Idle scooter detection method and system

Abstract

The invention relates to the technical field of image processing, and discloses an idle scooter detection method and system. The method comprises the following steps: acquiring a multi-source video through a city monitoring camera, preprocessing the multi-source video, identifying a scooter by using improved YOLOv8, constructing a self-adaptive region of interest through polar coordinate mapping, extracting pedestrian behavior characteristics by using a double-flow network, and calculating interaction strength in combination with space intersection, time continuity and posture identification. And finally, judging the idle state of the scooter based on the hierarchical decision-making tree and scoring. According to the method, the idle state of the shared scooter in the urban environment can be accurately identified and judged in real time, and high-precision use state evaluation is provided.

Description

Technical Field

[0001] The present application relates to the field of image processing technology, and in particular to a method and system for detecting idle scooters. Background Art

[0002] With the rapid development of the sharing economy, shared scooters have become a vital component of modern urban mobility. Traditional shared scooter management relies primarily on manual inspections and user feedback, requiring operators to regularly check scooter usage, location, and idleness. The number of shared scooters in cities continues to grow, spreading across streets, sidewalks, and public spaces, forming a complex urban transportation ecosystem.

[0003] Existing scooter management technologies have numerous limitations. Traditional methods, which primarily rely on GPS positioning and manual inspections, are unable to accurately and efficiently determine the specific usage status of scooters in real time. These methods struggle to precisely distinguish whether a scooter is in use, temporarily parked, or permanently idle, resulting in inefficient resource utilization. Furthermore, existing technologies lack in-depth analysis of the scooter's surroundings and human-vehicle interactions, making them unable to provide accurate usage status assessments. Summary of the Invention

[0004] The present application provides an idle scooter detection method and system for accurately and in real time identifying and judging the idle status of shared scooters in an urban environment, and providing high-precision usage status assessment.

[0005] In a first aspect, the present application provides a method for detecting idle scooters, the method comprising: collecting multi-source video streams through a city surveillance camera network, performing resolution normalization, inter-frame differencing, and illumination compensation processing on the multi-source video streams to obtain preprocessed video data; based on the preprocessed video data, performing scooter identification and time series association using an improved YOLOv8 deep learning model to obtain scooter position coordinates and trajectory data; based on the scooter position coordinates and trajectory data, constructing an elliptical adaptive region of interest through polar coordinate mapping and environmental complexity assessment; for the adaptive region of interest, performing pedestrian detection and behavior feature extraction using a dual-stream network architecture to obtain pedestrian position trajectory, behavior type, and dwell time data; combining the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory to obtain a human-vehicle interaction intensity score through spatial intersection ratio calculation, temporal continuity analysis, and posture sequence recognition; based on the human-vehicle interaction intensity score, classifying and determining the scooter state using a hierarchical decision tree model to obtain the scooter idle state and its confidence score.

[0006] In a second aspect, the present application provides an idle scooter detection system, the idle scooter detection system comprising:

[0007] An acquisition module is used to collect multi-source video streams through a city surveillance camera network, and perform resolution normalization, inter-frame difference and illumination compensation processing on the multi-source video streams to obtain pre-processed video data;

[0008] An identification module is used to perform scooter identification and time series association based on the preprocessed video data by improving the YOLOv8 deep learning model to obtain the scooter position coordinates and trajectory data;

[0009] An evaluation module, configured to construct an elliptical adaptive region of interest based on the scooter position coordinates and trajectory data through polar coordinate mapping and environment complexity evaluation;

[0010] An extraction module is used to detect pedestrians and extract their behavior features using a dual-stream network architecture for the adaptive region of interest, thereby obtaining pedestrian location trajectory, behavior type, and dwell time data.

[0011] An analysis module, configured to combine the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory to obtain a human-vehicle interaction intensity score through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition;

[0012] The determination module is used to classify and determine the scooter state according to the human-vehicle interaction intensity score through a hierarchical decision tree model to obtain the scooter idle state and its confidence score.

[0013] In a third aspect, an idle scooter detection device is provided, comprising: a memory and at least one processor, wherein the memory stores instructions; the at least one processor calls the instructions in the memory to enable the idle scooter detection device to execute the above-mentioned idle scooter detection method.

[0014] In a fourth aspect, a computer-readable storage medium is provided, wherein instructions are stored in the computer-readable storage medium, which, when executed on a computer, causes the computer to execute the above-mentioned idle scooter detection method.

[0015] The technical solution provided in this application significantly improves the intelligence and practicality of shared scooter idle state detection by innovatively integrating multi-source video stream processing, deep learning target recognition, and artificial intelligence interactive analysis technologies. Its core technical features demonstrate the unique value and transformative potential of artificial intelligence algorithms in complex urban environment monitoring. First, by collecting multi-source video streams through a network of urban surveillance cameras and performing resolution normalization, inter-frame differencing, and illumination compensation processing, this not only solves the data consistency issues caused by differences in shooting angles, lighting conditions, and resolution in traditional video analysis, but also provides high-quality, standardized input data for subsequent deep learning algorithms. This preprocessing step fully demonstrates the critical impact of data preprocessing on the performance of artificial intelligence models. Second, the improved YOLOv8 deep learning model greatly improves the accuracy and real-time performance of scooter recognition through multi-scale feature extraction, attention mechanism enhancement, and dedicated anchor frame configuration. In particular, the inter-frame matching and trajectory smoothing achieved by the Kalman filter algorithm not only enhance the stability of target tracking but also significantly reduce the jitter and error of target positioning, fully demonstrating the intelligent advantages of deep learning algorithms in target detection and motion trajectory analysis. Furthermore, the elliptical adaptive region of interest (ROI) constructed based on scooter position coordinates and trajectory data achieves intelligent perception and dynamic adaptation to urban scenarios through polar coordinate mapping and environmental complexity assessment. This not only accurately locates potential idle areas but also dynamically adjusts the ROI boundaries based on surrounding pedestrian density, demonstrating the unique value of artificial intelligence algorithms in complex environmental perception and adaptive analysis. Furthermore, the application of a dual-stream network architecture for pedestrian detection and behavioral feature extraction achieves precise identification and in-depth understanding of human-vehicle interactions by fusing position features extracted by a spatial stream network with motion features calculated by a temporal stream network. This architecture not only accurately captures the spatial relationship and dynamic changes between pedestrians and scooters, but also provides richer and more reliable semantic information for idle state determination through posture recognition and behavior classification. Finally, through spatial intersection ratio calculation, temporal persistence analysis, posture sequence recognition, and combined with the feature fusion of a multi-layer perceptron, a multi-dimensional, multi-level human-vehicle interaction intensity assessment model was constructed. This not only breaks through the traditional single-dimensional interaction analysis method, but also provides a more comprehensive and intelligent input for the hierarchical decision tree model. Ultimately, it achieves accurate identification and confidence scoring of the idle state of scooters, fully demonstrating the intelligent reasoning and state discrimination capabilities of artificial intelligence algorithms in complex decision-making scenarios.

[0016] It should be understood that the contents described in the Summary of the Invention are not intended to limit the key or important features of the embodiments of the present application, nor are they intended to limit the scope of the present application. Other features of the present application will become readily understood through the following description. BRIEF DESCRIPTION OF THE DRAWINGS

[0017] The above and other features, advantages and aspects of the embodiments of the present application will become more apparent with reference to the following detailed description in conjunction with the accompanying drawings. The accompanying drawings are provided for a better understanding of the present invention and do not constitute a limitation of the present application. In the accompanying drawings, the same or similar reference numerals represent the same or similar elements, among which:

[0018] Figure 1 This is a schematic diagram of an embodiment of a method for detecting an idle scooter in an embodiment of the present application;

[0019] Figure 2 This is a schematic diagram of an embodiment of an idle scooter detection system in an embodiment of the present application;

[0020] Figure 3 2 is a schematic block diagram of the structure of an idle scooter detection device in an embodiment of the present invention. DETAILED DESCRIPTION

[0021] To make the purpose, technical solutions, and advantages of the embodiments of this application more clear, the technical solutions in the embodiments of this application will be clearly and completely described below in conjunction with the drawings in the embodiments of this application. Obviously, the described embodiments are part of the embodiments of this application, not all of the embodiments. Based on the embodiments in this application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0022] In this document, the term "and / or" simply describes a relationship between related objects, indicating that three possible relationships exist. For example, "A and / or B" can represent: A exists alone, A and B exist simultaneously, or B exists alone. Furthermore, the character " / " in this document generally indicates that the related objects are in an "or" relationship.

[0023] For ease of understanding, the specific process of the embodiment of the present application is described below. Figure 1 In one embodiment of the present application, an idle scooter detection method includes:

[0024] Step S101: collecting multi-source video streams through a city surveillance camera network, performing resolution standardization, inter-frame difference and illumination compensation processing on the multi-source video streams to obtain pre-processed video data;

[0025] Step S102: Based on the pre-processed video data, the scooter is identified and time-series associated by improving the YOLOv8 deep learning model to obtain the scooter position coordinates and trajectory data;

[0026] Step S103: constructing an elliptical adaptive region of interest based on the scooter's position coordinates and trajectory data through polar coordinate mapping and environment complexity assessment;

[0027] Step S104: For the adaptive region of interest, pedestrian detection and behavior feature extraction are performed using a dual-stream network architecture to obtain pedestrian location trajectory, behavior type, and dwell time data;

[0028] Step S105: combining the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory, and obtaining a human-vehicle interaction intensity score through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition;

[0029] Step S106: Based on the human-vehicle interaction intensity score, the scooter status is classified and determined using a hierarchical decision tree model to obtain the scooter idle status and its confidence score.

[0030] It is understandable that the execution subject of this application can be an idle scooter detection system, or a terminal or a server, which is not limited here. The embodiment of this application is described by taking the server as the execution subject as an example.

[0031] Specifically, a multi-source video stream is collected through a city surveillance camera network. This network consists of surveillance cameras distributed throughout the city, covering major roads, public areas, and designated scooter parking areas. The raw video streams suffer from varying resolutions, varying lighting conditions, and complex backgrounds, necessitating preprocessing. During preprocessing, the raw video streams are first normalized to 1080p resolution to ensure uniform pixel density and clarity across all videos. A background difference algorithm is then used to calculate the difference between adjacent frames, extracting a sequence of dynamic regions containing moving objects and reducing computational redundancy in the static background. Gamma correction and histogram equalization are then applied to the dynamic region sequence to compensate for images under varying lighting conditions, generating a normalized image. The frame rate is then adaptively adjusted based on scene complexity, maintaining a high frame rate in crowded areas and appropriately reducing it in simple scenes, resulting in a streamlined video sequence. Finally, a spatiotemporal registration algorithm is used to fuse the streamlined video sequences from each acquisition node across multiple perspectives, ensuring consistent identification of the same scooter from different camera perspectives. The resulting preprocessed video data is then output. Based on preprocessed video data, this method uses a modified YOLOv8 deep learning model for scooter recognition. This model first feeds the preprocessed video data into a multi-scale feature extraction network to generate a map containing scooter features at different scales. An attention mechanism is applied to the scooter feature map to enhance the features of core components, such as the handlebars, pedals, and wheels, forming an enhanced feature map of key regions. Object detection is performed on this enhanced feature map using scooter-specific anchor frames. These anchor frames are designed based on the typical aspect ratio and dimensions of scooters to better reflect their morphological characteristics. A preliminary set of detection boxes is then output. Non-maximum suppression is then applied to this set of detection boxes to remove overlapping and redundant boxes, resulting in a high-confidence scooter bounding box. Unique instance IDs are assigned to high-confidence scooter bounding boxes to facilitate cross-frame tracking, generating the scooter's basic position data. Finally, a Kalman filter algorithm is used to perform inter-frame matching and trajectory smoothing on the basic scooter position data to address object loss due to occlusion or viewpoint changes, ultimately yielding the scooter's position coordinates and trajectory data.

[0032] An elliptical adaptive region of interest (ROI) is constructed based on the scooter's location coordinates and trajectory data. First, a polar coordinate system is established with the scooter's location coordinate center as the origin to define the monitoring space around the scooter. The forward direction vector is calculated based on the scooter's trajectory data, and the direction of the ellipse's major axis is determined, extending the ROI along the scooter's possible direction of motion. Initial ellipse parameters are set based on the scooter's location coordinates, with a range of 3 meters in front, 2 meters behind, and 1.5 meters on each side, forming the baseline monitoring range. The area within the initial ellipse parameters is scanned for environmental obstacles, identifying fixed obstacles such as curbs and railings to avoid misidentification as pedestrians. The initial ellipse parameters are dynamically adjusted based on the surrounding pedestrian density index, appropriately reducing the range in densely populated areas to improve detection accuracy and expanding the range in open areas to enhance detection coverage, thereby generating an environmentally adaptive boundary. Finally, the environmentally adaptive boundary is divided into three layers: a core area (0.5 meters around the scooter), an interaction area (0.5-2 meters), and a peripheral area (2-3 meters). Different weight coefficients are assigned to different areas to form an elliptical adaptive region of interest. For the constructed adaptive region of interest (ROI), this method uses a two-stream network architecture for pedestrian detection and behavioral feature extraction. This architecture consists of two branches: a spatial stream and a temporal stream. The image of the adaptive ROI is fed into the spatial stream network to extract pedestrian position features, including coordinates, posture, and appearance information. Optical flow is calculated on consecutive frames of the adaptive ROI, analyzing pixel-level motion information. This information is then fed into the temporal stream network to extract pedestrian motion features, such as direction, velocity, and acceleration. Pedestrian position features and motion features are integrated to generate a comprehensive pedestrian behavior feature map. Posture recognition is performed based on the behavior feature map, and pedestrian behaviors are classified into types such as walking, standing, squatting, approaching a scooter, and operating a scooter, generating a behavior type label. A sliding time window is established based on the behavior type label to record the duration of pedestrians' stays within the ROI, distinguishing between brief stops and intentional stays. The pedestrian position features, behavior type label, and stay duration are integrated to generate pedestrian position trajectories, behavior types, and stay duration data.

[0033] The method further combines the scooter's position coordinates, adaptive regions of interest, and pedestrian trajectory to generate a human-scooter interaction intensity score through a multi-dimensional intersection determination mechanism. First, the bounding box intersection ratio of the pedestrian's trajectory and the scooter's position coordinates is calculated to generate a spatial intersection index, quantifying the spatial proximity between the pedestrian and scooter. A distance decay function is applied to the spatial intersection index to give higher weight to close-range interactions, resulting in a distance-weighted intersection value. The pedestrian's stay continuity within the adaptive region of interest is analyzed to assess their stability and generate a temporal persistence index. Posture change sequences are extracted from the pedestrian's trajectory to identify key actions highly associated with scooter use, such as bending, reaching out to touch, and stepping on the pedals. The distance-weighted intersection value, temporal persistence index, and key action recognition results are input into a multi-layer perceptron. This network structure effectively integrates features from different dimensions to output a comprehensive score, namely the human-scooter interaction intensity score.

[0034] Based on the human-vehicle interaction intensity score, the scooter status is classified and determined through a hierarchical decision tree model. First, it is determined whether the scooter's location coordinates are within the designated parking area, and a location status identifier is generated to distinguish between normal parking and abnormal parking. The motion state parameters in the scooter trajectory data are calculated to determine whether the scooter is in a stationary state. The human-vehicle interaction intensity score is compared with the adaptive threshold set based on historical data to generate an interaction validity judgment result. The interaction duration is calculated to determine whether it exceeds the set time window (default 5 minutes). A decision path is constructed based on the location status identifier, stationary state, interaction validity judgment result, and interaction duration. Hierarchical judgment is performed along the decision path to classify the scooter status as "idle", "abnormal parking", "temporary interaction", or "in use", and the corresponding confidence score is calculated, and the scooter idle state and its confidence score are output.

[0035] For example: In an area with a high density of shared scooters, a surveillance camera captured a set of video streams. After preprocessing, the improved YOLOv8 model identified a scooter at image coordinates (320, 450). Tracking its trajectory, it was found that it had been stationary for more than 10 minutes. The system constructed an elliptical adaptive region of interest with the scooter as the center, 3 meters in front, 2 meters in the back, and 1.5 meters on both sides. Within the region of interest, the dual-stream network detected a pedestrian. Its position trajectory showed that it entered the interaction area from the peripheral area, stayed for 15 seconds, and then left. The behavior type label was "wait and see". The spatial intersection index of the pedestrian and the scooter was calculated to be 0.25, the distance weighted intersection value was 0.18, the time continuity index was 0.12, and no key usage actions were detected. The comprehensive calculation resulted in a human-vehicle interaction intensity score of 0.15. This score is lower than the current area adaptive threshold of 0.35, and the scooter is located in the designated parking area, the stationary state is confirmed, and the interaction duration does not exceed the set time window. The hierarchical decision tree model determines that the scooter status is "idle" with a confidence score of 0.92.

[0036] In the embodiments of this application, the intelligent level and practicality of shared scooter idle state detection are significantly improved by innovatively integrating multi-source video stream processing, deep learning target recognition, and artificial intelligence interactive analysis technologies. Its core technical features reflect the unique value and transformative potential of artificial intelligence algorithms in complex urban environment monitoring. First, by collecting multi-source video streams through the urban surveillance camera network and performing resolution normalization, inter-frame difference, and illumination compensation processing, it not only solves the data consistency problem caused by shooting angles, lighting conditions, and resolution differences in traditional video analysis, but also provides high-quality, standardized input data for subsequent deep learning algorithms. This preprocessing step fully demonstrates the key impact of data preprocessing on the performance of artificial intelligence models. Second, the improved YOLOv8 deep learning model greatly improves the accuracy and real-time performance of scooter recognition through multi-scale feature extraction, attention mechanism enhancement, and dedicated anchor frame configuration. In particular, the inter-frame matching and trajectory smoothing achieved by the Kalman filter algorithm not only enhance the stability of target tracking, but also significantly reduce the jitter and error of target positioning, fully demonstrating the intelligent advantages of deep learning algorithms in target detection and motion trajectory analysis. Furthermore, the elliptical adaptive region of interest (ROI) constructed based on scooter position coordinates and trajectory data achieves intelligent perception and dynamic adaptation to urban scenarios through polar coordinate mapping and environmental complexity assessment. This not only accurately locates potential idle areas but also dynamically adjusts the ROI boundaries based on surrounding pedestrian density, demonstrating the unique value of artificial intelligence algorithms in complex environmental perception and adaptive analysis. Furthermore, the application of a dual-stream network architecture for pedestrian detection and behavioral feature extraction achieves precise identification and in-depth understanding of human-vehicle interactions by fusing position features extracted by a spatial stream network with motion features calculated by a temporal stream network. This architecture not only accurately captures the spatial relationship and dynamic changes between pedestrians and scooters, but also provides richer and more reliable semantic information for idle state determination through posture recognition and behavior classification. Finally, through spatial intersection ratio calculation, temporal persistence analysis, posture sequence recognition, and combined with the feature fusion of a multi-layer perceptron, a multi-dimensional, multi-level human-vehicle interaction intensity assessment model was constructed. This not only breaks through the traditional single-dimensional interaction analysis method, but also provides a more comprehensive and intelligent input for the hierarchical decision tree model. Ultimately, it achieves accurate identification and confidence scoring of the idle state of scooters, fully demonstrating the intelligent reasoning and state discrimination capabilities of artificial intelligence algorithms in complex decision-making scenarios.

[0037] In a specific embodiment, the process of executing step S101 may specifically include the following steps:

[0038] The original video stream is obtained through multi-region distributed acquisition nodes, and the original video stream is converted to 1080p resolution to obtain standardized video data;

[0039] Based on the standardized video data, the difference values ​​of adjacent frames are calculated by background difference algorithm to extract the dynamic area sequence;

[0040] For dynamic area sequences, gamma correction and histogram equalization algorithms are used to compensate for illumination anomalies and generate illumination normalized images.

[0041] Adaptively adjust the frame rate of the illumination normalized image according to the scene complexity to obtain a streamlined video sequence;

[0042] The streamlined video sequences of each acquisition node are fused into multiple perspectives through a spatiotemporal registration algorithm to generate preprocessed video data.

[0043] Specifically, raw video streams are acquired through distributed collection nodes across multiple regions. These collection nodes include fixed surveillance cameras and some mobile surveillance equipment located throughout the city, covering major roads, public areas, and designated scooter parking areas. The acquired raw video streams have varying resolutions and frame rates. For example, some older devices may output 720p or even lower resolution video, while newer devices may provide 2K or 4K video streams. To ensure consistency in subsequent processing, all raw video streams are converted to 1080p resolution. Videos below 1080p are upsampled and interpolated, while videos above 1080p are downsampled, bringing all videos to a uniform 1920×1080 pixel resolution to generate standardized video data. A background difference algorithm is then used to calculate the difference between adjacent frames based on this standardized video data. This algorithm uses a mixture Gaussian model to build a background distribution model for each pixel. When the deviation between the pixel value of a new frame and the background model exceeds a threshold, the pixel is identified as a foreground pixel. By calculating the pixel difference between adjacent frames and applying dynamic thresholding, regions containing moving objects can be effectively identified. This step can separate the static background area from the dynamic foreground area, significantly reducing the amount of data that needs to be processed and extracting a sequence of dynamic areas containing potential scooters and pedestrians.

[0044] For the extracted dynamic region sequence, since the original video may have been shot at different times and in different weather conditions, there may be problems such as uneven lighting, underexposure, or overexposure, necessitating compensation for lighting anomalies. Gamma correction adjusts the pixel brightness distribution through power function transformation to correct image exposure issues. For underexposed dark areas, gamma correction with Y < 1 is applied to enhance details; for overexposed areas, correction with Y > 1 is applied to restore details. The histogram equalization algorithm enhances image contrast by redistributing the grayscale distribution of the image, making the image present similar visual features under different lighting conditions. The combined use of these two algorithms effectively solves the detection problems in cloudy, nighttime, or backlit environments, and generates a lighting-normalized image.

[0045] Next, the illumination-normalized image undergoes adaptive frame rate adjustment based on scene complexity. Scene complexity is calculated using multiple metrics, including the number of foreground objects in the image, object density, motion speed, and texture complexity. For highly complex scenes, such as crowded areas or areas of rapid movement, a higher frame rate (e.g., 30 fps) is maintained; for less complex scenes, such as open areas or static scenes, the frame rate is appropriately reduced (e.g., 10 fps). This adaptive adjustment ensures detection accuracy in key scenes while reducing data processing pressure, resulting in a streamlined video sequence. The streamlined video sequences from each acquisition node are then fused across multiple views using a spatiotemporal registration algorithm. This algorithm first determines the spatial correspondence between different viewpoints through feature point matching and constructs a view transformation matrix. It then uses timestamps to align the timing information of different video streams, ensuring temporal consistency across the multi-view data. For scenes where multiple cameras cover the same area, the algorithm performs view selection or information fusion to address view occlusion. Through multi-perspective fusion, the system obtains more comprehensive scene information, effectively solves the blind spot problem of a single perspective, generates complete pre-processed video data, and provides a high-quality data foundation for subsequent scooter detection.

[0046] In a specific embodiment, the process of executing step S102 may specifically include the following steps:

[0047] The preprocessed video data is fed into a multi-scale feature extraction network to generate a scooter feature map;

[0048] Apply the attention mechanism to the scooter feature map to enhance the core part information and form a key area enhanced feature map;

[0049] Based on the key area enhanced feature map, target detection is performed through the scooter-specific anchor frame configuration, and a preliminary detection frame set is output;

[0050] Perform non-maximum suppression on the preliminary detection box set to filter out high-confidence scooter bounding boxes;

[0051] Assign a unique instance ID to the high-confidence scooter bounding box to form the basic location data of the scooter;

[0052] The Kalman filter algorithm is used to perform inter-frame matching and trajectory smoothing on the basic position data of the scooter to obtain the scooter position coordinates and trajectory data.

[0053] Specifically, the preprocessed video data is fed into a multi-scale feature extraction network. The core goal of this network design is to accurately extract scooter features from complex urban surveillance videos. By constructing convolutional layers with different receptive fields, the multi-scale feature extraction network is able to simultaneously capture both the macroscopic outlines and microscopic details of scooter images. Specifically, the network generates multi-level feature maps for the input video frames, including edge, texture, and shape features. These feature maps capture the unique characteristics of scooters at different scales. After generating the feature maps, an attention mechanism is introduced to enhance the core features of the scooter. Through an adaptive weight allocation mechanism, the most critical areas for scooter identification, such as the handlebars, wheels, and body, are highlighted. The attention mechanism dynamically calculates the importance of each feature point, enabling the network to focus on the scooter's signature features while effectively filtering out background noise and irrelevant information. For example, in complex street scenes, the attention mechanism can accurately locate the scooter's key structural features, even in the presence of pedestrians, other vehicles, or complex street backgrounds.

[0054] The network performs object detection using a configuration of anchor boxes specifically designed for scooters. These anchor boxes are carefully tuned based on the typical geometry and size of scooters, covering a wide range of sizes and forms, from small children's scooters to adult shared scooters. A preliminary set of detection boxes is generated by convolution and projection of the key region-enhanced feature map. This process not only considers the overall outline of the scooter but also its subtle variations in different scenes and angles. To improve detection accuracy, the algorithm performs non-maximum suppression on the preliminary set of detection boxes. This technique suppresses overlapping and low-confidence detection boxes, ultimately selecting high-confidence scooter bounding boxes. Non-maximum suppression is essentially an optimization method that removes redundant detections, ensuring that each scooter is detected only once and retaining the highest-confidence detections.

[0055] After completing high-precision detection, the algorithm assigns a unique instance ID to each high-confidence scooter bounding box, forming the scooter's basic position data. This step solves the problem of individual identification when multiple scooters appear simultaneously, laying the foundation for subsequent trajectory tracking. The Kalman filter algorithm is introduced to perform inter-frame matching and trajectory smoothing on the scooter's basic position data. By combining current measurements with historical predictions, the Kalman filter can effectively reduce random noise in the detection process and generate more continuous and accurate scooter position coordinates and trajectory data. In practical applications, this means that even if the video frame is jittery or partially occluded, the algorithm can still accurately track the scooter's movement trajectory.

[0056] Imagine a shared scooter parked at the edge of a busy pedestrian street. Using this detection method, the video analysis system can precisely locate the scooter's location, track whether it has been moved, and identify surrounding pedestrians interacting with the scooter. This precise object detection and trajectory tracking technology provides powerful technical support for urban traffic management and shared scooter operations.

[0057] In a specific embodiment, the process of executing step S103 may specifically include the following steps:

[0058] Establish a polar coordinate system with the center point of the scooter's position coordinates as the origin;

[0059] Calculate the forward direction vector based on the trajectory data and determine the direction of the major axis of the ellipse;

[0060] Set the initial ellipse parameters based on the scooter's position coordinates, with a range of 3 meters in front, 2 meters behind, and 1.5 meters on both sides;

[0061] Scan the initial elliptical parameter area for environmental obstacles and mark the fixed obstacle area;

[0062] Dynamically adjust the initial ellipse parameters according to the surrounding pedestrian density index to generate an environment-adaptive boundary;

[0063] The environment adaptive boundary is divided into a three-layer structure of core area, interaction area and peripheral area to form an elliptical adaptive interest area.

[0064] Specifically, using the center point of the scooter's position coordinates as the origin of the polar coordinate system is a key spatial positioning strategy. Polar coordinate systems describe spatial positions using angles and distances, and this coordinate system can more flexibly capture the environmental characteristics surrounding the scooter. Establishing a polar coordinate system centered on the scooter means that all spatial analysis and region division will be based on the scooter's current position as a reference point, providing a precise spatial benchmark for subsequent environmental analysis. By analyzing the scooter's trajectory data, the algorithm calculates the forward direction vector. This vector is not only a simple direction indicator but also a key basis for determining the direction of the ellipse's major axis. The forward direction vector reflects the scooter's recent movement trend, allowing the adaptive region of interest to dynamically follow the scooter's actual direction of movement. This elliptical region construction based on movement direction allows the region of interest to more accurately capture the spatial environment surrounding the scooter.

[0065] The initial ellipse parameters are set to a carefully designed spatial range: 3 meters in front provides sufficient forward-looking space, 2 meters behind ensures that the scooter's possible parking or retreat areas are captured, and 1.5 meters on each side covers the interactive space around the scooter. These parameters are not randomly selected, but carefully determined based on the typical usage scenarios of shared scooters in cities and the spatial characteristics of sidewalks. Environmental obstacle scanning is a key step in the construction of adaptive regions of interest. By comprehensively scanning the initial ellipse parameter area, the algorithm is able to identify and mark fixed obstacles such as street furniture, roadside parking, and building edges. This obstacle marking not only provides static constraints on the environment but also provides an important spatial reference for subsequent pedestrian interaction analysis.

[0066] The introduction of a surrounding pedestrian density index enables dynamic adjustment of ellipse parameters, generating a truly adaptive boundary. The pedestrian density index analyzes the number, density, and movement characteristics of pedestrians within the region of interest (ROI) to adjust the ROI boundary in real time. In densely populated areas, the ROI automatically shrinks; in sparser areas, it expands appropriately, ensuring that critical interactions around the scooter are always captured. The adaptive boundary is divided into three layers: a core zone, an interaction zone, and a peripheral zone, forming a multi-layered, elliptical adaptive ROI. The core zone focuses on capturing the scooter's immediate surroundings, the interaction zone analyzes potential pedestrian-vehicle interactions, and the peripheral zone provides broader context. This layered design enables the algorithm to analyze scooter usage from different scales and perspectives. Imagine a shared scooter parked at the edge of a slightly crowded sidewalk. This adaptive ROI construction method allows the algorithm to accurately capture the spatial characteristics surrounding the scooter: the core zone may resize slightly based on the presence of pedestrians or street furniture, the interaction zone closely tracks whether pedestrians approach or touch the scooter, and the peripheral zone provides broader context.

[0067] In a specific embodiment, the process of executing step S104 may specifically include the following steps:

[0068] The adaptive region of interest image is input into the spatial flow network to extract pedestrian position features;

[0069] Calculate the optical flow field of the continuous frames of the adaptive interest area, input the time flow network, and extract the pedestrian motion features;

[0070] Fuse pedestrian position features and pedestrian motion features to generate pedestrian behavior feature maps;

[0071] Perform posture recognition based on pedestrian behavior feature mapping and classify to generate behavior type labels;

[0072] A sliding time window is established based on the behavior type label to record the duration of pedestrians staying in the area of ​​interest;

[0073] Pedestrian location features, behavior type labels and stay duration are integrated to obtain pedestrian location trajectory, behavior type and stay time data.

[0074] Specifically, the dual-stream network architecture plays a central role in the pedestrian behavior analysis phase of the idle scooter detection method. First, the image of the adaptive region of interest (ROI) is fed into the spatial stream network, a deep learning network specifically designed to capture spatial dimensional features. Using multiple layers of convolutional neural networks, the spatial stream network accurately extracts the position of pedestrians within the ROI. Each convolutional layer of the network gradually abstracts and refines the spatial features of pedestrians, moving from basic edge and outline information to more complex shape and posture features. This process goes beyond simple object localization and provides a deeper understanding of the pedestrian's specific position and posture in space. The algorithm calculates the optical flow field for consecutive frames within the adaptive ROI. The optical flow field is a mathematical model that describes the motion of objects in an image sequence, capturing the displacement vectors of pixels between consecutive frames. By accurately calculating the pedestrian's movement trajectory between video frames, the temporal stream network extracts the pedestrian's motion features. This step focuses not only on the pedestrian's instantaneous position but also on the continuity and dynamic changes in their motion. For example, distinguishing whether a pedestrian is passing quickly or lingering around a scooter is crucial for determining the scooter's usage status.

[0075] The fusion of pedestrian position and motion features is a key step in analysis. The two-stream network, through a carefully designed fusion strategy, integrates the feature maps of the spatial and temporal streams to generate a comprehensive pedestrian behavior feature map. This feature map not only contains the pedestrian's static position information but also incorporates their dynamic motion characteristics, providing a rich data foundation for subsequent behavior recognition. The fusion process utilizes a multi-layer perceptron or attention mechanism to ensure that features from different dimensions effectively complement each other.

[0076] Based on the generated pedestrian behavior feature map, the algorithm performs posture recognition and classifies it into behavior type labels. Posture recognition technology accurately classifies complex behavioral patterns by analyzing the pedestrian's skeletal key points, relative positions of body parts, and motion sequences. In the scooter detection scenario, this may include different behavior types, such as whether the pedestrian approaches the scooter, touches it, attempts to ride it, or simply passes by. To more accurately capture the interaction between pedestrians and scooters, the algorithm establishes a sliding time window based on the behavior type labels. Sliding time windows are a dynamic monitoring technique that continuously records the duration of pedestrians' stays within an area of ​​interest. This approach goes beyond simply recording duration and better reflects the dynamic process of pedestrian-scooter interaction. By adjusting the sliding window size and step size, the algorithm can flexibly capture interactions of varying durations and frequencies. The algorithm integrates pedestrian location features, behavior type labels, and dwell duration to generate comprehensive pedestrian location trajectory, behavior type, and dwell time data.

[0077] In a specific embodiment, the process of executing step S105 may specifically include the following steps:

[0078] Calculate the bounding box intersection ratio of the pedestrian position trajectory and the scooter position coordinates to generate a spatial intersection index;

[0079] Apply the distance decay function to the spatial intersection index to obtain the distance-weighted intersection value;

[0080] Analyze the continuity of pedestrians' stay in the adaptive region of interest based on their position trajectories and generate a temporal persistence index.

[0081] Extract the posture change sequence from the pedestrian's position trajectory and identify the key actions related to the scooter interaction;

[0082] Input the distance weighted intersection value, time duration index and key action recognition results into the multi-layer perceptron;

[0083] The input features are fused and calculated through a multi-layer perceptron to output the human-vehicle interaction intensity score.

[0084] Specifically, during the pedestrian-scooter interaction intensity analysis phase of the idle scooter detection method, calculating the bounding box intersection ratio is a primary task. By accurately calculating the overlap between the pedestrian's trajectory and the scooter's coordinates, the algorithm generates a spatial intersection index. This index not only reflects the physical spatial relationship between the pedestrian and scooter but also serves as a key indicator of potential interaction. Calculating bounding box intersection involves precise geometric analysis, taking into account the changing positions of the pedestrian and scooter across frames to ensure that the intersection ratio dynamically and accurately reflects the spatial relationship. The introduction of a distance decay function further optimizes the spatial intersection index. This function nonlinearly weights the intersection index based on the actual distance between the pedestrian and scooter. Distance decay means that the weight of the intersection index increases exponentially as the pedestrian gets closer to the scooter; conversely, the weight decays rapidly as the distance increases. This design highlights effective interactions at close range while minimizing the impact of distant, potentially irrelevant spatial overlap.

[0085] Generating a persistence index is another key analysis step. By tracking pedestrians' continuous movements within an adaptive region of interest, the algorithm assesses the stability and persistence of their interactions. This goes beyond simply recording dwell time; it provides a deeper analysis of the dynamics of pedestrian-scooter interactions. The algorithm detects pedestrian movement patterns and distinguishes between different types of interactions, including rapid transits, brief stops, and sustained stays. Each pattern contributes differently to the persistence index.

[0086] Posture change sequence analysis provides deeper behavioral insights for interaction intensity assessment. By extracting key posture changes in pedestrian position trajectories, the algorithm can identify interactions directly related to the scooter. These might include bending over to check the scooter, attempting to touch or move it, or adjusting its position. Each key action is accurately captured and classified, providing rich behavioral features for interaction intensity assessment.

[0087] The multilayer perceptron (MLP) serves as the final feature fusion and decision-making unit, integrating all previously calculated features. The distance-weighted intersection value, the temporal persistence index, and the key action recognition results are input into the MLP. This complex neural network structure is capable of learning and mapping the nonlinear relationships between these multidimensional features. Through multiple layers of nonlinear transformations and feature weight learning, the MLP ultimately outputs a comprehensive score for the intensity of the human-vehicle interaction.

[0088] For example, imagine a shared scooter sitting idle on the sidewalk of a city street. A cyclist approaches the scooter, pauses briefly, and carefully inspects its status. In this scenario, the algorithm captures details such as the increasing intersection of the pedestrian's bounding boxes with the scooter, highlighting the importance of close interaction through the distance decay function. The pedestrian's dwell time and movement trajectory within the area of ​​interest demonstrate sustained and purposeful behavior. Posture analysis captures key movements such as bending over to inspect and potentially touching the handlebars. These complex features are fused through a multi-layer perceptron to generate a score that accurately reflects the intensity of the scooter's interaction with the pedestrian, providing a key basis for determining whether the scooter is idle. This analysis goes beyond technical object detection and provides a deeper understanding of urban mobility and shared transportation usage patterns.

[0089] In a specific embodiment, the process of executing step S106 may specifically include the following steps:

[0090] Determine whether the scooter's location coordinates are within the designated parking area and generate a location status identifier;

[0091] Calculate motion state parameters based on trajectory data to determine the scooter's stationary state;

[0092] Compare the human-vehicle interaction intensity score with the adaptive threshold to generate the interaction effectiveness judgment result;

[0093] Calculate the interaction duration based on the interaction effectiveness determination result to determine whether it exceeds the set time window;

[0094] Construct a decision path based on location status identification, static state, interaction effectiveness determination results, and interaction duration;

[0095] Hierarchical judgment is performed along the decision path, outputting the scooter idle state and its confidence score.

[0096] Specifically, in the final determination phase of the idle scooter detection method, generating a location status identifier is the primary task. By precisely comparing the scooter's location coordinates with pre-defined parking area boundaries, the scooter is determined to be within the specified parking range. This step goes beyond simple coordinate matching and also considers the geographic boundary characteristics of the parking area, including road markings, lined areas, and the specific layout of urban infrastructure. The location status identifier accurately classifies the scooter's location, providing a critical spatial reference for subsequent idle status determination. Calculating motion state parameters is the core step in determining the scooter's stationary state. By analyzing scooter trajectory data, the algorithm extracts a series of kinematic features, including position change rate, acceleration, and angular change. These parameters not only reflect the scooter's instantaneous state but also reveal its long-term motion patterns. Determining the stationary state requires complex threshold design to distinguish between true stationary state and minor position fluctuations. The algorithm comprehensively considers the amplitude, frequency, and duration of position deviations to ensure the accuracy of the stationary state determination.

[0097] Comparing the human-scooter interaction intensity score with the adaptive threshold is a key step in the judgment process. The adaptive threshold is not fixed but dynamically adjusts based on environmental complexity, time period, and historical data. When the human-scooter interaction intensity score exceeds the threshold, a positive judgment is generated regarding the interaction's validity; otherwise, it is considered invalid. This dynamic threshold mechanism adapts to the complexities of scooter use in different scenarios and avoids misjudgments caused by simple static thresholds.

[0098] Interaction duration calculation further refines the interaction effectiveness assessment. Based on the interaction effectiveness determination results, the algorithm accurately measures the duration of pedestrian-scooter interactions. The set time window is not a simple fixed value, but is dynamically adjusted based on scenario characteristics and historical data. Interactions exceeding the time window are considered more meaningful potential usage behaviors, providing a more detailed temporal analysis for determining whether the scooter is idle. Decision path construction is the integration point of the entire decision process. Location status identification, stationary state, interaction effectiveness determination results, and interaction duration are integrated into a multi-dimensional decision matrix. This complex decision path is not a simple linear decision, but a multi-layered, multi-conditional decision network. The characteristics of each dimension have varying degrees of influence on the final idle state determination, ensuring the comprehensiveness and accuracy of the determination results. Hierarchical determination is a key step in the final idle state output. The algorithm follows the pre-constructed decision path, performing detailed state assessments layer by layer. Each layer of determination gradually converges to the final idle state conclusion based on the weighting of the characteristics of different dimensions. The confidence score is a key output of this process, not only providing a judgment on whether the scooter is idle but also quantifying the reliability of this judgment.

[0099] The above describes the idle scooter detection method in the embodiment of the present application. The following describes the idle scooter detection system in the embodiment of the present application. Figure 2 In one embodiment of the present application, an idle scooter detection system includes:

[0100] An acquisition module 201 is configured to acquire multi-source video streams through a city surveillance camera network, perform resolution normalization, frame difference and illumination compensation processing on the multi-source video streams, and obtain pre-processed video data;

[0101] Identification module 202, configured to perform scooter identification and time series association based on the pre-processed video data by improving the YOLOv8 deep learning model to obtain scooter position coordinates and trajectory data;

[0102] An evaluation module 203 is configured to construct an elliptical adaptive region of interest based on the scooter position coordinates and trajectory data through polar coordinate mapping and environment complexity evaluation;

[0103] Extraction module 204, configured to perform pedestrian detection and behavior feature extraction on the adaptive interest region through a dual-stream network architecture to obtain pedestrian location trajectory, behavior type, and dwell time data;

[0104] An analysis module 205 is configured to combine the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory to obtain a human-vehicle interaction intensity score through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition;

[0105] The determination module 206 is configured to classify and determine the scooter state according to the human-vehicle interaction intensity score through a hierarchical decision tree model to obtain the scooter idle state and its confidence score.

[0106] Through the collaborative efforts of these components, and through the innovative integration of multi-source video stream processing, deep learning object recognition, and AI interactive analysis technologies, the intelligent level and practicality of shared scooter idle state detection have been significantly improved. Its core technical features demonstrate the unique value and transformative potential of AI algorithms in monitoring complex urban environments. First, by collecting multi-source video streams from a network of urban surveillance cameras and performing resolution normalization, inter-frame differencing, and illumination compensation, this not only addresses the data consistency issues inherent in traditional video analysis due to differences in shooting angles, lighting conditions, and resolution, but also provides high-quality, standardized input data for subsequent deep learning algorithms. This preprocessing step fully demonstrates the critical impact of data preprocessing on the performance of AI models. Second, the improved YOLOv8 deep learning model significantly improves the accuracy and real-time performance of scooter recognition through multi-scale feature extraction, enhanced attention mechanisms, and dedicated anchor frame configuration. In particular, the inter-frame matching and trajectory smoothing implemented by the Kalman filter algorithm not only enhance the stability of target tracking but also significantly reduce jitter and error in target positioning, fully demonstrating the intelligent advantages of deep learning algorithms in target detection and motion trajectory analysis. Furthermore, the elliptical adaptive region of interest (ROI) constructed based on scooter position coordinates and trajectory data achieves intelligent perception and dynamic adaptation to urban scenarios through polar coordinate mapping and environmental complexity assessment. This not only accurately locates potential idle areas but also dynamically adjusts the ROI boundaries based on surrounding pedestrian density, demonstrating the unique value of artificial intelligence algorithms in complex environmental perception and adaptive analysis. Furthermore, the application of a dual-stream network architecture for pedestrian detection and behavioral feature extraction achieves precise identification and in-depth understanding of human-vehicle interactions by fusing position features extracted by a spatial stream network with motion features calculated by a temporal stream network. This architecture not only accurately captures the spatial relationship and dynamic changes between pedestrians and scooters, but also provides richer and more reliable semantic information for idle state determination through posture recognition and behavior classification. Finally, through spatial intersection ratio calculation, temporal persistence analysis, posture sequence recognition, and combined with the feature fusion of a multi-layer perceptron, a multi-dimensional, multi-level human-vehicle interaction intensity assessment model was constructed. This not only breaks through the traditional single-dimensional interaction analysis method, but also provides a more comprehensive and intelligent input for the hierarchical decision tree model. Ultimately, it achieves accurate identification and confidence scoring of the idle state of scooters, fully demonstrating the intelligent reasoning and state discrimination capabilities of artificial intelligence algorithms in complex decision-making scenarios.

[0107] above Figure 2 The idle scooter detection system in the embodiment of the present invention is described in detail from the perspective of modular functional entities. The idle scooter detection device in the embodiment of the present invention is described in detail from the perspective of hardware processing.

[0108] Figure 3FIG3 is a schematic diagram of the structure of an idle scooter detection device provided by an embodiment of the present invention. The idle scooter detection device 300 may vary significantly depending on configuration or performance. The device may include one or more central processing units (CPUs) 310 (e.g., one or more processors), a memory 320, and one or more storage media 330 (e.g., one or more mass storage devices) storing application programs 333 or data 332. The memory 320 and storage media 330 may be either transient or persistent storage. The program stored in the storage medium 330 may include one or more modules (not shown), each of which may include a series of instruction operations on the idle scooter detection device 300. Furthermore, the processor 310 may be configured to communicate with the storage medium 330, and the idle scooter detection device 300 may execute the series of instruction operations stored in the storage medium 330 to implement the steps of the idle scooter detection method described above.

[0109] The idle scooter detection device 300 may further include one or more power supplies 340, one or more wired or wireless network interfaces 350, one or more input and output interfaces 360, and / or one or more operating systems 331, such as Windows Server, Mac OS X, Unix, Linux, FreeBSD, etc. It will be appreciated by those skilled in the art that Figure 3 The structure of the idle scooter detection device shown does not constitute a limitation on the idle scooter detection device provided by the present invention, and may include more or fewer components than shown in the figure, or combine certain components, or arrange the components differently.

[0110] The present invention also provides a computer-readable storage medium, which may be a non-volatile computer-readable storage medium or a volatile computer-readable storage medium. The computer-readable storage medium stores instructions. When the instructions are executed on a computer, the computer executes the steps of the idle scooter detection method.

[0111] Those skilled in the art will clearly understand that, for the convenience and brevity of description, the specific working processes of the above-described systems, systems and units can refer to the corresponding processes in the aforementioned method embodiments and will not be repeated here.

[0112] If the integrated unit is implemented in the form of a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, or all or part of the technical solution can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions for enabling an idle scooter detection device (which can be a personal computer, server, or network device, etc.) to execute all or part of the steps of the method described in each embodiment of the present invention. The aforementioned storage medium includes: U disk, mobile hard disk, read-only memory (ROM), random access memory (RAM), magnetic disk or optical disk, etc., various media that can store program code.

[0113] The above embodiments are only used to illustrate the technical solutions of the present invention, rather than to limit the same. Although the present invention has been described in detail with reference to the aforementioned embodiments, those skilled in the art should understand that they can still modify the technical solutions described in the aforementioned embodiments, or make equivalent replacements for some of the technical features therein. However, these modifications or replacements do not deviate the essence of the corresponding technical solutions from the spirit and scope of the technical solutions of the various embodiments of the present invention.

Claims

1. A method for detecting an idle scooter, characterized in that: The method comprises: Collecting multi-source video streams through a city surveillance camera network, performing resolution standardization, inter-frame difference and illumination compensation processing on the multi-source video streams to obtain pre-processed video data; Based on the pre-processed video data, scooter identification and time series association are performed by improving the YOLOv8 deep learning model to obtain the scooter position coordinates and trajectory data; Based on the scooter position coordinates and trajectory data, an elliptical adaptive region of interest is constructed through polar coordinate mapping and environment complexity assessment; For the adaptive region of interest, pedestrian detection and behavior feature extraction are performed through a dual-stream network architecture to obtain pedestrian location trajectory, behavior type and residence time data; Combining the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory, a human-vehicle interaction intensity score is obtained through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition; According to the human-vehicle interaction intensity score, the scooter status is classified and determined through a hierarchical decision tree model to obtain the scooter idle status and its confidence score.

2. The idle scooter detection method according to claim 1, characterized in that: The method includes collecting multi-source video streams through a city surveillance camera network, performing resolution standardization, inter-frame difference and illumination compensation processing on the multi-source video streams to obtain pre-processed video data, including: Obtaining the original video stream through multi-region distributed acquisition nodes, and performing 1080p resolution conversion processing on the original video stream to obtain standardized video data; According to the standardized video data, the difference values ​​of adjacent frames are calculated by a background difference algorithm to extract a dynamic area sequence; For the dynamic region sequence, compensation for illumination anomalies is performed through gamma correction and histogram equalization algorithms to generate an illumination normalized image; Adaptively adjusting the frame rate of the illumination normalized image according to scene complexity to obtain a streamlined video sequence; The simplified video sequence of each acquisition node is subjected to multi-view fusion through a spatiotemporal registration algorithm to generate the pre-processed video data.

3. The idle scooter detection method according to claim 1, characterized in that: The method of performing scooter identification and time series association based on the pre-processed video data by improving the YOLOv8 deep learning model to obtain the scooter position coordinates and trajectory data includes: Inputting the preprocessed video data into a multi-scale feature extraction network to generate a scooter feature map; Applying an attention mechanism to the scooter feature map to enhance core part information and form a key area enhanced feature map; According to the key area enhanced feature map, target detection is performed through the scooter-specific anchor frame configuration, and a preliminary detection frame set is output; Performing non-maximum suppression on the preliminary detection frame set to screen out a high-confidence scooter bounding box; Assigning a unique instance ID to the high-confidence scooter bounding box to form basic scooter position data; The Kalman filter algorithm is used to perform inter-frame matching and trajectory smoothing on the basic position data of the scooter to obtain the position coordinates and trajectory data of the scooter.

4. The idle scooter detection method according to claim 1, characterized in that: The method of constructing an elliptical adaptive region of interest based on the scooter position coordinates and trajectory data through polar coordinate mapping and environment complexity assessment includes: Establishing a polar coordinate system with the center point of the scooter position coordinate as the origin; Calculate the forward direction vector according to the trajectory data to determine the direction of the major axis of the ellipse; Based on the scooter's position coordinates, the initial ellipse parameters are set to 3 meters in front, 2 meters behind, and 1.5 meters on both sides; Scanning the initial elliptical parameter area for environmental obstacles and marking fixed obstacle areas; Dynamically adjusting the initial ellipse parameters according to the surrounding pedestrian density index to generate an environment-adaptive boundary; The environment adaptive boundary is divided into a three-layer structure of a core area, an interactive area and a peripheral area to form the elliptical adaptive interest area.

5. The idle scooter detection method according to claim 1, characterized in that: The method uses a dual-stream network architecture to perform pedestrian detection and behavior feature extraction for the adaptive region of interest, and obtains pedestrian location trajectory, behavior type, and dwell time data, including: Inputting the adaptive region of interest image into a spatial stream network to extract pedestrian position features; Performing optical flow calculation on the continuous frames of the adaptive region of interest, inputting the optical flow field into a temporal flow network, and extracting pedestrian motion features; fusing the pedestrian position features and the pedestrian motion features to generate a pedestrian behavior feature map; Performing posture recognition based on the pedestrian behavior feature map and classifying and generating behavior type labels; Establishing a sliding time window based on the behavior type label to record the duration of pedestrians staying in the area of ​​interest; The pedestrian location features, the behavior type label and the stay duration are integrated to obtain the pedestrian location trajectory, behavior type and stay time data.

6. The idle scooter detection method according to claim 1, characterized in that: The combination of the scooter position coordinates, the adaptive interest region, and the pedestrian position trajectory, through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition, obtains a human-vehicle interaction intensity score, including: Calculating the bounding box intersection ratio of the pedestrian position trajectory and the scooter position coordinates to generate a spatial intersection index; Applying a distance decay function to the spatial intersection index to obtain a distance-weighted intersection value; Analyzing the pedestrian's stay continuity within the adaptive area of ​​interest based on the pedestrian's position trajectory to generate a time continuity index; Extracting a posture change sequence from the pedestrian's position trajectory and identifying key actions related to interaction with the scooter; Inputting the distance weighted intersection value, the time duration index and the recognition result of the key action into a multi-layer perceptron; The multi-layer perceptron performs fusion calculation on the input features and outputs the human-vehicle interaction intensity score.

7. The idle scooter detection method according to claim 1, characterized in that: The scooter status is classified and determined by a hierarchical decision tree model based on the human-vehicle interaction intensity score to obtain the scooter idle status and its confidence score, including: Determining whether the scooter's position coordinates are within a designated parking area and generating a position status identifier; Calculating motion state parameters based on the trajectory data to determine the stationary state of the scooter; Comparing the human-vehicle interaction intensity score with an adaptive threshold to generate an interaction effectiveness determination result; Calculate the interaction duration based on the interaction validity determination result to determine whether it exceeds the set time window; Constructing a decision path based on the location state identifier, the stationary state, the interaction validity determination result, and the interaction duration; A hierarchical decision is performed along the decision path to output the scooter idle state and its confidence score.

8. An idle scooter detection system, characterized in that: For implementing the idle scooter detection method according to any one of claims 1 to 7, the idle scooter detection system comprises: An acquisition module is used to collect multi-source video streams through a city surveillance camera network, and perform resolution normalization, inter-frame difference and illumination compensation processing on the multi-source video streams to obtain pre-processed video data; An identification module is used to perform scooter identification and time series association based on the preprocessed video data by improving the YOLOv8 deep learning model to obtain the scooter position coordinates and trajectory data; An evaluation module, configured to construct an elliptical adaptive region of interest based on the scooter position coordinates and trajectory data through polar coordinate mapping and environment complexity evaluation; An extraction module is used to detect pedestrians and extract their behavior features using a dual-stream network architecture for the adaptive region of interest, thereby obtaining pedestrian location trajectory, behavior type, and dwell time data. An analysis module, configured to combine the scooter position coordinates, the adaptive region of interest, and the pedestrian position trajectory to obtain a human-vehicle interaction intensity score through spatial intersection ratio calculation, time persistence analysis, and posture sequence recognition; The determination module is used to classify and determine the scooter state according to the human-vehicle interaction intensity score through a hierarchical decision tree model to obtain the scooter idle state and its confidence score.

9. An idle scooter detection device, characterized in that: The method comprises a memory and a processor, wherein the memory stores a computer program that can be run on the processor, and when the processor executes the computer program, the idle scooter detection method according to any one of claims 1 to 7 is implemented.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that: When the computer program is executed by a processor, the processor is caused to perform the idle scooter detection method according to any one of claims 1 to 7.

Images