Intelligent monitoring method and system for overload and manned of motorcycle

Abstract

The invention belongs to the technical field of computer vision, motorcycle detection and motorcycle head detection, and discloses a motorcycle overloading and manned intelligent monitoring system which comprises a motorcycle detection model module, a head detection model module, a reasoning module, a processing module, a distance judgment module and a pushing module. According to the method, a specially customized motorcycle detection model is adopted, and the model is trained through a large number of electric vehicle and bicycle negative samples, so that the probability of misclassification of electric vehicles and bicycles as motorcycles is remarkably reduced. Besides, a specially customized motorcycle head detection model (the model focuses on detecting the head of the passenger sitting on the motorcycle) is used for replacing a traditional pedestrian detection model, so that the detection rate of the motorcycle passenger is improved, and misjudgment on nearby pedestrians is reduced. In the post-processing stage, a finer judgment mechanism is introduced, for example, the detection accuracy is further ensured by analyzing the relative position relation between the head and the motorcycle.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of computer vision, motorcycle detection, and motorcycle head detection, and particularly relates to a motorcycle overloading and passenger intelligent monitoring method and system. BACKGROUND

[0002] The existing technical solutions usually use snapshot devices equipped with motorcycle and pedestrian detection functions. When these devices detect a motorcycle, they automatically crop out the image of the motorcycle area and further perform pedestrian detection on the image. If the number of detected pedestrians exceeds the pre-set threshold, the system will generate a warning information. The purpose of this method is to improve the monitoring efficiency and accuracy of motorcycle overloading and passenger phenomenon through automated image processing technology.

[0003] Current pedestrian detection algorithms often face accuracy challenges in complex scenarios, such as dense streets or uneven lighting conditions. In these situations, the algorithm may misjudge and incorrectly classify nearby pedestrians as motorcycle passengers, or fail to accurately identify the presence of motorcycle passengers due to their partial exposure, resulting in missed detection. In addition, due to the similarity in appearance between motorcycles and electric bicycles, the algorithm often incorrectly identifies electric bicycles as motorcycles. These situations significantly reduce the accuracy and reliability of detection, affecting the overall performance of the system.

[0004] For example, Chinese invention patent application CN112395976B discloses a motorcycle passenger recognition method, device, equipment and storage medium, which gives a complete motorcycle passenger recognition link: first, the target is locked as a motorcycle through the exhaust pipe or motorcycle license plate, then the counting mode is selected according to the camera collection direction - when collecting forward / backward, the head and shoulder or leg contour edge is extracted, and the number of people is obtained after removing the "edge spacing + key point spacing"; when collecting laterally, the head detection or leg counting is directly run. If the final number of people exceeds the threshold, an overload signal is sent. This method considers the shielding difference under different viewing angles, and for the first time introduces "contour de-duplication" and "viewing angle self-adaptation" into the overstaff recognition, but it still exposes defects that are difficult to avoid. First, the vehicle type distinguishing feature is too thin. This method only uses a single signal of "exhaust pipe" or "motorcycle license plate" to filter electric vehicles, and for the working conditions of new electric motorcycles without exhaust pipes, license plate damage, folding or intentional shielding, the classifier will inevitably fail, resulting in that all subsequent counting results are based on the wrong vehicle type. Second, the geometric consistency check is missing. In the lateral collection mode, the method directly counts all the heads in the motorcycle box without setting a "head-vehicle body" relative position threshold, so the heads of roadside pedestrians and parallel vehicle passengers are easily included in the box, resulting in "false overstaff". Finally, the de-duplication strategy is empirical and fragile. Using fixed thresholds K1 and K2 to judge "ghost contours" does not consider the scale changes caused by the camera's downward angle, nor does it provide a compensation mechanism for contour breaks in low-illumination and motion-blur conditions, resulting in the head and shoulder edges of the same passenger being split into multiple edges, or the edges of different passengers being combined into one edge, and the final number of people often has an error of ±1.

[0005] Therefore, although this method takes the first step in "viewing angle self-adaptive counting", it still has false positives and false negatives in real road scenes such as electric vehicle mixed driving, dense parallel driving, and downward angle shooting, and a systematic solution is needed to solve the problems of vehicle type confusion, spatial verification, and shielding de-duplication. SUMMARY

[0006] In view of the problems existing in the prior art, the present application provides a motorcycle overstaff passenger intelligent monitoring system.

[0007] The present application is implemented as follows: a motorcycle overstaff passenger intelligent monitoring system comprises:

[0008] a motorcycle detection model module, a head detection model module, an inference module, a processing module, a distance determination module, and a pushing module;

[0009] The motorcycle detection model module is connected with the head detection model module and the distance determination module, and is used to customize a motorcycle detection model.

[0010] The human head detection model module is connected with the motorcycle detection model module, the reasoning module, the processing module and the distance judgment module, and is used for customizing the motorcycle detection model.

[0011] The reasoning module is connected with the human head detection model module and the processing module, and is used for model reasoning.

[0012] The processing module is connected with the human head detection model module, the reasoning module and the distance judgment module, and is used for post-processing.

[0013] The distance judgment module is connected with the motorcycle detection model module, the human head detection model module and the processing module, and is used for distance judgment of the driver head frame and the vehicle body frame.

[0014] The pushing module is connected with the processing module, and is used for pre-warning data pushing Kafka.

[0015] Further, the customized motorcycle detection model comprises:

[0016] A large number of actual scene images are collected, including motorcycles, electric vehicles, bicycles and tricycles; in these images, the motorcycle is used as a positive sample, and other vehicles are used as negative samples;

[0017] The model is trained by using a YOLOv5 framework, and through continuous fine-tuning, a customized motorcycle detection model is finally obtained.

[0018] Further, the customized motorcycle human head detection model comprises:

[0019] 1) The motorcycle detection model is used to crop the motorcycle part from the original image, and artificial verification is performed;

[0020] 2) The head of the occupant on the motorcycle is labeled, and the driver and the passenger are distinguished;

[0021] 3) The YOLOv5 framework is used for model training, and through fine-tuning, a customized model focusing on detecting the head of the motorcycle occupant is obtained.

[0022] Further, the model reasoning comprises:

[0023] 1) After accessing the video stream, first, motorcycle detection is performed to obtain the motorcycle image and the vehicle body frame coordinates;

[0024] 2) The detected motorcycle image is cropped by a certain proportion to avoid the misinclusion of pedestrians beside the motorcycle into the detection range, and then motorcycle human head detection is performed to obtain the human head frame coordinates.

[0025] Further, the post-processing comprises:

[0026] 1) Remove incomplete vehicles

[0027] Excluding those only partially appearing in the image, ensuring that only complete vehicles are used for analysis;

[0028] 2) Remove overlapping vehicles

[0029] If a vehicle identifies multiple drivers, it means that two motorcycles are squeezed together, and the data detected in this case is incorrect, which is directly filtered out;

[0030] 3) Distance judgment between passenger head frame and driver head frame;

[0031] If the x-coordinate of the center point of the passenger head frame deviates significantly from the x-coordinate of the center point of the driver head frame (i.e., the x-coordinate of the center point of the passenger head frame minus the x-coordinate of the center point of the driver head frame is greater than 1.3 times the width of the driver head frame) or the y-coordinate of the center point of the passenger head frame is significantly higher than the y-coordinate of the center point of the driver head frame (i.e., the y-coordinate of the center point of the passenger head frame minus the y-coordinate of the center point of the driver head frame is greater than 1 times the height of the driver head frame), the application considers that this position relationship is unreasonable and is excluded from further analysis to avoid misjudgment;

[0032] 4) Distance judgment between driver head frame and vehicle body frame

[0033] Pay attention to the proportional relationship between the height of the center point of the head frame and the height of the center point of the vehicle body frame.

[0034] Further, the pre-warning data is pushed to Kafka:

[0035] After data cleaning and quality control, the system will count the number of motorcycle passengers' heads; according to the preset threshold, if the number of counted heads exceeds the threshold, the system will trigger a warning message and push it to the Kafka message queue; then, the business application system will consume these pre-warning data and display them on the interface, so that relevant personnel can respond and handle the overloading situation in a timely manner.

[0036] Another object of the present application is to provide a motorcycle overloading intelligent monitoring method comprising:

[0037] Step 1, customizing a motorcycle detection model through a motorcycle detection model module;

[0038] Step 2, customizing a motorcycle detection model through a head detection model module;

[0039] Step 3, model inference through an inference module;

[0040] Step 4, post-processing through a processing module;

[0041] Step 5, the distance between the driver's head frame and the vehicle body frame is judged by the distance judgment module;

[0042] Step 6, the warning data is pushed to Kafka by the pushing module.

[0043] Another object of the present application is to provide a computer device comprising a memory and a processor, the memory storing a computer program, the computer program being executed by the processor to make the processor execute the steps of the motorcycle overloading intelligent monitoring method.

[0044] Another object of the present application is to provide a computer readable storage medium storing a computer program, the computer program being executed by a processor to make the processor execute the steps of the motorcycle overloading intelligent monitoring method.

[0045] Another object of the present application is to provide an information data processing terminal for implementing the motorcycle overloading intelligent monitoring system.

[0046] In combination with the above technical solutions and the technical problems solved, the technical solutions to be protected by the present application have the following advantages and positive effects:

[0047] First, the core technical problems and difficulties solved by the present application in view of the defects of the prior art

[0048] 1. Vehicle type confusion: electric motorcycles have no exhaust pipe and there are many license plate pollution / folding / suiting phenomena, so the single feature classifier will inevitably fail. The present application first constructs a "motorcycle exclusive sample library" to train YOLOv5 with 87,000 negative samples of electric vehicles, bicycles and tricycles to reduce the vehicle type misclassification rate to 2.1% (internal test set, IoU >= 0.5 condition).

[0049] 2. Pedestrian interference: the roadside pedestrians are counted as passengers by the existing system, resulting in false positives. The present application restricts the "head" detection range within "vehicle body frame + 0.15 times frame height", and introduces the "driver-passenger" relative position priori (X-direction offset <= 1.3 times driver head width, Y-direction not higher than 1 times head height), which reduces the pedestrian false positive rate from 15.4% to 0.9%.

[0050] 3. Occlusion and missed detection: some passengers only show the top of the helmet, and the general pedestrian model recall is insufficient. The present application customizes a "motorcycle head detection model", which uses 6 layers of pyramid + small target anchor under 640x640 input, and the recall rate of 40x40 pixel helmet is improved by 22.6%, and the overall head mAP@0.5 reaches 91.7%.

[0051] 4. Lack of Spatial Verification: Existing technologies lack geometric consistency verification between the "person's head" and the vehicle body. This invention proposes a "relative height ratio" criterion: the y-coordinate of the person's head center must be located within 2 to 5 times the y-coordinate of the vehicle body center; otherwise, it is judged as an abnormal target and removed. This can further reduce false alarms caused by "pedestrians standing next to the vehicle" by 38%.

[0052] II. Key Data and Validation Results of the R&D Process

[0053] Data scale: 420,000 kilometers of real road video were collected, covering 6 provinces and 11 cities, and 316,000 motorcycles and 682,000 people were labeled.

[0054] Training strategy: Mosaic+MixUp+HSV is used to enhance SIoU loss, and the iteration is performed for 300 epochs; it can converge after 38 hours of training on Tesla V100.

[0055] Evaluation results:

[0056] – Vehicle model accuracy rate: 97.9%;

[0057] - Recall rate 91.7%, precision rate 94.5%, overall F193.1%;

[0058] – The overall overcrowding detection accuracy was 96.4%, with a false positive rate of 0.8% and a false negative rate of 1.5%.

[0059] Edge deployment: After TensorRT INT8 quantization, the inference time for a single Jetson Xavier is 22ms (1920×1080), which meets the requirements of 25fps real-time streaming processing, and the power consumption is only 18W.

[0060] III. Creative Technical Effects

[0061] 1. For the first time, "negative sample enhancement + spatial geometry verification" was introduced into the motorcycle overload scenario, solving two major industry pain points: vehicle type confusion and pedestrian interference;

[0062] 2. A three-level bounding box association model of "driver's head - passenger's head - vehicle body" is proposed to achieve joint optimization of pixel-level semantic and geometric consistency, so that the system can still maintain a false alarm rate of <1% in extreme scenarios such as dense traffic, low light at night, and partial occlusion.

[0063] 3. The end-to-end link is entirely based on edge computing, eliminating the need to upload original videos, which complies with the requirements of the Data Security Law and the Personal Information Protection Law, and is ready for large-scale deployment.

[0064] 4. It has the capability of zero-modification and reuse of existing equipment, and can be connected to the existing public security video platform. The single-point deployment cost is reduced by 73% compared with the traditional "checkpoint + ground sensor" solution, and the maintenance workload is reduced by 60%, providing grassroots traffic management departments with an economical and feasible non-on-site law enforcement method.

[0065] In summary, this invention not only overcomes the common defects of high false alarms, high false negatives, and high costs in existing patented technologies, but also achieves breakthroughs in detection accuracy, real-time performance, and deployment cost through a dual-drive approach of "customized depth model + geometric posterior rules," demonstrating significant inventiveness and engineering promotion value.

[0066] Second, the expected benefits and commercial value of the technical solution of this invention after transformation are as follows:

[0067] 1. Direct Economic Benefits: Based on a single edge device (Jetson Xavier) cost of 4,500 yuan and software licensing cost of 12,000 yuan, deploying two sets per kilometer of urban road is sufficient to cover a two-way four-lane road. According to the projection of 180,000 kilometers of national and provincial highways nationwide by 2025 and a 70% penetration rate, the hardware + software market size is 2.56 billion yuan; adding 5 years of operation and maintenance and upgrades, the total market space is 4.2 billion yuan. The company maintains a gross profit margin of over 50%, with an estimated cumulative net profit of 630 million yuan over 3 years and an investment payback period of 1.1 years.

[0068] 2. Cost reduction and efficiency improvement value: Replacing the traditional "checkpoint + ground sensor" solution (single point cost of 120,000-150,000 yuan), this invention reuses old cameras, with a single point renovation cost of ≤20,000 yuan, which can save the government more than 70% of construction funds; at the same time, it reduces the man-hours of manual inspection by 60%. Based on the annual inspection budget of 180 million yuan for 100 counties, it saves 108 million yuan of fiscal expenditure every year.

[0069] 3. Data operation extension: Overload capture data can be connected with insurance and credit platforms to form "driving behavior scores", which is expected to generate an additional SaaS revenue of 240 million yuan per year; further derivative scenarios such as used car residual value assessment and accurate car insurance pricing are expected to be developed. The data operation revenue is expected to increase to 35% in five years.

[0070] The technical solution of this invention fills a technological gap in the industry both domestically and internationally:

[0071] This invention proposes for the first time a pixel-level detection framework with a three-level visual association of "driver's head - passenger's head - vehicle body," upgrading motorcycle overload detection from "counting heads on the entire vehicle" to a fine-grained paradigm of "distinguishing driver and passenger identities + spatial geometry verification." Using a single RGB video stream, it simultaneously completes driver and passenger head detection, identity labeling, relative position constraints, and overload determination. Without requiring additional modules such as exhaust pipe recognition, license plate recognition, or infrared illumination, it reduces the false alarm rate to 0.8%, filling the technological gap in "visually distinguishing driver and passenger identities and quantifying spatial consistency." This provides the first penalizable pure vision solution for non-site enforcement of motorcycle overload regulations. Attached Figure Description

[0072] Figure 1 This is a system structure block diagram of the intelligent monitoring system for motorcycle overloading provided in an embodiment of the present invention.

[0073] Figure 2 This is a flowchart of the customized motorcycle detection model method provided in the embodiments of the present invention.

[0074] Figure 3 This is a flowchart of the intelligent monitoring method for motorcycle overloading provided in an embodiment of the present invention.

[0075] Figure 4 This is a detailed flowchart of the intelligent monitoring method for motorcycle overloading provided in an embodiment of the present invention. Detailed Implementation

[0076] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0077] like Figure 1 As shown, an intelligent monitoring system for motorcycle overloading provided in this embodiment of the invention includes:

[0078] Motorcycle detection model module 1 is connected to head detection model module 2 and distance judgment module 5, and is used to customize the motorcycle detection model;

[0079] The head detection model module 2 is connected to the motorcycle detection model module 1, the inference module 3, the processing module 4, and the distance judgment module 5, and is used to customize the motorcycle detection model.

[0080] Inference module 3, connected to head detection model module 2 and processing module 4, is used for model inference;

[0081] Processing module 4 is connected to head detection model module 2, inference module 3, and distance judgment module 5, and is used for post-processing;

[0082] The distance judgment module 5 is connected to the motorcycle detection model module 1, the head detection model module 2, and the processing module 4, and is used to judge the distance between the driver's head frame and the vehicle body frame.

[0083] Push module 6, connected to processing module 4, is used to push early warning data to Kafka.

[0084] The working principle of the intelligent monitoring system for overloaded motorcycles:

[0085] First, the motorcycle detection model module 1 performs real-time parsing of the input video stream based on pre-trained and customized datasets to identify the motorcycle's bounding box in the image. This module, as a fundamental component of the system, ensures accurate differentiation between motorcycles and other vehicles in complex traffic scenarios, providing a reliable vehicle baseline for subsequent head recognition and people counting.

[0086] Secondly, under the constraint of the motorcycle detection results, the head detection model module 2 uses a deep learning head detection algorithm to scan the image and extract the head bounding boxes of the rider and passengers. This module not only effectively avoids interference from irrelevant pedestrians in the environment, but also ensures that the detected head objects are indeed on the motorcycle through linkage with the motorcycle bounding box.

[0087] Next, the reasoning module 3 inputs the results from the motorcycle detection module and the head detection module into the reasoning engine, and performs logical judgment based on the scenario rules. Through in-depth fusion and reasoning analysis of the detection results, the system can calculate the actual number of people on the motorcycle and form a preliminary assessment of the risk of overcrowding.

[0088] Then, processing module 4 performs post-processing on the inference results, including data denoising, bounding box optimization, and overcrowding threshold judgment. When the number of heads detected exceeds legal or safety standards, this module will generate an abnormal event identifier and, combined with the analysis results of distance judgment module 5, further verify the matching relationship between the head bounding box and the vehicle bounding box to avoid false detections or false alarms.

[0089] Finally, the processed and verified results are transmitted to push module 6. The system pushes the over-capacity warning data to the Kafka message queue in the form of a message stream, achieving real-time connection with the back-end monitoring platform or traffic enforcement system. In this way, the system can not only automatically identify on-site issues but also upload abnormal data in real time, providing strong support for law enforcement evidence collection, traffic safety management, and accident prevention.

[0090] like Figure 2 As shown, the customized motorcycle detection model provided in this embodiment of the invention:

[0091] S101 collects a large number of real-world scene images, including motorcycles, electric vehicles, bicycles, and tricycles; in these images, motorcycles are used as positive samples, while other vehicles are used as negative samples.

[0092] S102 uses the YOLOv5 framework to train the model, and through continuous fine-tuning, a customized motorcycle detection model is finally obtained.

[0093] The customized motorcycle head detection model provided in this embodiment of the invention:

[0094] 1) Use a motorcycle detection model to crop out the motorcycle portion from the original image and perform manual verification;

[0095] 2) Mark the heads of motorcycle riders to distinguish between drivers and passengers;

[0096] 3) The YOLOv5 framework was used for model training, and through fine-tuning, a customized model focused on detecting the head of a motorcycle rider was obtained.

[0097] The model reasoning provided in this embodiment of the invention:

[0098] 1) After receiving the video stream, the motorcycle is first detected to obtain the motorcycle image and the coordinates of the vehicle body bounding box;

[0099] 2) The detected motorcycle image is cropped by a certain proportion to avoid pedestrians being mistakenly included in the detection range. Then, the motorcycle head detection is performed to obtain the head bounding box coordinates.

[0100] Post-processing provided in this embodiment of the invention:

[0101] 1) Remove incomplete vehicles

[0102] Exclude vehicles that only partially appear in the images to ensure that only complete vehicles are used for analysis;

[0103] 2) Remove overlapping vehicles

[0104] If a vehicle identifies multiple drivers, it means that two motorcycles are squeezed together. In this case, the detected data is incorrect and should be filtered out.

[0105] 3) Determining the distance between the passenger's head frame and the driver's head frame;

[0106] If the x-coordinate of the center point of the passenger's head frame deviates significantly from the x-coordinate of the center point of the driver's head frame (i.e., the difference between the x-coordinate of the passenger's head frame and the x-coordinate of the driver's head frame is greater than 1.3 times the width of the driver's head frame), or if the y-coordinate of the center point of the passenger's head frame is significantly higher than the y-coordinate of the center point of the driver's head frame (i.e., the difference between the y-coordinate of the passenger's head frame and the y-coordinate of the driver's head frame is greater than 1 times the height of the driver's head frame), this application considers such a positional relationship unreasonable and excludes it from further analysis to avoid misjudgment.

[0107] 4) Distance judgment between the driver's head frame and the vehicle body frame

[0108] Pay attention to the proportional relationship between the height of the center point of the human head frame and the height of the center point of the vehicle body frame.

[0109] The early warning data push to Kafka provided in this embodiment of the invention:

[0110] After data cleaning and quality control, the system will count the number of motorcycle passengers. If the number of passengers exceeds a preset threshold, the system will trigger an alert and push it to a Kafka message queue. Subsequently, the business application system will consume this alert data and display it on the interface so that relevant personnel can respond to and handle the overloading situation in a timely manner.

[0111] like Figure 3 As shown in the figure, an intelligent monitoring method for motorcycle overloading provided by an embodiment of the present invention includes:

[0112] S201, customize the motorcycle detection model through the motorcycle detection model module;

[0113] S202, a motorcycle detection model is customized through the human head detection model module;

[0114] S203, Model inference is performed through the inference module;

[0115] S204, post-processing is performed through the processing module;

[0116] S205, the distance judgment module judges the distance between the driver's head frame and the vehicle body frame;

[0117] S206 pushes early warning data to Kafka via the push module.

[0118] In steps S201 and S202, the system receives video signals from the front-end camera as raw input data. The video stream is first fed into the motorcycle detection model module and the head detection model module, where it undergoes feature extraction and classification by a deep convolutional neural network to generate detection signals for the motorcycle body bounding box and the head bounding box. These detection signals are output in a structured data format, including bounding box coordinates, confidence scores, and timestamps.

[0119] In step S203, the inference module performs data fusion and logical reasoning processing on the detected signals. The inference module jointly calculates the spatial relationship between the motorcycle frame and the head frame to obtain the number of heads on the motorcycle. High-speed inference is achieved through tensor quantization and parallel computing, enabling multi-target recognition and headcount counting to be completed in milliseconds, ensuring the real-time performance and accuracy of signal processing.

[0120] In step S204, the processing module receives the inference result signal and performs post-processing. This module includes noise suppression, bounding box overlap elimination, and confidence threshold filtering to reduce false detections and duplicate detection signals. Simultaneously, the processing module performs feature labeling on abnormal signals and converts detection events that meet the over-capacity condition into standardized alarm signals for subsequent transmission.

[0121] In step S205, the distance judgment module calculates the spatial distance between the driver's head frame and the vehicle body frame. This module uses a matching algorithm based on geometric relationships and pixel coordinates to convert the detected signal into a distance matrix. It then compares this matrix with preset threshold conditions and outputs a judgment signal indicating whether the driver's head position is reasonable and the number of people exceeds the limit, thereby further improving the accuracy and reliability of the recognition.

[0122] In step S206, the push module receives the processed and verified warning signal, encapsulates it into a data packet, and pushes it to the backend monitoring platform in real time via the message middleware Kafka. The push signal includes the event number, timestamp, video segment index, and overcrowding determination result, forming a complete monitoring data stream. Through this signal transmission process, the linkage between the front-end detection and back-end monitoring systems is realized, providing data support for traffic safety enforcement and accident prevention.

[0123] like Figure 4 1. Customized motorcycle inspection model

[0124] A large number of real-world scene images were collected, including motorcycles, electric vehicles, bicycles, and tricycles. In these images, motorcycles were used as positive samples, while other vehicles were used as negative samples.

[0125] Using the YOLOv5 framework, this application trained the model and, through continuous fine-tuning, obtained a customized motorcycle detection model that significantly reduced the possibility of misidentifying electric vehicles as motorcycles.

[0126] 2. Customized motorcycle head detection model

[0127] 1) Use a motorcycle detection model to crop out the motorcycle part from the original image and perform manual verification to ensure accuracy.

[0128] 2) This application marks the heads of occupants on motorcycles to distinguish between drivers and passengers.

[0129] 3) The YOLOv5 framework was used for model training, and through fine-tuning, a customized model focused on detecting the head of a motorcycle rider was obtained.

[0130] 3. Model Reasoning

[0131] 1) After receiving the video stream, the motorcycle is first detected to obtain the motorcycle image and the coordinates of the vehicle body frame.

[0132] 2) The detected motorcycle image is cropped by a certain proportion to avoid pedestrians being mistakenly included in the detection range. Then, the motorcycle head detection is performed to obtain the head bounding box coordinates.

[0133] 4. Post-processing

[0134] 1) Remove incomplete vehicles

[0135] Exclude vehicles that only partially appear in the images to ensure that only complete vehicles are used for analysis.

[0136] 2) Remove overlapping vehicles

[0137] If a vehicle identifies multiple drivers, it means that two motorcycles are squeezed together. In this case, the detected data is incorrect and should be filtered out.

[0138] 3) Distance judgment between passenger head frame and driver head frame

[0139] If the x-coordinate of the center point of the passenger's head frame deviates significantly from the x-coordinate of the center point of the driver's head frame (i.e., the difference between the x-coordinate of the passenger's head frame and the x-coordinate of the driver's head frame is greater than 1.3 times the width of the driver's head frame), or if the y-coordinate of the center point of the passenger's head frame is significantly higher than the y-coordinate of the center point of the driver's head frame (i.e., the difference between the y-coordinate of the passenger's head frame and the y-coordinate of the driver's head frame is greater than 1 times the height of the driver's head frame), this application considers such a positional relationship unreasonable and excludes it from further analysis to avoid misjudgment.

[0140] 4) Distance judgment between the driver's head frame and the vehicle body frame

[0141] This application focuses on the proportional relationship between the height of the center point of the human head frame and the height of the center point of the vehicle body frame. To avoid the uncertainty caused by relying on absolute values, this application sets a relative threshold: the height of the center point of the human head frame must be greater than twice the height of the vehicle body frame, but not more than five times.

[0142] 5. Alert data push to Kafka

[0143] After data cleaning and quality control, the system will count the number of motorcycle passengers. Based on a preset threshold, if the calculated number exceeds the threshold, the system will trigger an alert and push it to a Kafka message queue. Subsequently, the business application system will consume this alert data and display it on the interface so that relevant personnel can respond and handle overloading situations promptly.

[0144] Example 1: Public Security Electronic Police Capture Products

[0145] Product Form: An "edge smart box" (JetsonXavier + this invention's algorithm) is attached to existing urban traffic enforcement poles, utilizing the video stream from existing 2-megapixel PTZ cameras. No road construction or additional sensors are required. Deployment Results: In March 2024, 32 sets were installed on a 14km test section of G312 National Highway in Suzhou. After 90 days of continuous operation, 382,614 motorcycles were captured, generating 7,823 overload warnings. Of these, 7,548 were manually verified, resulting in a false alarm rate of 2.9% and a false negative rate of 1.4%. Each set generated an average of 2.8 valid tickets per day, with a payback period of 3.2 months per set.

[0146] Technical Value: For the first time, "ordinary public security video" has been upgraded to "law enforcement-grade evidence source," directly proving that the "motorcycle detection model + head detection model + distance judgment module" described in claim 1 can simultaneously meet the three major requirements of accuracy, cost, and real-time performance after being commercialized, demonstrating its inventiveness.

[0147] Example 2: Highway Entrance Dissuasion Product

[0148] Product Formulation: A 5-megapixel camera and the algorithm box of this invention are installed on the side of the ETC antenna pole at the toll station, linked with the toll collection system. When the system detects an overloaded motorcycle, the automatic barrier fails to lift and a voice warning is issued to turn it back. Deployment Results: Trial operation was conducted at Huizhou Station of the Guanghui Expressway in May 2024. The average daily traffic flow was 4,200 motorcycles, and 127 overloaded vehicles were turned back per day, with a 100% success rate and no complaints. Compared with the manual inspection mode, it saves 5 police officers per shift, resulting in annual labor cost savings of over 900,000 yuan.

[0149] Technical value: This proves that the "model reasoning" in claim 4 and the "post-processing" in claim 5 can complete the "detection-counting-linkage" closed loop within 600ms, overcoming the bias of traditional solutions that require additional infrared or weighing sensors, and demonstrating inventiveness.

[0150] Evidence related to the technical effects obtained by the embodiments of the present invention.

[0151] 1. Comparison Test of Vehicle Classification Accuracy

[0152] Experimental Design:

[0153] Test set: 10,087 frames were randomly selected from traffic police electronic surveillance videos from Suzhou, Huizhou, and Yuncheng in 2024Q1, including 3,421 motorcycles, 2,876 electric bicycles, 2,530 bicycles, and 1,260 tricycles.

[0154] Evaluation metrics: Accuracy (Acc), False Positive Rate (FPR).

[0155] Benchmarks: ① YOLOv5 official model (COCO); ② This invention's negative sample enhancement model.

[0156] result:

[0157] Table 1. Comparison Test Results of Vehicle Classification Accuracy

[0158]

[0159] Conclusion: The negative sample enhancement scheme reduces the number of "electric vehicles being mistakenly identified as motorcycles" by an order of magnitude, meeting law enforcement requirements.

[0160] 2. Recall rate test for head detection

[0161] Experimental Design:

[0162] 1847 human heads (minimum 32×32px) were manually labeled as "visible only at the top of the helmet".

[0163] The model presented in this paper, "6-layer pyramid + small target anchor", is compared with the original YOLOv5-s.

[0164] Metric: Recall@0.5.

[0165] result:

[0166] Original model Recall = 69.1%

[0167] The recall rate of this invention is 91.7%.

[0168] This represents an increase of 22.6 percentage points, directly addressing the problem of "passengers being obscured and missed during detection."

[0169] 3. Post-geometric verification filtering effect

[0170] Test method:

[0171] A total of 1,293 cases were identified as "false positives" by manually marking the heads of pedestrians and vehicles traveling side by side on the road.

[0172] Run two separate processes, one with and one without geometric verification, and count the remaining false alarms.

[0173] result:

[0174] 1293 false alarms without verification

[0175] 802 false alarms remained after geometric verification.

[0176] Filtration rate = 38.0%

[0177] The creative value of the rule "x offset ≤ 1.3 times head width & y offset ≤ 1 times head height" has been verified.

[0178] 4. On-site joint debugging of end-to-end accuracy, speed, and power consumption.

[0179] Test environment: 14km of G312 National Highway, 32 edge boxes, 90 consecutive days and nights.

[0180] Evaluation method: Traffic police motorcycle escorts collecting evidence are considered Ground Truth.

[0181] result:

[0182] Table 2 Comparison of recall rates for head detection (difficult samples)

[0183]

[0184] In summary, the multi-dimensional experimental data and third-party acceptance reports jointly prove that the present invention is significantly superior to the existing technology in the three core indicators of "accuracy-speed-power consumption", and the effect is quantifiable and reproducible, providing solid evidence for its inventiveness.

[0185] It should be noted that embodiments of the present invention can be implemented in hardware, software, or a combination of both. The hardware portion can be implemented using dedicated logic; the software portion can be stored in memory and executed by a suitable instruction execution system, such as a microprocessor or dedicated-design hardware. Those skilled in the art will understand that the above-described devices and methods can be implemented using computer-executable instructions and / or included in processor control code, for example, such code provided on a carrier medium such as a disk, CD, or DVD-ROM, a programmable memory such as read-only memory (firmware), or a data carrier such as an optical or electronic signal carrier. The devices and modules of the present invention can be implemented by hardware circuitry such as very large-scale integrated circuits or gate arrays, semiconductors such as logic chips, transistors, or programmable hardware devices such as field-programmable gate arrays, programmable logic devices, etc., or by software executed by various types of processors, or by a combination of the above-described hardware circuitry and software, such as firmware.

[0186] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any modifications, equivalent substitutions, and improvements made by those skilled in the art within the scope of the technology disclosed in the present invention, and within the spirit and principles of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A motorcycle overload monitoring system, characterized in that, The intelligent monitoring system for motorcycle overloading includes: Motorcycle detection model module, head detection model module, inference module, processing module, distance judgment module, push module; The motorcycle detection model module, connected to the head detection model module and the distance judgment module, is used to customize the motorcycle detection model. The head detection model module is connected to the motorcycle detection model module, inference module, processing module, and distance judgment module, and is used to customize the motorcycle detection model. The inference module, connected to the head detection model module and the processing module, is used for model inference. The processing module, connected to the head detection model module, inference module, and distance judgment module, is used for post-processing. The distance judgment module is connected to the motorcycle detection model module, the head detection model module, and the processing module 4, and is used to judge the distance between the driver's head frame and the vehicle body frame. The push module, connected to the processing module, is used to push early warning data to Kafka.

2. The intelligent monitoring system for motorcycle overloading as described in claim 1, characterized in that, The customized motorcycle detection model: Collect a large number of real-world scene images, including motorcycles, electric vehicles, bicycles, and tricycles; in these images, motorcycles are used as positive samples, while other vehicles are used as negative samples. The model was trained using the YOLOv5 framework, and through continuous fine-tuning, a customized motorcycle detection model was finally obtained.

3. The intelligent monitoring system for motorcycle overloading as described in claim 1, characterized in that, The customized motorcycle head detection model: 1) Use a motorcycle detection model to crop out the motorcycle portion from the original image and perform manual verification; 2) Mark the heads of motorcycle riders to distinguish between drivers and passengers; 3) The YOLOv5 framework was used for model training, and through fine-tuning, a customized model focused on detecting the head of a motorcycle rider was obtained.

4. The intelligent monitoring system for motorcycle overloading as described in claim 1, characterized in that, The model inference: 1) After receiving the video stream, the motorcycle is first detected to obtain the motorcycle image and the coordinates of the vehicle body bounding box; 2) The detected motorcycle image is cropped by a certain proportion to avoid pedestrians being mistakenly included in the detection range. Then, the motorcycle head detection is performed to obtain the head bounding box coordinates.

5. The intelligent monitoring system for motorcycle overloading as described in claim 1, characterized in that, The post-processing: 1) Remove incomplete vehicles Exclude vehicles that only partially appear in the images to ensure that only complete vehicles are used for analysis; 2) Remove overlapping vehicles If a vehicle identifies multiple drivers, it means that two motorcycles are squeezed together. In this case, the detected data is incorrect and should be filtered out. 3) Determining the distance between the passenger's head frame and the driver's head frame; If the x-coordinate of the center point of the passenger's head frame deviates significantly from the x-coordinate of the center point of the driver's head frame (i.e., the difference between the x-coordinate of the passenger's head frame and the x-coordinate of the driver's head frame is greater than 1.3 times the width of the driver's head frame), or if the y-coordinate of the center point of the passenger's head frame is significantly higher than the y-coordinate of the center point of the driver's head frame (i.e., the difference between the y-coordinate of the passenger's head frame and the y-coordinate of the driver's head frame is greater than 1 times the height of the driver's head frame), this application considers such a positional relationship unreasonable and excludes it from further analysis to avoid misjudgment. 4) Distance judgment between the driver's head frame and the vehicle body frame Pay attention to the proportional relationship between the height of the center point of the human head frame and the height of the center point of the vehicle body frame.

6. The intelligent monitoring system for motorcycle overloading as described in claim 1, characterized in that, The early warning data is pushed to Kafka: After data cleaning and quality control, the system will count the number of motorcycle riders. If the number of riders exceeds a preset threshold, the system will trigger an alert and push it to the Kafka message queue. Subsequently, the business application system will consume this early warning data and display it on the interface so that relevant personnel can respond to and handle the situation of overloading in a timely manner.

7. A method for intelligent monitoring of motorcycle overloading, implementing the intelligent monitoring system for motorcycle overloading as described in any one of claims 1-6, characterized in that, The intelligent monitoring method for motorcycle overloading includes: Step 1: Customize the motorcycle detection model using the motorcycle detection model module; Step 2: Customize the motorcycle detection model using the human head detection model module; Step 3: Perform model reasoning through the reasoning module; Step 4: Post-processing is performed through the processing module; Step 5: Determine the distance between the driver's head frame and the vehicle body frame using the distance judgment module; Step 6: Push the alert data to Kafka via the push module.

8. A computer device, characterized in that, The computer device includes a memory and a processor. The memory stores a computer program, which, when executed by the processor, causes the processor to perform the steps of the intelligent monitoring method for overloading motorcycles as described in claim 7.

9. A computer-readable storage medium storing a computer program, which, when executed by a processor, causes the processor to perform the steps of the intelligent monitoring method for overloading motorcycles as described in claim 7.

10. An information data processing terminal, characterized in that, The information data processing terminal is used to implement the intelligent monitoring system for overloaded motorcycles as described in any one of claims 1-6.

Images