Highway motorcycle identification system

Abstract

The invention discloses a highway motorcycle recognition system, and belongs to the technical field of traffic management. An intermediate intelligent identification module; and a back-end support module. A front-end high-definition observation module is used for acquiring moving target data of a designated area of a current road section, so that recognition work can be completed more efficiently in a subsequent processing link; the middle intelligent recognition module is used for loading an intelligent recognition algorithm in advance according to the detection information, obtaining moving target data of a designated area and recognizing the moving target data to obtain motorcycle information, recognition preparation can be conducted in advance, and the problems that in a high-speed environment, an image obtaining window is small, and the image obtaining effect is poor are solved; and the rear-end storage and optimization module is used for providing remote data processing and algorithm optimization capability and can overcome the problem that the information transmission capability and the data processing capability are limited in a special highway environment.

Description

Technical Field

[0001] The present invention relates to the technical field of traffic management, in particular to a highway motorcycle recognition system. Background Art

[0002] With the rapid development of intelligent transportation systems, highway traffic monitoring and management are of great significance for improving road safety and reducing traffic accidents. However, traditional monitoring methods face many challenges in motorcycle recognition, such as motorcycles are small in size, fast in speed, and susceptible to light and environmental interference. The system relies on cloud servers to process video data, which has the following defects: 1. High latency: The average round-trip latency to the cloud is 300-500ms, which is difficult to meet the real-time recognition requirements of high-speed motorcycles (IEEE ITS 2022 data); 2. Low accuracy: Conventional target detection algorithms (such as HOG+SVM) have a recognition rate of less than 75% for small moving targets (CVPR 2021 study); 3. Poor environmental adaptability: Light changes (such as tunnel entrances and exits) and motion blur lead to a license plate miss rate of over 30%. Summary of the Invention

[0003] The present invention proposes a highway motorcycle recognition system to solve the problems of poor recognition efficiency and low precision in the prior art.

[0004] In order to achieve the above object, the technical solution adopted by the present invention is:

[0005] The highway motorcycle recognition system includes: a front-end high-definition detection module, which is used to obtain mobile target data in a specified area of ​​the current road section and process it to obtain predicted arrival information; an intermediate intelligent recognition module, which is used to load the recognition algorithm in advance based on the predicted arrival information, obtain mobile target data in a specified area of ​​the current road section and identify motorcycle information; and a back-end support module, which is used to provide remote data processing capabilities.

[0006] Furthermore, the front-end high-definition observation module includes a high-definition camera deployed on the road section for collecting real-time video streams; the high-definition camera uses a Sony IMX585 sensor, supports 4K resolution and 120fps high frame rate shooting, a dynamic range of ≥120dB, and an integrated active infrared fill light.

[0007] Furthermore, the front-end high-definition observation module and / or the intermediate intelligent recognition module include an edge computing unit that integrates a high-performance processor and a customized AI algorithm, wherein the customized AI algorithm includes a YOLO algorithm and an SSD algorithm; the front-end high-definition observation module and / or the intermediate intelligent recognition module are communicatively connected with the back-end support module and transmit structured data through the lightweight MQTT protocol.

[0008] Furthermore, the recognition algorithm includes: using Radon transform to detect the inclination angle of the license plate, and correcting the tilted license plate through affine transformation through bilinear interpolation; based on the MSRCR algorithm, automatically activating low-light enhancement when the illumination is <10lux; using a deep learning network to extract image features from motorcycle images, and then using a classifier to identify the vehicle model, license plate number and driver characteristics, wherein the image features include shape, color and texture.

[0009] Furthermore, the back-end support module includes a cloud server; the cloud server regularly aggregates data from each edge node based on a federated learning algorithm, updates the global model, and optimizes the recognition algorithm and / or the AI ​​algorithm based on big data analysis.

[0010] Furthermore, the optimization of the recognition algorithm and / or the AI ​​algorithm based on big data analysis includes: training a deep learning model based on historical data, and regularly updating the model to adapt to newly emerging motorcycle types and environmental changes; the updating model includes using public data sets and self-collected highway motorcycle data to train the deep learning model.

[0011] Furthermore, the public dataset and / or the self-collected highway motorcycle data are optimized, including: randomly adding 0-15px displacement blur to simulate motion blur; and simulating lighting disturbance by ±30% brightness change.

[0012] Furthermore, the intermediate intelligent recognition module is used for: video stream decoding, identifying moving targets based on inter-frame difference method, ROI area interception, YOLOv8s-Mobile reasoning, license plate correction and CRNN recognition, and outputting results.

[0013] Furthermore, the back-end support module is used to output warning information and algorithm update and optimization based on the recognition results.

[0014] Furthermore, the front-end high-definition observation unit determines the type and movement data of the moving target in the specified area of ​​the current road section based on appearance feature recognition; calculates the predicted time and predicted motion posture of the motorcycle arriving at the observation position corresponding to the intermediate intelligent recognition module unit according to the current position, road data and the position of the intermediate intelligent recognition module; outputs the appearance recognition feature, the predicted time and the predicted motion posture to obtain the predicted arrival information.

[0015] Due to the adoption of the above technical solution, the present invention has the following beneficial effects:

[0016] 1. The present invention uses a front-end high-definition observation module to obtain mobile target data in a specified area of ​​the current road section, and processes it to obtain predicted arrival information, which can provide early warning and facilitate subsequent processing links to complete the identification work more efficiently; the intermediate intelligent recognition module is used to load the recognition algorithm in advance according to the predicted arrival information, obtain mobile target data in a specified area of ​​the current road section and identify motorcycle information, and make recognition preparations in advance to adapt to the problems of small image acquisition window and poor image acquisition effect in high-speed environment; the back-end support module is used to provide remote data processing capabilities, which can overcome the problems of limited information transmission capacity and data processing capacity in the special environment of highways.

[0017] 2. The present invention uses a front-end high-definition observation unit to determine the type and movement data of the mobile target in the designated area of ​​the current road section based on appearance feature recognition, and can obtain basic information of the identified target in advance; according to the current position, road data and the position of the intermediate intelligent recognition module, the predicted time and predicted movement posture of the motorcycle arriving at the observation position corresponding to the intermediate intelligent recognition module unit are calculated, and the predicted information of the motorcycle can be obtained; the appearance recognition features, the predicted time and the predicted movement posture are output to obtain the predicted arrival information, and the data can be integrated to facilitate subsequent processing. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] Figure 1 This is a schematic diagram of the highway motorcycle recognition system proposed by the present invention. DETAILED DESCRIPTION

[0019] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0020] The highway motorcycle recognition system includes: a front-end high-definition observation module 1, which is used to obtain mobile target data in a specified area of ​​the current road section and process it to obtain predicted arrival information; an intermediate intelligent recognition module 2, which is used to load the recognition algorithm in advance based on the predicted arrival information, obtain mobile target data in a specified area of ​​the current road section and identify motorcycle information; and a back-end support module 3, which is used to provide remote data processing capabilities.

[0021] Most highways are located in non-urban areas. Due to their distance from cities, maintenance is often difficult. In practice, specific locations are chosen for electronic equipment such as cameras and speedometers. These locations are generally chosen to ensure easy access for maintenance personnel, access to a power source, and protection from external damage (such as exposure to natural disasters, flooding, or physical damage caused by accidents). However, existing technologies are insufficient for identifying high-speed moving objects on highways. Key bottlenecks include the increased difficulty in acquiring high-definition images if the target moves too quickly, and the increased processing power required for image processing. Installing high-end equipment for image acquisition and data processing to overcome these challenges would be prohibitively expensive, making it unsuitable for use in highway environments.

[0022] This solution is equipped with a front-end high-definition observation module to obtain mobile target data in the designated area of ​​the current road section and process it to obtain predicted arrival information. Its principle is to roughly identify the mobile targets in the designated area of ​​the current road section and then hand it over to the mid-end for processing. At this time, since precise identification is not required, the corresponding image acquisition equipment, processing algorithms and processors do not need to be too advanced, thus saving costs. For the special environment of highways, there may not be many places where it is more suitable to set up cameras and other equipment. However, in practice, there are too many places that need to be inspected, such as the need to prevent intruders (humans or animals) outside the road, or to record possible accidents. The use of a low-cost front-end high-definition observation module can meet multiple needs at the same time.

[0023] The intermediate intelligent recognition module set up in this solution can load the recognition algorithm in advance according to the predicted arrival information, obtain the moving target data in the specified area of ​​the current road section, and identify the motorcycle information. At this time, since the preparation time is more sufficient, on the one hand, the hardware equipment can be controlled to prepare in advance, such as turning on a specific working mode, or moving the camera to limit the shooting area, etc.; on the other hand, appropriate processing algorithms can be selected to improve the efficiency and accuracy of image processing.

[0024] The back-end support module set up in this solution can make up for the inconvenient communication in the highway area, the insufficient data processing capabilities of the intermediate intelligent recognition module / front-end high-definition observation module, and provide other additional functions.

[0025] The front-end high-definition observation module includes a high-definition camera deployed on the road section for collecting real-time video streams; the high-definition camera uses a Sony IMX585 sensor, supports 4K resolution and 120fps high frame rate shooting, a dynamic range of ≥120dB, and an integrated active infrared fill light.

[0026] High-definition cameras deployed along key highway sections are responsible for capturing real-time video streams. These cameras feature a wide dynamic range and low-light design. Powered by the Sony IMX585 sensor, they support 4K resolution (3840×2160) and 120fps high frame rate shooting. With a dynamic range of ≥120dB, they ensure license plate recognition in backlit environments and capture clear motorcycle images in all lighting conditions. Integrated active infrared fill light (850nm wavelength) increases the effective nighttime recognition range to 50 meters.

[0027] Due to design and cost considerations, large areas of highways are directly exposed to the sun. Furthermore, shadows may appear due to surrounding trees, mountains, and clouds. At night, vehicle headlights and camera fill light create a contrast between light and dark. Wide dynamic range (WDR) is a technology used to enhance the image quality of cameras by facilitating strong contrast. When bright areas illuminated by strong light sources (sunlight, lamps, or reflections) coexist with relatively low-brightness areas such as shadows and backlight, the camera's output image may appear white due to overexposure, while dark areas may appear black due to underexposure, severely impacting image quality. A camera's ability to capture both the brightest and darkest areas in a scene is limited; this limitation is commonly referred to as "dynamic range." Using a WDR camera can improve the efficiency and accuracy of daytime image acquisition. Low-light cameras can meet the demands of nighttime image acquisition.

[0028] The front-end high-definition observation module and / or the intermediate intelligent recognition module include an edge computing unit that integrates a high-performance processor and a customized AI algorithm, wherein the customized AI algorithm includes a YOLO algorithm and an SSD algorithm; the front-end high-definition observation module and / or the intermediate intelligent recognition module are communicatively connected with the back-end support module and transmit structured data through the lightweight MQTT protocol.

[0029] An edge computing unit integrating high-performance processors and customized AI algorithms: Deploys the NVIDIA Jetson AGX Orin edge computing unit, which has a computing power of 200TOPS and supports INT8 quantitative reasoning; deployed near the front-end monitoring equipment (front-end high-definition observation module), it is responsible for receiving video streams and performing pre-processing, feature extraction, and motorcycle recognition.

[0030] Edge computing devices utilize deep learning models such as the YOLO (You Only Look Once) series and SSD (Single Shot MultiBox Detector). The improved YOLOv8s-Mobile model introduces a Ghost module to replace traditional convolutions, reducing the number of parameters by 40%. A motion blur compensation layer (MBLC) is embedded to enhance feature extraction of high-speed targets through frequency-domain filtering. The license plate location module integrates color space clustering (HSV threshold: H∈[200,260], S>0.3, V>0.5) with geometric constraints (aspect ratio of 3:1±0.2) to rapidly detect and track motorcycles in video frames. Edge computing devices ensure low latency and secure data transmission.

[0031] The recognition algorithm includes: using Radon transform to detect the license plate tilt angle, and correcting the tilted license plate through affine transformation through bilinear interpolation; automatically activating low-light enhancement when the illumination is less than 10 lux based on the MSRCR algorithm; using a deep learning network to extract image features from motorcycle images, and then using a classifier to identify the vehicle model, license plate number and driver characteristics, wherein the image features include shape, color and texture.

[0032] Radon transform is used to detect the tilt angle of the license plate, and affine transformation is implemented through bilinear interpolation to correct the tilted license plate. The MSRCR algorithm (Multi-Scale Retinex with Color Restoration) based on Retinex theory automatically activates low-light enhancement when the illumination is less than 10 lux. A deep learning network is used to extract features from motorcycle images, including key features such as shape, color, and texture. Subsequently, a classifier (deep neural network) is used to identify the vehicle model, license plate number, and even driver characteristics (such as whether a helmet is worn).

[0033] After obtaining the vehicle model, license plate number and even driver characteristics, it is also possible to combine historical data such as the speed indicators of different vehicle models, driving habits of different license plates in corresponding areas, and driving habits of different drivers (gender, age, etc.) to facilitate subsequent driving predictions.

[0034] The back-end support module includes a cloud server; the cloud server regularly aggregates data from each edge node based on a federated learning algorithm, updates the global model, and optimizes the recognition algorithm and / or the AI ​​algorithm based on big data analysis.

[0035] The edge computing device uploads recognition results (including motorcycle images, identification information, and timestamps) to a cloud server in real time for remote monitoring and management. Structured data (only the license plate number, timestamp, and confidence level) is transmitted using the lightweight MQTT protocol, reducing bandwidth usage by 80%. Federated learning regularly aggregates data from edge nodes to update the global model. Furthermore, the cloud platform uses big data analysis to further optimize the AI ​​model and improve recognition accuracy.

[0036] The optimization of the recognition algorithm and / or the AI ​​algorithm based on big data analysis includes: training a deep learning model based on historical data, and regularly updating the model to adapt to newly emerging motorcycle types and environmental changes; the updating model includes using public data sets and self-collected highway motorcycle data to train the deep learning model.

[0037] A deep learning model is trained based on historical data and regularly updated to adapt to emerging motorcycle types and environmental changes. This model integrates a public dataset (UA-DETRAC) with in-house highway motorcycle data (100,000 images, covering rain, fog, and nighttime scenes). To increase the number of training samples and improve model recognition accuracy, the dataset can be processed, such as by simulating motion blur (randomly adding 0-15px displacement blur) and lighting perturbations (±30% brightness variation).

[0038] Optimization processing is performed on the public dataset and / or the self-collected highway motorcycle data, including: randomly adding 0-15px displacement blur to simulate motion blur; and simulating lighting disturbance by ±30% brightness change.

[0039] The intermediate intelligent recognition module is used for: video stream decoding, identifying moving targets based on the inter-frame difference method, ROI area interception, YOLOv8s-Mobile reasoning, license plate correction and CRNN recognition, and outputting results.

[0040] Real-time processing flow: video stream decoding (H.265 hard decoding) → moving target detection (inter-frame difference method) → ROI area interception → YOLOv8s-Mobile inference → license plate correction and CRNN recognition → result output (the entire process takes ≤35ms).

[0041] The back-end support module is used to output warning information based on the recognition results.

[0042] Recognition results are uploaded to the cloud platform in real time, triggering an early warning mechanism to send alerts to traffic management and drivers. Based on the collected data, the cloud platform continuously optimizes the AI ​​model to improve recognition accuracy and efficiency. This real-time early warning mechanism helps promptly detect and address traffic violations, enhancing road safety. The system is easily scalable and adaptable to different highway scenarios, providing strong technical support for intelligent transportation systems.

[0043] The front-end high-definition observation unit determines the type and movement data of the mobile target in the specified area of ​​the current road section based on appearance feature recognition; calculates the predicted time and predicted motion posture of the motorcycle arriving at the observation position corresponding to the intermediate intelligent recognition module unit according to the current position, road data and the position of the intermediate intelligent recognition module; outputs the appearance recognition feature, the predicted time and the predicted motion posture to obtain the predicted arrival information.

[0044] Vehicle appearance generally conforms to certain patterns. This is especially true in countries that don't support vehicle modifications. With sufficient samples and mature image recognition models, it's possible to identify the type of moving target. Furthermore, since the highway serves as a fixed background, simple image position changes and highway dimensional parameters can be used to calculate speed, position, and other data. Based on this movement and position data, it's possible to roughly estimate the time it will take for the current target to reach a subsequent designated location.

[0045] Furthermore, highways don't always have a single moving target. In practice, it's quite common for multiple vehicles to follow one another. In this case, the interplay of vehicles can lead to target overlap. Even without this overlap, if the target isn't lost, rapid left-right movement can still affect image acquisition and analysis. Predicting motion posture involves predicting the motorcycle's tendency to move. For example, if the front-end high-definition observation unit previously observed only one motorcycle, there's no need for it to specifically change lanes on the highway. In any case, if multiple vehicles are present, the likelihood of the motorcycle overtaking and changing lanes is high. Based on information such as the number, speed, and type of vehicles, combined with a preset predictive analysis model, a predicted motion posture can be derived. Controlling the appropriate camera (if multiple cameras are used) or using a suitable data processing algorithm can improve recognition efficiency and accuracy.

[0046] The above description is a detailed description of the preferred embodiments of the present invention, but the embodiments are not intended to limit the scope of the patent application of the present invention. Any equivalent changes or modifications completed under the technical spirit suggested by the present invention should fall within the patent scope covered by the present invention.

Claims

1. Highway motorcycle identification system, characterized by: include: The front-end high-definition observation module is used to obtain the moving target data in the specified area of ​​the current road section and process it to obtain the predicted arrival information; The intermediate intelligent recognition module is used to load the recognition algorithm in advance according to the predicted arrival information, obtain the moving target data in the specified area of ​​the current road section and identify the motorcycle information; Back-end support module, used to provide remote data processing and algorithm optimization capabilities.

2. The highway motorcycle identification system according to claim 1, characterized in that: The front-end high-definition observation module includes a high-definition camera deployed on the road section to collect real-time video streams; The high-definition camera uses a Sony IMX585 sensor, supports 4K resolution and 120fps high frame rate shooting, has a dynamic range of ≥120dB, and integrates an active infrared fill light.

3. The highway motorcycle identification system according to claim 2, characterized in that: The front-end high-definition observation module and / or the intermediate intelligent recognition module include an edge computing unit that integrates a high-performance processor and a customized AI algorithm, wherein the customized AI algorithm includes the YOLO algorithm and the SSD algorithm; The front-end high-definition observation module and / or the intermediate intelligent recognition module are communicatively connected with the back-end support module and transmit structured data via the lightweight MQTT protocol.

4. The highway motorcycle identification system according to claim 3, characterized in that: The recognition algorithm includes: Radon transform is used to detect the tilt angle of the license plate, and affine transform is implemented through bilinear interpolation to correct the tilted license plate; Based on the MSRCR algorithm, low-light enhancement is automatically activated when the illumination is <10lux; A deep learning network is used to extract image features from motorcycle images, and then a classifier is used to identify the vehicle model, license plate number and driver characteristics, where the image features include shape, color and texture.

5. The highway motorcycle identification system according to claim 4, characterized in that: The backend support module includes a cloud server; The cloud server regularly aggregates data from each edge node based on a federated learning algorithm, updates the global model, and optimizes the recognition algorithm and / or the AI ​​algorithm based on big data analysis.

6. The highway motorcycle identification system according to claim 5, characterized in that: The optimizing the recognition algorithm and / or the AI ​​algorithm based on big data analysis includes: Based on historical data, deep learning models are trained and regularly updated to adapt to new motorcycle types and environmental changes; The updated model includes using public data sets and self-collected highway motorcycle data to train a deep learning model.

7. The highway motorcycle identification system according to claim 6, characterized in that: Optimizing the public dataset and / or the self-collected highway motorcycle data includes: Simulate motion blur by randomly adding 0-15px displacement blur; The brightness was varied by ±30% to simulate light disturbances.

8. The highway motorcycle identification system according to claim 7, characterized in that: The intermediate intelligent identification module is used to: Video stream decoding, moving target recognition based on inter-frame difference method, ROI region interception, YOLOv8s-Mobile inference, license plate correction and CRNN recognition, and output results.

9. The highway motorcycle identification system according to claim 8, characterized in that: The back-end support module is used to output warning information based on the recognition results.

10. The highway motorcycle identification system according to claim 9, characterized in that: The front-end high-definition observation unit determines the type and movement data of the mobile target in the designated area of ​​the current road section based on appearance feature recognition; Calculate the predicted time and predicted motion posture of the motorcycle arriving at the observation position corresponding to the intermediate intelligent recognition module unit based on the current position, road data and the position of the intermediate intelligent recognition module unit; The appearance recognition feature, the predicted time, and the predicted motion posture are output to obtain the predicted arrival information.

Images