unit consumes 8 watts, occupying the bottom of the bar. Above it are stacked segments: HD camera (2 watts), millimeter-wave radar (0 watts), environmental sensor (0.2 watts), communication module (1 watt), and other loads (1 watt), for a total power consumption of 12.2 watts. The total power consumption value of 12.2W is marked at the top of the bar. In Level 1 energy-saving mode, the edge computing unit's power consumption increases to 10 watts, the HD camera's power consumption is 3 watts, the millimeter-wave radar's power consumption is 1 watt, the environmental sensor's power consumption is 0.5 watts, the communication module's power consumption is 1.5 watts, and other loads' power consumption is 1 watt, for a total power consumption of 17.0 watts, marked 17.0W at the top of the bar. In Level 2 high-performance mode, the edge computing unit's power consumption reaches 15 watts, the HD camera's power consumption is 5 watts, the millimeter-wave radar's power consumption is 3 watts, the environmental sensor's power consumption is 0.8 watts, the communication module's power consumption is 2.5 watts, and other loads' power consumption is 1.5 watts, for a total power consumption of 27.8 watts, marked 27.8W at the top of the bar. In Level 3 high-performance mode, the power consumption of each component is the same as in Level 2, and the total power consumption is also 27.8 watts. The graph connects the total power consumption levels of Level 0 and Level 3 with horizontal dashed lines. The dashed line on the left indicates that the power consumption level of Level 0 is approximately 12 watts, and the dashed line on the right indicates that the power consumption levels of Level 2 and Level 3 are approximately 28 watts. The energy-saving rate calculation results are labeled on the right side of the graph, showing that switching from high-performance mode to energy-saving mode can achieve an energy saving rate of 56.1%. Different stacked segments of the bar chart are distinguished by different gray levels or textures, with the bottom edge calculation unit segments being darker in color, gradually transitioning to lighter colors upwards. The legend is located in the upper left corner of the graph, labeling the component names corresponding to each stacked segment. Figure 3 The impact of energy state level on overall system power consumption is clearly demonstrated. When the battery is fully charged, the system operates in high-performance mode, with the edge computing unit running at its highest frequency, the HD camera acquiring images at the highest resolution and frame rate, the millimeter-wave radar and environmental sensors all participating, and the communication module maintaining high-frequency data transmission. At this time, the total power consumption is close to 28 watts. When the battery level drops to a lower level, the system switches to energy-saving mode. The edge computing unit reduces its processing frequency, the HD camera reduces its resolution and frame rate, the millimeter-wave radar is activated only when necessary, the environmental sensor extends its sampling interval, and the communication module adopts a data merging and reporting strategy. At this time, the total power consumption is reduced to 12-17 watts, achieving significant energy savings.

[0069] While specific embodiments of the present invention have been described above, those skilled in the art should understand that these specific embodiments are merely illustrative. Those skilled in the art can omit, substitute, and modify the details of the above methods and systems in various ways without departing from the principles and essence of the present invention. For example, combining the above method steps to perform substantially the same function and achieve substantially the same result according to substantially the same method falls within the scope of the present invention. Therefore, the scope of the present invention is defined only by the appended claims.

Claims

1. A solar-powered edge computing monitoring device for highway equipment, characterized in that, include: Solar panels, a power management module, and a battery; the solar panels supply power to the battery and edge computing loads via the power management module. A high-definition camera is used to capture monitoring images of the highway construction area; an edge computing unit, electrically connected to the power management module and the high-definition camera, is configured to preprocess the monitoring images and generate boundary crossing and illegal parking event elements based on a multi-exit anchor-free target detection network; a 4G or 5G wireless communication module, connected to the edge computing unit, is used to send comprehensive alarm data encapsulated based on boundary crossing and illegal parking event elements to the management center platform; a local alarm unit, connected to the edge computing unit, is used to generate audible and visual alarms based on the comprehensive alarm data; the multi-exit anchor-free target detection network is built-in. The edge computing unit includes a shared backbone structure, a central candidate exit, a size estimation exit, a category determination exit, and a region segmentation exit. The shared backbone structure takes a top-view equivalent image as input, and sequentially sets convolutional and pooling layers to extract texture features, structural features, and semantic features, and outputs a first feature map, a second feature map, and a third feature map at three different spatial scales. When designing, the shared backbone structure establishes the main direction axis of the lane based on the road markings. In the convolution operation, it expands the sensing area along the main direction axis of the lane, so that vehicles, cones, and protective facilities arranged along the lane direction form a continuous response on the first feature map, the second feature map, and the third feature map, providing a unified spatial feature basis for each exit.

2. The apparatus according to claim 1, characterized in that, The power management module is electrically connected to the solar panel, the battery and the edge computing unit respectively. The power management module collects the output voltage and output current of the solar panel and the remaining power of the battery. It divides the collected results into multiple discrete energy state levels according to the segmentation rules and sends them to the edge computing unit. The edge computing unit selects between a high-performance working mode and an energy-saving working mode based on the energy state level. In the high-performance working mode, it acquires monitoring images at the preset highest resolution and frame rate and enables all configured sensors. In the energy-saving working mode, it reduces the acquisition resolution and frame rate of the monitoring images and shuts down some non-critical sensors, thereby achieving adaptive control of the overall power consumption of the solar-powered edge computing monitoring device for highway equipment under different energy state levels.

3. The apparatus according to claim 2, characterized in that, It includes millimeter-wave radar and environmental sensors, both of which are electrically connected to the edge computing unit. The edge computing unit controls the high-definition camera to acquire monitoring images and performs lens distortion correction, perspective transformation based on road markings, block brightness equalization, and color deviation compensation to obtain the original perspective image. Based on perspective transformation, the original view image is mapped to a top-view equivalent image with the road surface as a reference. When configuring millimeter-wave radar, distance point sequences corresponding to the current monitoring screen are collected. The distance point sequences are mapped to the top-view equivalent image coordinate system through a calibration matrix, and each distance point is assigned a top-view coordinate. Environmental sensors output information on illumination, fog, precipitation, and wind speed. The edge computing unit associates the environmental information with the original view image and the top-view equivalent image at the corresponding time, which is used as the environmental state when generating event elements.

4. The apparatus according to claim 1, characterized in that, The candidate center exit is connected to the first feature map, the second feature map, and the third feature map respectively, generating a center response map on each scale feature map. Each pixel position in the center response map corresponds to a position in the top-view equivalent image. The edge computing unit selects pixel positions with response values ​​greater than a preset response threshold as center candidate points in the center response map and assigns them top-view coordinates to form a single-scale center candidate set. Each single-scale center candidate set is sorted in descending order of response value, and only the center candidate point with the largest response value is retained in each preset neighborhood window to obtain a single-scale center set. The single-scale center sets are merged into a full-scale candidate set, and candidate points are selected as priority centers in descending order of response value. Other candidate points with a distance less than a preset distance threshold from the priority centers are found in the full-scale candidate set. These candidate points are assigned to the same target center group, and only the candidate point with the largest response value in each target center group is retained as the target center to form a full-scale target center set.

5. The apparatus according to claim 4, characterized in that, The size estimation output is connected to the first feature map, the second feature map, and the third feature map, respectively. At each scale, a size response map containing four channels is output. These four channels represent the estimated pixel distances from the target center to the forward, backward, left, and right edges in the top-view equivalent image. For each target center in the full-scale target center set, the edge computing unit reads the four estimated pixel distances at the target center location from the size response map at the corresponding scale. Using the target center as the geometric center, it expands the rectangle along the forward, backward, left, and right directions in the top-view equivalent image based on the estimated pixel distances, forming a rectangular bounding area. This rectangular bounding area covers the outlines of vehicles, personnel, cones, construction equipment, or protective facilities. When the rectangular bounding area exceeds the boundary of the top-view equivalent image, the edge computing unit clips the rectangular bounding area according to the top-view equivalent image boundary, ensuring that the rectangular bounding area is completely within the top-view equivalent image.

6. The apparatus according to claim 5, characterized in that, The category determination output is connected to the first, second, and third feature maps, respectively, and outputs a multi-channel category response map on each scale feature map. Each channel corresponds to the categories of vehicle, personnel, cone, construction equipment, and protective facilities, respectively. For each target center in the full-scale target center set, the edge computing unit reads the response values ​​of each channel at the target center location in the category response map of the corresponding scale, and determines the category corresponding to the channel with the largest response value as the target category. When the difference between the largest channel response value and the second largest channel response value is less than a preset difference threshold, the edge computing unit reads the aspect ratio of the rectangle bounding area corresponding to the target center. When the aspect ratio is greater than the first aspect ratio threshold, the target category is adjusted to vehicle. When the aspect ratio is within the range of the second aspect ratio threshold and the area of ​​the rectangle bounding area is within the range of the cone area threshold, the target category is adjusted to cone. The edge computing unit combines the target center, the rectangle bounding area, and the target category to form a structured target set for boundary construction and event element generation.

7. The apparatus according to claim 6, characterized in that, The region segmentation exit is connected to the highest-scale feature map of the shared backbone structure to generate a region label map with the same size as the top-view equivalent image. Each pixel in the region label map is labeled as a construction area, lane area, or buffer area. The edge computing unit performs connected component analysis on the region label map, merging connected components of regions with areas smaller than a first area threshold into adjacent regions of the same type with areas not smaller than a second area threshold. Dilation and erosion operations are sequentially performed at the boundaries of the construction area and the buffer area to form continuous closed contour lines, generating a construction area mask and a buffer area mask. The edge computing unit... The calculation unit selects targets categorized as cones and protective facilities from the structured target set, sorts the target center coordinates according to the main direction of the lane, and divides targets on one side of the lane centerline into a left-side marker chain and targets on the other side into a right-side marker chain, using the lane centerline as a reference. In each marker chain, the centers of two adjacent targets are connected to form a broken line sequence. Curve fitting is performed on the broken line sequence to obtain a continuous boundary curve. A strip warning zone is formed on both sides of the boundary curve by buffering and expanding according to a fixed pixel width. The strip warning zone is aligned with the outer boundary of the construction area mask to obtain a dynamic virtual warning boundary covering the edge of the construction area.

8. The apparatus according to claim 7, characterized in that, The edge computing unit generates boundary crossing event elements based on the dynamic virtual warning boundary: it selects targets categorized as personnel and vehicles from the structured target set, reads the coordinates of the target center in the top-view equivalent image, and queries the region type corresponding to the coordinates in the region label map; when the region type is a construction area or a buffer area, it calculates the minimum pixel distance from the target center to the inner and outer edge lines of the dynamic virtual warning boundary, and sets a safe distance threshold for determining boundary crossing based on the illumination and fog values ​​given by the environmental sensors. When the illumination value is lower than the illumination threshold or the fog value is higher than the fog threshold, the safe distance threshold is increased; when the minimum pixel distance is less than the safe distance threshold and the target center is located on one side of the construction area or within the buffer area, the target is determined to be a boundary crossing target, and boundary crossing event elements are generated. The boundary crossing event elements include at least the target category, target center coordinates, rectangular bounding range, dynamic virtual warning boundary index, and current environmental state.

9. The apparatus according to claim 8, characterized in that, The process by which the edge computing unit generates illegal parking event elements and encapsulates comprehensive alarm data based on evidence of stationary vehicles includes: for targets located within the lane area mask and whose target category is vehicle, constructing wheel texture evidence, ground shadow evidence, and contour clarity evidence within the corresponding rectangular bounding area; dividing the lower part of the vehicle's rectangular bounding area into four wheel candidate sub-regions according to proportions, performing edge detection in each wheel candidate sub-region and counting the number of edge pixels and fine texture lines; when the number of edge pixels and fine texture lines in at least two wheel candidate sub-regions are not less than the corresponding thresholds, the wheel texture evidence is established; within the narrow band area extending downwards from the bottom edge of the vehicle's rectangular bounding area, converting the narrow band area into a grayscale image and dividing it into dark and bright regions according to a grayscale threshold, performing connected component detection, and when a connection is formed near the bottom of the vehicle along the vehicle's horizontal direction... When a continuous dark connected band with a width falling within a preset width range is formed, grounding shadow evidence is established. Edge detection is performed and the edge direction distribution is statistically analyzed within the outer band area extending around the vehicle's rectangular outer range. When the edge direction is within a preset dispersion range and the vehicle's outline is continuous within the outer band area, outline clarity evidence is established. When the number of such evidences reaches the evidence quantity threshold, the target vehicle is identified as an illegally parked vehicle, and illegal parking event elements are generated. The illegal parking event elements include at least the vehicle target center coordinates, the rectangular outer range, the evidence type list, and the lane number. When a millimeter-wave radar is configured, the edge computing unit searches for distance point columns within the top-view equivalent image coordinate range near the rectangular outer range of the illegally parked vehicle. When a group of distance points with a number of points not less than the point quantity threshold is detected in multiple consecutive sampling periods, a millimeter-wave radar consistency marker is added to the illegal parking event elements.

Images