Inspection bicycle system based on multi-sensor fusion

Abstract

The invention discloses a bicycle inspection system and method based on multi-sensor fusion, and relates to the technical field of smart city sensing equipment and city management informatization. The system comprises a mobile acquisition module, a data processing and event judging module and an event grading and responding module, the mobile acquisition module integrates an image acquisition device, an ultrahigh frequency RFID reader-writer, a millimeter wave radar and a vehicle-mounted control unit, and is carried on a bicycle platform to realize multi-modal data synchronous acquisition. The vehicle-mounted control unit guarantees that the timestamp deviation of the three types of signals is smaller than or equal to 100 microseconds through the PTP precision timing module, and data are uploaded to the cloud through the wireless communication module. The data processing and event judgment module executes four-order fusion processing, calculates comprehensive confidence coefficients according to 35%, 40% and 25% weights, and generates a structured city management event; and the event grading and response module allocates response levels according to confidence and triggers corresponding disposal processes. The system solves the problems of insufficient coverage, low identification precision, weak ownership association and the like of an existing inspection scheme, has the advantages of high maneuverability, high reliability and full-process automation, can be widely applied to the scenes of municipal facility inspection, standard management of shared bicycles and the like, and greatly improves the fine and intelligent level of urban management.

Description

Technical Field

[0001] This invention relates to the field of smart city sensing equipment and urban management information technology, and in particular to a lightweight inspection bicycle system and method based on the fusion of multiple sensors including image, UHF RFID and millimeter-wave radar. Background Technology

[0002] As my country's urbanization process continues to deepen, cities are expanding in scale and becoming increasingly complex, placing unprecedented demands on the refinement, intelligence, and routine operation and management of cities. The effectiveness of urban operation and management inspections is crucial to the health, safety, and orderly operation of cities.

[0003] However, the traditional manual inspection model is facing serious challenges: on the one hand, the management elements are complex, ranging from road manhole covers to overhead cables, from urban order to facility safety, highlighting the contradiction between heavy tasks and limited human resources, inevitably leading to blind spots in inspections; on the other hand, recurring problems make it difficult to achieve a fundamental shift from passive handling to proactive prevention. Furthermore, existing intelligent inspection equipment generally suffers from poor environmental adaptability and crude event judgment.

[0004] Patent CN110378352A discloses a four-wheeled electric inspection vehicle equipped with a high-definition camera and a GPU edge computing unit, which uses the YOLOv5 model to identify targets such as missing manhole covers, road damage, and illegal advertisements. However, this patent relies on a single visual modality, and its accuracy drops sharply to below 62% under conditions of rain, fog, nighttime, strong glare, or obstructions such as green belts or parked vehicles. Furthermore, it cannot obtain the unique identifier and ownership information of the target, only outputting "a damaged manhole cover exists at a certain location," without associating it with specific road segment numbers, property owners, and maintenance personnel, leading to delays in dispatching incident work orders and unclear responsibilities. The four-wheeled vehicle is large, with a turning radius greater than 3.5 m, making it difficult to enter old alleyways and non-motorized vehicle lanes with a width of less than 2.5 m, resulting in a coverage rate of less than 45% of the total length of the urban road network.

[0005] Patent CN109886242B proposes deploying UHF RFID tags on facilities such as bus stops, trash cans, and fire hydrants. Inspection personnel use handheld readers to scan along the route, and the facility status is obtained by matching the tag ID with the backend database. This inspection method relies entirely on RFID. When the tags are damaged, obscured, or detached, the system cannot determine whether the physical facility still exists at that location, which can easily lead to missed inspections. It also lacks target positioning and motion perception, and cannot distinguish between "stationary and intact facilities" and "facilities temporarily obscured by vehicles".

[0006] Current technologies have not yet solved the core problem of how to simultaneously acquire three-dimensional information of urban management objectives using a low-cost, highly mobile platform, and automatically generate executable, traceable, and hierarchical structured event work orders. Summary of the Invention

[0007] The purpose of this invention is to provide a reliable, accurate, and intelligent bicycle inspection system and method to address the problems mentioned in the background art.

[0008] The above-mentioned objective of the present invention is achieved through the following technical solution: a multi-sensor fusion-based inspection bicycle system, characterized in that it includes a mobile acquisition module, a data processing and event determination module, and an event classification and response module; The mobile acquisition module, which uses multiple sensors to simultaneously acquire multimodal data, includes an image acquisition device, an ultra-high frequency RFID reader, radar, and a vehicle control unit, and is installed on a bicycle. The image acquisition device is fixedly installed at the front of the bicycle and acquires visual data of the environment in front or to the side. The RFID reader is fixedly installed on the front or middle part of the bicycle frame and connected to a high-frequency or ultra-high-frequency antenna to read RFID tag information deployed on urban vehicles or public facilities. The radar is fixedly installed at the front of the bicycle, adjacent to the image acquisition device, together covering the detection area in front, and measuring the distance, speed and angle information of the target. The vehicle control unit is communicatively connected to the image acquisition device, the RFID reader and the radar, and synchronously receives and temporarily stores the visual data, RFID tag information and radar data. The event classification and response module is used to assign response levels to the urban management events according to preset rules and automatically trigger the corresponding handling process.

[0009] Preferably, the RFID reader is an ultra-high frequency reader, whose radio frequency field covers the main field of view of the image acquisition device, and reads RFID tag information in batches within the field of view during movement; The image acquisition device and the millimeter-wave radar have a lateral deviation of ≤2 cm and a pitch angle deviation of ≤0.5°, ensuring the engineering basis for a field-of-view overlap of ≥90%.

[0010] Preferably, the vehicle control unit has a built-in PTP precision timing module to synchronize the image acquisition frame start signal, RFID command trigger signal, and radar sampling trigger signal, with a timestamp deviation of ≤100 μs.

[0011] Preferably, the UHF RFID reader operates in the frequency band of 902–928 MHz, and the effective reading radius of the radio frequency field is 3.0–3.5 m in the moving state, which meets the requirements for batch and no-missed reading of tags within the main field of view of the image; The millimeter-wave radar operates in the 76–81 GHz frequency band, with a range accuracy of ±0.1 m, an angular resolution of ±1°, and a point cloud output frame rate of ≥10 Hz, supporting continuous tracking of dynamic targets.

[0012] Preferably, the vehicle control unit and the data processing and event determination module interact via a wireless communication network; the data processing and event determination module and the event classification and response module are deployed on a remote server or cloud platform.

[0013] Preferably, the data processing and event determination module is deployed on a remote server or cloud platform to receive synchronized visual data, RFID tag information, and radar data, and perform the following fourth-order fusion processing: First stage: Using image recognition results as a trigger, extract the pixel coordinates, size, color histogram, and YOLOv8s confidence of the potential event target; The second step involves calling the RFID-ID mapping database and returning management attributes such as EPC code, affiliated unit, and maintenance contact person based on pixel coordinate matching, thereby realizing semantic association between form and identity. The third stage involves using target distance, velocity, and azimuth data measured by millimeter-wave radar to perform status verification on the image-RFID matching results. Fourth stage: Calculate the overall confidence level using a weighted method: image attributes contribute 35%, RFID matching degree contributes 40%, and radar status consistency contributes 25%. The overall confidence level serves as the core field for structured urban management events. The event classification and response module is deployed on a remote server or cloud platform to allocate response levels one to four based on the overall confidence level and event type, and automatically trigger the corresponding handling process.

[0014] The corresponding handling process is as follows: Level 1 response involves the 110 command center; Level 2 response generates a work order and dispatches it to the responsible unit; Level 3 response is included in the daily patrol plan; and Level 4 response is only archived for future reference.

[0015] Preferably, when the data processing and event determination module performs fusion analysis, it identifies potential event targets and their image attributes through the visual data; obtains the identification information or management attributes of the potential event targets through the RFID information; and matches and associates the image attributes with the identification information or management attributes to generate structured urban management events.

[0016] A method for inspecting bicycles based on multi-sensor fusion, characterized by the following steps: S1. Acquire video streams using an image acquisition device, identify potential event targets using the YOLOv8s model, and extract pixel coordinates, image attributes, and confidence levels; S2. Control the UHF RFID reader to scan in the coordinate area obtained in S1, and attempt to obtain RFID tag information associated with the target; S3. Simultaneously acquire range, velocity, and azimuth data of the target from the millimeter-wave radar; S4. Perform fourth-order fusion: (a) The first-order trigger is based on the image recognition result; (b) Secondary liability is determined based on RFID matching results; (c) Use radar status data as the third-order verification; (d) Calculate the overall confidence level using a weighting of 35%+40%+25% to generate structured urban management events; S5. Based on the confidence level and event type obtained in S4, the event classification and response module assigns the response level; S6. Automatically execute corresponding processing procedures, including SMS push, voice reminder, work order dispatch, GIS heat map update or archiving for future reference.

[0017] Preferably, when at least one sensor in the mobile acquisition module identifies a potential event target, the image attributes are matched and associated with the identification information or management attributes to generate a structured urban management event. When the image acquisition device identifies a potential event target, it acquires the target's location and image features; controls the RFID reader to perform directional reading in the area where the potential event target is located, attempting to acquire RFID tag information associated with the target; simultaneously, it acquires the radar's ranging and speed measurement data of the potential event target; based on at least two of the image features, RFID tag information, and radar data, it performs cross-validation and attribute supplementation on the potential event target to determine the type, attributes, and confidence level of the urban management event.

[0018] The beneficial effects of this invention are as follows: 1. In view of the problem that existing four-wheeled inspection vehicles have difficulty entering narrow areas, resulting in insufficient inspection coverage, this system adopts a bicycle platform, which can enter old alleys with a width of less than 2.2 meters, greatly increasing the inspection coverage and reducing blind spots.

[0019] To address the issues of low recognition rate of single visual sensors in environments such as rain and fog, and the missed detection problems of pure RFID solutions, this invention improves the overall event recognition accuracy to 96.7% and reduces the false alarm rate by fusing three-modal data.

[0020] To address the problem of delays in work order dispatch caused by the inability of existing technologies to link facility identity and ownership information, this invention uses forced association between images and RFID to automatically generate structured events containing information on the property owner and responsible person, achieving 100% accuracy in ownership matching and 0% error rate in work order dispatch. Attached Figure Description

[0021] Figure 1 This is a flowchart of the method of the present invention. Table 1: Performance Comparison of the Invention Scheme with Comparative Examples 1 and 2 Detailed Implementation

[0022] The present invention will be further described in detail below with reference to the accompanying drawings.

[0023] This specific embodiment is merely an explanation of the present invention and is not intended to limit the invention. After reading this specification, those skilled in the art can make modifications to this embodiment without contributing any inventive step, but such modifications are protected by patent law as long as they are within the scope of the claims of the present invention.

[0024] Example 1: 1. System Deployment The inspection bicycle is equipped with a Hikvision DS-2CD3T47WDA3-L 4K high-definition infrared camera fixed at the front, supporting 1080P@60fps real-time acquisition, a 120° field of view, and infrared supplementary lighting for nighttime operation, making it suitable for all-weather inspections. It is paired with a Continental AR1221-BA00 77GHz millimeter-wave radar, with the lateral deviation between the optical center and beam center adjusted to 1.2cm and the elevation angle deviation adjusted to 0.3°, achieving a 95% field-of-view overlap. An Impinjee R420 UHF RFID reader is installed in the middle of the bicycle, equipped with an Anxinjie AX-915-8dBi directional flat panel antenna, covering the camera's main field of view with a radio frequency field. The effective reading radius is 3.2m in motion, and it supports the ISO 18000-6C protocol.

[0025] The vehicle control unit uses an Advantech UNO-2484G embedded industrial computer, equipped with an Intel Celeron N5105 processor, 8GB DDR4 memory, and 128GB SSD. It has a built-in Renesas PT7C4339 PTP precision timing module to synchronize the camera frame start signal, RFID command trigger signal, and radar sampling trigger signal, with the timestamp deviation of the three controlled within 60μs. The industrial computer integrates a Huawei ME909s-821 4G / 5G dual-mode module to achieve high-speed data upload to the cloud.

[0026] The municipal management department pre-attaches ImpinjMonza R6 UHF RFID tags (512 bits storage capacity, reading distance up to 5m, operating temperature -40℃~85℃) to every manhole cover, lamppost, and traffic sign within its jurisdiction. The tags contain management attribute information such as the facility's EPC code, the road section it belongs to, the installation time, the maintenance responsibility unit, and the contact person's phone number. An RFID-ID mapping database is also established and deployed on an Alibaba Cloud ECS c5.large instance server. The data processing and event judgment module, as well as the event classification and response module, are all deployed on the municipal management cloud server, relying on NVIDIA T4 GPUs for YOLOv8s model inference acceleration.

[0027] 2. Inspection Process S1. Inspection personnel ride inspection bicycles along a preset route. Hikvision DS-2CD3T47WDA3-L cameras collect real-time video streams of road facilities ahead. The YOLOv8s model built into the vehicle-mounted Advantech UNO-2484G industrial control computer identifies potential event targets. When abnormal conditions such as missing manhole covers, tilted street light poles, or damaged signs are detected, the target pixel coordinates, size, color histogram, and YOLOv8s confidence level are extracted (e.g., the confidence level for identifying missing manhole covers is 92%).

[0028] S2. The vehicle-mounted Advantech UNO-2484G industrial control computer controls the Impinj R420 RFID reader to scan the target pixel coordinates output by S1 in the coordinate area, read the Impinj Monza R6 tag information corresponding to the missing manhole cover, and match the road section to which the manhole cover belongs, "XX Avenue K3+200", the maintenance unit is "Municipal Engineering Co., Ltd.", and the contact person's phone number is "XXX" through the mapping database. The tag reading response time is ≤50ms.

[0029] S3. The mainland AR1221-BA00 millimeter-wave radar simultaneously acquired distance, speed, and azimuth data for the area with the missing manhole cover. The target distance to the inspection bicycle was measured to be 5.8m, the azimuth angle to be 2.1°, and the speed to be 0km / h (confirmed to be static). There were no dynamic objects obstructing the view. The radar point cloud output frame rate was 15Hz, which meets the requirements for continuous monitoring.

[0030] S4. The vehicle-mounted industrial control computer uploads the synchronized visual data, RFID tag information, and radar data to the cloud data processing and event determination module via the Huawei ME909s-821 module, where it performs fourth-order fusion processing: (a) Using an image recognition confidence level of 92% as the first-order trigger criterion; (b) Using a 100% RFID tag matching success rate as the basis for second-level liability attribution, the main body responsible for facility management shall be clearly identified; (c) Using radar data to confirm that the target is statically abnormal and unobstructed as the basis for the third-order verification; (d) Calculate the overall confidence level by weight: 92%×35% + 100%×40% + 100%×25% = 97.2%, generate a structured urban management event, and the event type is "missing manhole cover".

[0031] S5. The event classification and response module determines the "missing manhole cover" event as a Level 1 response event (highest priority) based on preset rules, with a comprehensive confidence level ≥ 90%.

[0032] S6. The system automatically triggers the Level 1 response procedure: immediately sends a warning SMS to the contact person of the maintenance unit, pushes a voice reminder to the inspection team's dispatch center, automatically generates an electronic work order through the Fanwei e-cology OA system and dispatches it to the mobile terminal of the maintenance personnel, and synchronously updates the fault location in the urban facility GIS heat map built on ArcGIS Server.

[0033] Application effect By fusing three-modal data from images, UHF RFID, and millimeter-wave radar, combined with YOLOv8s model inference and a fourth-order fusion algorithm, the overall fault identification accuracy reaches 99.6%, with a fault false detection rate as low as 0.4%. Even under adverse conditions such as nighttime or light rain, the identification accuracy remains at 98.9%, solving the technical bottleneck of a single-vision inspection vehicle experiencing a sharp drop in identification rate (only 62.5%) in complex environments.

[0034] Example 2: Inspection and monitoring of standardized parking of shared bicycles in the city This embodiment discloses the application of a multi-sensor fusion-based inspection bicycle system in the standardized parking management of shared bicycles in cities. The inspection bicycles are deployed in the special inspection teams of urban traffic management departments for shared bicycles, and are used to automatically check and deal with violations such as disorderly parking and obstruction of blind paths of shared bicycles in key areas such as urban business districts, subway stations, and bus stops.

[0035] 1. System Deployment The inspection bicycle is equipped with a Dahua DH-IPC-HFW5449M-I2 wide-angle HD camera (image acquisition device) at the front of the frame. It has a 110° field of view, supports 4MP@30fps acquisition, and features strong light suppression. It is paired with a Bosch LRR3-76GHz 76GHz millimeter-wave radar. The lateral deviation between the optical center and beam center of the two devices is 1.8cm, the elevation angle deviation is 0.4°, the field of view overlap is 92%, the radar range accuracy is ±0.1m, the angular resolution is ±1°, and the point cloud output frame rate is 12Hz. The front of the vehicle is equipped with a Shenzhen Quanshunhong F800 UHF RFID reader (operating frequency band 902-928MHz, supporting simultaneous reading of multiple tags, with a maximum reading rate of 200 tags / second), and a Quanshunhong QS-HU900 circularly polarized antenna, with an effective reading radius of 3.0m, covering the shared bicycle parking area within the main field of view of the camera. The on-board control unit uses an Advantech M50 embedded industrial computer (Intel Core i3-10110U processor, 16GB DDR4 memory, 256GB NVMe SSD), and has a built-in Xibashi PT82B715 PTP precision timing module, which controls the time stamp deviation of image, RFID, and radar signals to within 80μs, and integrates a Quectel EC200S 4G module for data transmission.

[0036] Shared bicycle operators pre-equip each shared bicycle with a Fudan Micro FM11NC08 UHF RFID tag (passive design, 100,000 read / write cycles, 25mm×25mm size, suitable for harsh outdoor environments). The tag contains information such as the vehicle's EPC code, the company to which it belongs, the vehicle number, and the deployment area. The city's traffic management department establishes a shared bicycle RFID-ID mapping database and connects it to a Tencent Cloud CVM S6 instance server. The data processing and event judgment module, as well as the event classification and response module, are deployed on the traffic management department's cloud server, using an Intel Xeon E5-2680 v4 processor to ensure data processing efficiency.

[0037] 2. Inspection Process S1. Inspection personnel patrol the area around the subway station exit using inspection bicycles. Dahua DH-IPC-HFW5449M-I2 wide-angle high-definition cameras collect real-time video streams of the shared bicycle parking area. The YOLOv8s model identifies multiple shared bicycles illegally parked on the blind path, and extracts the target pixel coordinates, number of vehicles, parking angle, and confidence level (e.g., the confidence level for identifying obstruction of the blind path is 89%).

[0038] S2. The vehicle-mounted Yanxiang M50 industrial control computer controls the Shenzhen Quanshunhong F800 RFID reader to scan the RFID tags of shared bicycles in the area according to the target pixel coordinates. It reads the EPC codes of 12 illegally parked vehicles in batches, and matches them with the mapping database to find that 8 of them belong to company A and 4 belong to company B. The tag matching time is ≤200ms.

[0039] S3. The Bosch LRR3-76GHZ millimeter-wave radar simultaneously acquired distance and azimuth data of shared bicycles in the area. It measured that the illegal vehicles were 10.5m away from the inspected bicycles and the azimuth was -3.2°. It was confirmed that the vehicles were all statically parked and occupied areas that covered the blind path, with no dynamic interference targets.

[0040] S4. The vehicle-mounted industrial control computer uploads the synchronized data to the cloud data processing and event determination module via the Quectel EC200S module, performing fourth-order fusion processing: (a) Using an image recognition confidence level of 89% as the first-order trigger criterion; (b) Using the 98% RFID tag matching success rate as the basis for the second-level liability attribution, the company to which the violating vehicle belongs is clearly identified; (c) Use radar data to confirm the static parking of vehicles and the area occupied by blind paths as the basis for the third-level verification; (d) Calculate the overall confidence level by weight: 89%×35% + 98%×40% + 95%×25% = 94.1%, and generate a structured urban management event, the event type of which is "shared bicycles illegally parked on blind paths".

[0041] S5. The event classification and response module determines "occupying the blind path" events with a comprehensive confidence level of ≥90% as Level 2 response events based on preset rules.

[0042] S6. The system automatically triggers the Level 2 response and handling process: It sends a list of violating vehicles and their location information to the regional operation and maintenance managers of Enterprise A and Enterprise B respectively through Alibaba Cloud SMS service, generates an inspection report and archives it to Tencent Cloud COS object storage, and synchronizes the violation points to the electronic fence management system built by the shared bicycle company based on Baidu Map API.

[0043] Application effect The accuracy of handling violations reaches 95.3%, with a false judgment rate of ≤3.8%. Through radar data verification, it can accurately distinguish between static illegal parking and dynamic interference targets, avoiding the subjective errors of the traditional manual mode. At the same time, it can achieve linkage response between violation incidents and operating companies within 5 minutes, and shorten the time for rectification of illegal vehicles to 35 minutes. Compared with the traditional mode, the handling efficiency is improved by more than 70%.

[0044] Comparative Example: A test comparing the traditional inspection method with this system's method. 1. Experimental Objective By comparing the core performance of the multi-sensor fusion-based inspection bicycle system of this invention with existing traditional inspection solutions (single-vision inspection vehicle, manual inspection + handheld RFID reader) in municipal facility inspection, the superiority of this system is verified.

[0045] 2. Test subjects and conditions Three municipal roads of the same length, each 5 kilometers long, with the same type of facilities, including 50 manhole covers, 50 streetlights, and 50 traffic signs, were selected as test sections. The following three inspection schemes were used respectively, and each test was repeated 3 times, and the average value was taken.

[0046] Group A (Comparative Example 1): Single-vision four-wheeled inspection vehicle: adopts the solution disclosed in patent CN110378352A, equipped with a high-definition camera and GPU edge computing unit, identifies targets based on YOLOv5 model, without RFID and radar modules, and the turning radius of the four-wheeled inspection vehicle is 3.5m.

[0047] Group B (Comparative Example 2): Manual inspection + handheld RFID reader: Inspectors conduct foot patrols, use handheld UHF RFID readers to read facility tag information, visually identify facility malfunctions, record them, and manually upload them to the management system.

[0048] Group C (Invention): Multi-sensor fusion inspection bicycle system: adopts the system configuration and process in Example 1.

[0049] Test environment: Three municipal roads with a length of 5km and identical facilities (50 manhole covers, 50 streetlights, and 50 traffic signs each) were selected as test sections. Each test was repeated 3 times, and the average value was taken. The test environment covered sunny days, nighttime, and light rain conditions. The road sections included 30% old narrow alleys with a width of 2.0m and 70% main roads. Ten facility failures were preset (2 missing manhole covers, 3 tilted streetlights, and 5 damaged signs). Table 1 As shown in Table 1, the present invention achieves breakthrough improvements in fault identification accuracy, data processing efficiency, environmental adaptability, and automation level, completely resolving the technical shortcomings of the comparative examples. Specifically, the overall fault identification accuracy reaches 99.6%, an increase of 21.4 percentage points compared to comparative example 1 (78.2%) and 10.1 percentage points compared to comparative example 2 (89.5%); the fault false negative rate is only 0.4%, far lower than the 21.8% of comparative example 1 and the 10.5% of comparative example 2, greatly reducing safety hazards caused by false negatives. For harsh environments (night / light rain), the accuracy of this invention reaches 98.9%, which is 36.5 and 26.1 percentage points higher than Comparative Example 1 (62.4%) and Comparative Example 2 (72.8%), respectively, demonstrating stable operation capability in all weather conditions. In a narrow alleyway of 2.0m, the turning radius of this invention is ≤1.0m, allowing for flexible passage, while Comparative Example 1 cannot pass due to its excessively large turning radius (3.5m), and Comparative Example 2 can only meet basic passage requirements. This invention is more suitable for complex and narrow operations.

Claims

1. A bicycle inspection system based on multi-sensor fusion, characterized in that, It includes a mobile data acquisition module, a data processing and event determination module, and an event classification and response module; The mobile acquisition module, which uses multiple sensors to simultaneously acquire multimodal data, includes an image acquisition device, an ultra-high frequency RFID reader, radar, and a vehicle control unit, and is installed on a bicycle. The image acquisition device is fixedly installed at the front of the bicycle body to acquire visual data of the environment in front or to the side. The RFID reader is fixedly installed at the front or middle of the bicycle body and connected to a high-frequency or ultra-high-frequency antenna to read RFID tag information deployed on urban vehicles or public facilities. The radar is fixedly installed at the front of the bicycle, adjacent to the image acquisition device, together covering the detection area in front, and measuring the distance, speed and angle information of the target. The vehicle control unit is communicatively connected to the image acquisition device, the RFID reader and the radar, and synchronously receives and temporarily stores the visual data, RFID tag information and radar data. The event classification and response module is used to assign response levels to the urban management events according to preset rules and automatically trigger the corresponding handling process.

2. The system according to claim 1, characterized in that, The RFID reader is an ultra-high frequency reader, whose radio frequency field covers the main field of view of the image acquisition device, and reads RFID tag information in batches within the field of view during movement; The image acquisition device and the millimeter-wave radar have a lateral deviation of ≤2 cm and a pitch angle deviation of ≤0.5°, ensuring the engineering basis for a field-of-view overlap of ≥90%.

3. The system according to claim 1, characterized in that, The vehicle control unit has a built-in PTP precision timing module to synchronize the image acquisition frame start signal, RFID command trigger signal, and radar sampling trigger signal, with a timestamp deviation of ≤100 μs.

4. The system according to claim 1, characterized in that, The ultra-high frequency RFID reader operates in the frequency band of 902–928 MHz, and the effective reading radius of the radio frequency field is 3.0–3.5 m in the moving state, which meets the requirements for batch reading of tags within the main field of view of the image; The millimeter-wave radar operates in the 76–81 GHz frequency band, with a range accuracy of ±0.1 m, an angular resolution of ±1°, and a point cloud output frame rate of ≥10 Hz, supporting continuous tracking of dynamic targets.

5. The system according to claim 1, characterized in that, The vehicle control unit and the data processing and event determination module interact with each other via a wireless communication network; the data processing and event determination module and the event classification and response module are deployed on a remote server or cloud platform.

6. The system according to claim 1, characterized in that, The data processing and event determination module is deployed on a remote server or cloud platform to receive synchronized visual data, RFID tag information, and radar data, and performs the following fourth-order fusion processing: First stage: Using image recognition results as a trigger, extract the pixel coordinates, size, color histogram, and YOLOv8s confidence of the potential event target; The second step involves calling the RFID-ID mapping database and returning management attributes such as EPC code, affiliated unit, and maintenance contact person based on pixel coordinate matching, thereby realizing semantic association between form and identity. The third stage involves using target distance, velocity, and azimuth data measured by millimeter-wave radar to perform status verification on the image-RFID matching results. Fourth stage: Calculate the overall confidence level using a weighted method: image attributes contribute 35%, RFID matching degree contributes 40%, and radar status consistency contributes 25%. The overall confidence level serves as the core field for structured urban management events. The event classification and response module is deployed on a remote server or cloud platform to allocate response levels one to four based on the overall confidence level and event type, and automatically trigger the corresponding handling process.

7. The system according to claim 1, characterized in that, When the data processing and event determination module performs fusion analysis, it identifies potential event targets and their image attributes through the visual data; it obtains the identification information or management attributes of the potential event targets through the RFID information; and it matches and associates the image attributes with the identification information or management attributes to generate structured urban management events.

8. A method for inspecting bicycles based on multi-sensor fusion, characterized in that, Includes the following steps:   S1. Acquire video streams using an image acquisition device, identify potential event targets using the YOLOv8s model, and extract pixel coordinates, image attributes, and YOLOv8s confidence scores; S2. Control the UHF RFID reader to scan in the coordinate area obtained in S1, and attempt to obtain RFID tag information associated with the target; S3. Simultaneously acquire range, velocity, and azimuth data of the target from the millimeter-wave radar; S4. Perform fourth-order fusion: (a) The first-order trigger is based on the image recognition result; (b) Secondary liability is determined based on RFID matching results; (c) Use radar status data as the third-order verification; (d) Calculate the overall confidence level using a weighting of 35%+40%+25% to generate structured urban management events; S5. Based on the confidence level and event type obtained in S4, the event classification and response module assigns the response level; S6. Automatically execute corresponding processing procedures, including SMS push, voice reminder, work order dispatch, GIS heat map update or archiving for future reference.

Images