A parking lot oil theft detection system based on computer vision technology

Abstract

This invention discloses a parking lot oil theft detection system based on computer vision technology. It is characterized by comprising: a camera module, installed within the parking lot to acquire image data; and an image processing module, which preprocesses and analyzes the image data acquired by the camera to extract feature information of vehicles and personnel. This invention achieves automatic monitoring of vehicles within the parking lot using computer vision technology, helping managers better understand vehicle usage and promptly identify and resolve problems. This parking lot oil theft detection system, based on computer vision technology, uses deep learning algorithms to track human trajectories. By tracking human trajectories, it can monitor personnel activity in real time, determine the presence of suspicious behavior, and, combined with human posture analysis, analyze illicit oil theft activities.

Description

Technical Field

[0001] This invention relates to the field of computer vision technology, specifically to a parking lot oil theft detection system based on computer vision technology. Background Technology

[0002] Computer vision is a science that studies how to make machines "see". More specifically, it refers to machine vision, which uses cameras and computers to replace human eyes to identify, track and measure targets, and then performs image processing to make the computer-processed images more suitable for human observation or transmission to instruments for detection.

[0003] CN114445855A discloses a method and device for detecting fuel theft from a car's fuel tank based on sentry mode. This method acquires pedestrian images of the area to be detected, first identifying suspected fuel thieves according to preset conditions, and then determining the actual fuel thieves based on the time these suspected thieves spend in the protected and alarm areas. This efficient method ensures the accuracy of fuel theft detection, prevents fuel theft, and improves vehicle parking security.

[0004] With the continuous rise in gasoline prices, the probability of gasoline theft in parking lots is also increasing. Currently, traditional gasoline theft detection methods usually require manual patrols or the use of pressure sensors installed on the ground to monitor the status of fuel tanks. However, these methods often require a lot of manpower or cost and are easily affected by environmental factors. Compared with the above-mentioned technical solutions, which simply distinguish gasoline thieves by detecting the time when the person is in the image, the accuracy is not high. Furthermore, judging gasoline theft behavior by the pedestrian feature vectors of several pedestrian images obtained within a preset time period is too broad. To solve the above problems, a parking lot gasoline theft detection system based on computer vision technology is urgently needed. Summary of the Invention

[0005] The purpose of this invention is to provide a parking lot oil theft detection system based on computer vision technology, which has the advantages of automatic monitoring and reduced labor costs, and solves the problems mentioned in the background technology.

[0006] To achieve the above objectives, the present invention provides the following technical solution: a parking lot oil theft detection system based on computer vision technology, comprising:

[0007] Camera module: Cameras are installed in the parking lot to acquire image data of the parking lot.

[0008] The image processing module preprocesses and analyzes the image data acquired by the camera, extracting feature information of vehicles and people. The image processing module is configured as follows:

[0009] (i) Trajectory tracking algorithm: Person detection is achieved by using the single-stage target detection algorithm YOLOv5. After outputting the bounding box position, each bounding box is assigned a random unique identifier. A multi-target tracker is used to track the bounding boxes in consecutive frames. A vote is performed every 10 frames. If the IOU of the corresponding bounding boxes in 5 of the 10 frames is greater than the threshold, then the same unique identifier is assigned to this series of bounding boxes in these 10 frames, thereby realizing the reserve of target IDs.

[0010] (II) Crouching Detection Algorithm: The output of the trajectory tracking algorithm serves as the basis for detecting crouching behavior. Based on the bounding box output by the trajectory tracking algorithm, the corner points of the rectangular box enclose the target. The corner points include, but are not limited to, the upper left and lower right corners of the bounding box. The center point coordinates c and aspect ratio r of the target's bounding box are calculated. According to the tracking algorithm, the unique identifier of the target is obtained. For each target, the following information is recorded: P(frame) = [unique identifier, center point coordinates c, aspect ratio r, category, frame_ID], where frame_ID is the unique name of the video frame. As time or frame rate changes, P is continuously updated. The change of P in consecutive frames is calculated using the difference method. If the y-coordinate of the center point changes abruptly, the center of gravity sinks, and at the same time, the aspect ratio r also changes abruptly, and the height decreases. At this time, the target exhibits crouching behavior.

[0011] (iii) A loitering detection algorithm is used to determine loitering behavior. If a person stays under a camera for more than a certain threshold, it is determined that the person is loitering.

[0012] (iv) Personnel wearing detection algorithm: Image classification algorithm is used to judge personnel wearing; For authorized loitering personnel, work clothes of a specific color are issued, and staff are asked to wear the designated work clothes and walk in the warehouse in advance. The background completes data collection, and two types of image data, safety vest and non-safety vest, are set up. The googLeNet algorithm is trained to classify the images. If the classification result is the designated work clothes, it is determined to be a patrol personnel and excluded from the list of suspected oil theft personnel.

[0013] (v) Engine hood detection algorithm: Based on the fact that people stealing oil must open the engine hood, GoogLeNet is used to classify whether the engine hood of the vehicle is open.

[0014] The vehicle status comparison module compares and analyzes the current status of people in the vehicle's surrounding environment with historical data to determine whether fuel theft has occurred.

[0015] The alarm notification module automatically sends an alarm notification to the management personnel when an oil theft incident is detected, and records the time and location details of the incident.

[0016] The data storage module is used to store historical data and records of oil theft incidents.

[0017] Preferably, in the trajectory tracking algorithm, when a person is occluded by a large vehicle while walking in a parking lot, each detected category is first classified and tracked separately. Different categories use different tracking systems to ensure that trajectory calculations are completed in different vector spaces. Secondly, during the partial occlusion process before complete occlusion, the target and vehicles in the foreground can still be detected. At this time, the IOU value between the target and the vehicle is calculated. However, when this value exceeds threshold_occlusion=0.5, the tracking_frames parameter is automatically doubled to tracking_frames=20 to improve the tracker's tolerance and ultimately improve the survival rate of bounding box IDs, thus achieving person tracking under occlusion conditions.

[0018] Preferably, the squatting detection algorithm may also use a skeleton or a cascaded classification algorithm.

[0019] Preferably, in the personnel loitering detection algorithm, the threshold judgment logic for dwell time is as follows: According to the warehouse scenario, the main work of personnel in the parking lot is to retrieve and drive the car. The walking time of personnel in the same camera frame will not exceed 1 minute. If the unique identifier of a target survives for more than 1 minute or 1400 consecutive frames, it is determined that the personnel are loitering.

[0020] Preferably, in the crouching detection algorithm, threshold_c is generally set to (width + height) / 2*20% of the target, and threshold_r is set to r*20%, that is, a target whose center of gravity drops by 20% and whose height decreases by 20% in each frame is judged as a "crouching" behavior.

[0021] Preferably, to ensure the algorithm can run in real time, the following optimization methods are adopted:

[0022] Lightweight Neural Networks: Using lightweight neural network models, such as MobileNet and ShuffleNet, can speed up model training and inference by having fewer parameters and less computation.

[0023] Hardware optimization: GPU hardware acceleration technology is used to improve the training and inference speed of the model. Nvidia's GPU and TensorRT acceleration library are used to double the running speed.

[0024] Data augmentation: Data augmentation techniques can be used to expand datasets, increase the diversity and quantity of data, thereby improving the accuracy of models;

[0025] Model distillation: Model distillation can be used to distill large models into smaller models, thereby improving the speed and efficiency of the model, while also reducing the requirements for model storage space and computing resources.

[0026] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0027] This invention utilizes computer vision technology to achieve automatic monitoring of vehicles in parking lots, reducing the cost and workload of manual patrols while improving the accuracy and real-time performance of monitoring. Furthermore, comparing and analyzing historical data further enhances detection accuracy and helps managers better understand vehicle usage within the parking lot, enabling timely identification and resolution of problems. This computer vision-based parking lot oil theft detection system processes surveillance video within the parking lot using deep learning algorithms to track human trajectories. By tracking these trajectories, it can monitor personnel activity in real time, identify suspicious behavior, and, combined with human posture analysis, analyze illicit activities such as oil theft. Ultimately, it achieves automatic monitoring of personnel and vehicles within the parking lot, avoiding the cost and workload of manual patrols while improving monitoring accuracy and real-time performance. Compared to traditional manual patrols, this technology not only improves efficiency but also better protects personnel safety. Attached Figure Description

[0028] Figure 1 This is a schematic diagram of the logic of the present invention. Detailed Implementation

[0029] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0030] Please see Figure 1 This invention provides a technical solution: a parking lot oil theft detection system based on computer vision technology, comprising:

[0031] A camera module is used to install cameras in parking lots to acquire image data of the parking lot area.

[0032] The image processing module preprocesses and analyzes the image data acquired by the camera, extracting feature information of vehicles and people. The image processing module includes a GPU module, a CPU module, and a cache module.

[0033] (i) Trajectory tracking algorithm: Personnel detection is achieved by using the single-stage target detection algorithm YOLOv5. After outputting the bounding box position, each bounding box is assigned a random unique identifier. A multi-target tracker is used to track the bounding boxes in consecutive frames. A vote is performed every 10 frames. If the IOU of the corresponding bounding boxes in 5 of the 10 frames is greater than the threshold, then the same unique identifier is assigned to this series of bounding boxes in these 10 frames, thereby realizing the storage of target IDs.

[0034] (II) Crouching Detection Algorithm: The trajectory tracking algorithm output serves as the basis for crouching behavior detection. Specifically, based on the bounding box output by the trajectory tracking algorithm (the rectangle enclosing the target's corners, including but not limited to the top left and bottom right corners), the center point coordinates c and aspect ratio r of the target's bounding box are calculated. According to the tracking algorithm, a unique identifier for the target is obtained. For each target, the following information is recorded: P(frame) = [unique identifier, center point coordinates c, aspect ratio r, category, frame_ID], where frame_ID is the unique name of the frame in this video, changing over time or... As the frame rate changes, P is constantly updated. The difference method is used to calculate the change of P in consecutive frames. If the y-coordinate of the center point changes abruptly, that is, the center of gravity drops, i.e., delta_c(y)threshold_c, and at the same time, the aspect ratio r also changes abruptly, i.e., the height becomes shorter, i.e., delta_r>threhold_r, the target exhibits crouching behavior. Threshold_c is generally set to (width + height) / 2*20% of the target, and threshold_r is set to r*20%. That is, a target whose center of gravity drops by 20% and whose height decreases by 20% every frame is judged as "crouching" behavior.

[0035] (III) Person loitering detection algorithm: determining that a person loitering is defined as the time a person spends under a certain camera exceeds a certain threshold, that is, determining that the person has loitering behavior. In the person loitering detection algorithm, the threshold judgment logic for the time spent is as follows: According to the warehouse scenario, the main work of people in the parking lot is to retrieve and drive cars. People will not walk for more than 1 minute in the same camera frame. If the unique identifier of a target survives for more than 1 minute or 1400 consecutive frames, it is determined that the person has loitering behavior.

[0036] (iv) Personnel Wearing Detection Algorithm: The personnel wearing detection module adopts an image classification algorithm. For authorized loitering personnel, a work uniform of a specific color is issued, usually a bright yellow safety vest. The staff are asked to wear the designated work uniform and walk around the warehouse in advance. The background completes data collection and creates two datasets of image data: one for safety vests and one for non-safety vests. The googLeNet algorithm is trained to classify the images. If the classification result is the designated clothing, the person is judged to be a patrol personnel and is excluded from the list of suspected oil thieves.

[0037] (v) Engine hood detection algorithm: Based on the fact that people stealing oil must open the engine hood, GoogLeNet is used to classify whether the engine hood of the vehicle is open.

[0038] The vehicle status comparison module compares and analyzes the current status of people and the surrounding environment of the vehicle with historical data to determine whether fuel theft has occurred.

[0039] The alarm notification module automatically sends an alarm notification to the management personnel when an oil theft incident is detected, and records the time and location details of the incident.

[0040] The data storage module is used to store historical data and records of oil theft incidents.

[0041] People walking in parking lots are often obstructed by large vehicles. When a target is obstructed by a vehicle, the trajectory tracking algorithm will temporarily fail, the target's unique identifier will disappear from the ID pool, and the multi-target tracker will be unable to track it. The target ID will be lost, which will also cause the trajectory to be broken.

[0042] In this invention: when a person is occluded by a large vehicle or object while walking in a parking lot in the trajectory tracking algorithm: firstly, each detected category is categorized and tracked. Different categories use different tracking systems to ensure that trajectory calculations are completed in different vector spaces. Secondly, the "occlusion" of the target during movement is a dynamic problem. During partial occlusion before complete occlusion, the target and vehicles in the foreground can still be detected. At this time, the IOU value between the target and the vehicle is calculated. However, when this value exceeds threshold_occlusion=0.5, the tracking_frames parameter is automatically doubled to tracking_frames=20 to improve the tracker's tolerance and ultimately improve the survival rate of bounding box IDs, thus achieving personnel tracking under occlusion conditions.

[0043] In this invention, the squatting detection algorithm can also use a body skeleton or a cascaded classification algorithm.

[0044] In this invention, to ensure the algorithm can run in real time, the following optimization methods are adopted:

[0045] Lightweight Neural Networks: Using lightweight neural network models, such as MobileNet and ShuffleNet, can speed up model training and inference by having fewer parameters and less computation.

[0046] Hardware optimization: GPU hardware acceleration technology is used to improve the training and inference speed of the model. Nvidia's GPU and TensorRT acceleration library are used to double the running speed.

[0047] Data augmentation: Data augmentation techniques can be used to expand datasets, increase the diversity and quantity of data, thereby improving the accuracy of models;

[0048] Model distillation: Model distillation can be used to distill large models into smaller models, thereby improving the speed and efficiency of the model, while also reducing the requirements for model storage space and computing resources.

[0049] It should be noted that, in this document, relational terms such as "first" and "second" are used only to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such process, method, article, or apparatus.

[0050] All electrical components mentioned in this article are connected to an external main controller and 220V AC mains power, and the main controller can be a conventional known device such as a computer that can control it.

[0051] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A computer vision technology based detection system for siphoning in parking lots, characterized in that, The application relates to a parking lot oil stealing behavior detection system, which comprises the following modules: a camera module, which is installed in a parking lot and used for acquiring image data in the parking lot; an image processing module, which is used for pre-processing and analyzing the image data acquired by the camera and extracting feature information of vehicles and personnel, and is configured as follows: (1) a trajectory tracking algorithm, which realizes personnel detection based on a single-stage target detection algorithm YOLOv5, and after outputting a boundary box position, each boundary box is assigned a random unique identifier, a multi-target tracker is used to track the boundary boxes in continuous frames, and voting is performed once every tracking_frames=10 frames; if the IOU of the corresponding boundary boxes in 5 frames of the 10 frames is greater than a threshold value, the same unique identifier is assigned to the series of boundary boxes in the 10 frames, so that the target ID is reserved; (2) a crouching detection algorithm, which uses the output of the trajectory tracking algorithm as the basis for crouching behavior detection, calculates the center point coordinates c and the height-width ratio r of the target according to the corner points of the rectangular frame of the target contained in the boundary box output by the trajectory tracking algorithm, and records the following information P (frame) = [unique identifier, center point coordinates c, height-width ratio r, category, frame_ID] for each target, wherein frame_ID is the unique name of the frame of the video; P is continuously updated with the change of time or frame rate, and the change of P in continuous frames is calculated by using a difference method; if the y coordinate of the center point suddenly changes, the center of gravity sinks, and the height-width ratio r also suddenly changes, the height becomes lower, and the target appears a crouching behavior; (3) a personnel loitering detection algorithm, which is used for judging personnel loitering behavior; if a personnel stays under a camera for more than a threshold value, it is determined that the personnel has a loitering behavior; (4) a personnel wearing detection algorithm, which uses an image classification algorithm to judge personnel wearing; for authorized loitering personnel, a specific color of work clothes is issued, and the staff is required to wear the specified work clothes to walk in the warehouse in advance, and the background completes data acquisition, formulates a data set of two types of image data of safety vests and non-safety vests, trains a googLeNet algorithm to classify the images, and if the classification result is the specified work clothes, the result is determined as a patrol personnel and is excluded from the list of suspicious oil stealing personnel; (5) an engine cover detection algorithm, which is based on the personnel oil stealing behavior that must open the engine cover and uses GoogLeNet to classify whether the engine cover of the vehicle is open; a vehicle state comparison module, which compares and analyzes the current vehicle surrounding environment personnel state with historical data to judge whether the oil tank has an oil stealing event; an alarm notification module, which automatically sends an alarm notification to the management personnel when detecting the oil stealing event, and records the time and place details of the event; a data storage module, which is used for storing historical data and oil stealing event record information.

2. The system according to claim 1, wherein: In the trajectory tracking algorithm, when a person is blocked by a large object during walking in the parking lot, first, each detected category is classified and tracked, different categories use different tracking systems to ensure that trajectory calculation is completed in different vector spaces; second, during the partial occlusion before complete occlusion, the target and the vehicle in the foreground can still be detected, at this time the IOU value of the target and the vehicle is calculated, but when this value exceeds threshold_occlusion=0.5, the tracking_frames parameter is automatically doubled to tracking_frames=20, improving the tolerance of the tracker and ultimately improving the survival rate of the bounding box ID to achieve personnel tracking under occlusion.

3. The system as claimed in claim 1, wherein the system is based on computer vision technology. The crouching detection algorithm can also use body skeleton or re-cascade classification algorithm.

4. The system according to claim 1, wherein: In the personnel loitering detection algorithm, the threshold judgment logic of the stay time is: according to the warehouse scene, the main work of the personnel in the parking lot is to take the car and drive the car, and the personnel walking in the same camera picture will not exceed 1 minute, if the unique identifier of a target survives for more than 1 minute or 1400 consecutive frames, it is determined that the personnel have loitering behavior.

5. The system as claimed in claim 1, wherein the system is based on computer vision technology. In the crouching detection algorithm, threshold_c is set to target (width+height) / 2*20%, and threshold_r is set to r*20%, that is, a target with a 20% reduction in height and a 20% reduction in center of gravity is determined as a "crouching" behavior.

6. The computer vision technology based siphoning detection system for parking lots as claimed in claim 1 wherein: In order to ensure that the algorithm can run in real time, the following optimization methods are adopted: Lightweight neural network: use lightweight neural network models, including MobileNet and ShuffleNet, which have fewer parameters and less computation, can speed up the training and inference speed of the model; Hardware optimization: use GPU hardware acceleration technology to improve the training and inference speed of the model, use Nvidia's GPU and TensorRT acceleration library, the running speed is doubled; Data enhancement: use data enhancement technology to expand the data set, increase the diversity and quantity of data, and improve the accuracy of the model; Model distillation: use model distillation technology to distill large models into small models, thereby improving the speed and efficiency of the model, while also reducing the model storage space and computing resource requirements.

Images