A system and method for identifying abnormal behavior in a station based on multi-scale vision

Abstract

The present invention relates to the technical field of railway safety assurance, and specifically to a system and method for identifying abnormal behavior in a station based on multi-scale vision. The system consists of an image acquisition and preprocessing module, a feature extraction and fusion module, a behavior recognition module and an alarm module; the image acquisition and preprocessing module is responsible for collecting video data in a railway station from a monitoring camera or other image acquisition equipment, and comprises a digital-to-analog converter, an image acquisition card, and a data transmission interface component, which is used to convert a real-time scene into a digital image signal for processing; the beneficial effects are as follows: the system and method for identifying abnormal behavior in a station based on multi-scale vision proposed by the present invention further solve the problems of low accuracy and efficiency in the existing recognition of multiple categories of abnormal behaviors in complex scenes; accurate recognition of abnormal behaviors from large-scale to small-scale can be achieved, and strong technical support is provided for safety monitoring in railway stations.

Description

Technical Field

[0001] The present invention relates to the field of railway safety assurance, and in particular to a system and method for identifying abnormal behavior within a station based on multi-scale vision. Background Art

[0002] Abnormal behaviors in railway stations include (1) safety-threatening behaviors; (2) illegal activities; (3) emergencies; and (4) uncivilized behaviors. These abnormal behaviors are characterized by diverse categories, complex recognition backgrounds, and variable target sizes. However, traditional machine vision methods have difficulty in fine-grained classification of multi-category abnormal behaviors, and have difficulty separating human bodies from complex backgrounds. In addition, the detection accuracy of small target samples is low.

[0003] In the existing technology, the invention patent "A behavior recognition method driven by synchronous and cross-domain asynchronous fusion" (patent number CN202211433310.3) does not have a structure for identifying small target samples, making it difficult to achieve accurate recognition. The invention patent "A dangerous behavior recognition method based on deep reinforcement learning" (patent number CN202210362415.8) can only realize a single behavior, such as the recognition of dangerous actions such as raising or shaking hands, and does not have the ability to recognize multiple types of abnormal behaviors. At the same time, there is currently a lack of abnormal behavior recognition methods for railway stations.

[0004] Therefore, we need a system and method for identifying abnormal behavior within the station based on multi-scale vision to solve the existing problems of low accuracy and efficiency in identifying multiple categories of abnormal behaviors in complex scenarios; it can achieve accurate identification of abnormal behaviors from large-scale to small-scale, and provide strong technical support for safety monitoring within railway stations. Summary of the invention

[0005] The purpose of the present invention is to provide a system and method for identifying abnormal behavior in a station based on multi-scale vision, so as to solve the problems of low accuracy and efficiency in the existing recognition of multiple categories of abnormal behaviors in complex scenes proposed in the above-mentioned background technology; it can realize accurate recognition of abnormal behaviors from large-scale to small-scale, and provide strong technical support for safety monitoring in railway stations.

[0006] To achieve the above-mentioned purpose, the present invention provides the following technical solutions: a system for identifying abnormal behavior in a station based on multi-scale vision, the system comprising an image acquisition and preprocessing module, a feature extraction and fusion module, a behavior recognition module and an alarm module;

[0007] The image acquisition and preprocessing module is responsible for collecting video data in the railway station from monitoring cameras or other image acquisition devices, including a digital-to-analog converter, an image acquisition card, and a data transmission interface component, which are used to convert real-time scenes into digital image signals for processing; the monitoring camera is rotatably connected to a rotating seat, the rotating seat is rotatably connected to the top of the stage, a cleaning plate is slidably provided inside the rotating seat, the stage is fixed on the surface of the lifting plate, the lifting plate is movably plugged into the installation frame, one side of the lifting plate is provided with teeth, the teeth and the gear disc are meshed, the gear disc passes through the side plate of the installation frame, and the gear disc is rotatably connected to the protective frame, one end of the protective frame is movably plugged with an oil storage box, the interior of the oil storage box is filled with lubricating oil, the surface of the oil storage box is provided with a second perforation, the lubricating oil seeps out from the second perforation and is smeared on the gear disc, a display screen one is fixed on the surface of the installation frame, and a display screen two is fixed on the surface of the lifting plate.

[0008] The feature extraction and fusion module is connected to the image acquisition and preprocessing module and the behavior recognition module, receives the collected signals, extracts and fuses the features of the preprocessed video information, and facilitates the classification of the behavior recognition module; it is composed of an edge computing graphics processing unit, a cache unit, and a data wireless transmission unit;

[0009] The behavior recognition module is connected with the feature extraction and fusion and alarm modules, receives multi-scale abnormal behavior features within the station, and accelerates the response to emergency situations by synchronizing the information to relevant managers in other stations; it consists of a central graphics processing unit, a cache unit, a data wireless transmission unit, and a data transmission interface;

[0010] The alarm module is connected to the behavior recognition module, stores the collected abnormal situation images and abnormal situation categories, and issues a sound alarm; it is composed of a control chip, a storage controller, a data wireless transmission unit, an alarm and an interface circuit.

[0011] Preferably, a storage groove is provided on the top surface of the rotating seat, and connecting shafts are fixed on both sides of the monitoring camera, and the two connecting shafts respectively pass through the two sides of the storage groove, and a servo motor 1 is fixed on the surface of the rotating seat, and a driving shaft of the servo motor 1 is fixed on the surface of a connecting shaft, and an assembly groove is provided on the bottom surface of the storage groove, and a plug rod is fixed on the surface of the assembly groove, and the plug rod is inserted in a perforation 1, and the perforation 1 is provided on the surface of the cleaning plate, and the cleaning plate is a "T"-shaped plate structure, and bristles and a sponge wiper are fixed on the top surface of the cleaning plate, and two elastic traction plates are fixed at one end of the cleaning plate, and the two elastic traction plates are both slidably connected in the assembling groove, and a card strip is fixed on the surface of the elastic traction plate, and the card strip is inserted in the card slot, and the card slot is provided on the side wall of the assembling groove, and the height of the card slot is less than the height of the assembling groove, and a pull ring is fixed on the surface of the elastic traction plate, and the pull ring and the card strip are symmetrically distributed about the elastic traction plate.

[0012] Preferably, a servo motor 2 is fixed to the bottom surface of the loading platform, and a rotating shaft of the servo motor 2 passes through the loading platform and is fixed to the bottom surface of the rotating seat, the loading platform is slidably connected in the groove, the groove is provided on the top surface of the mounting frame, a baffle is fixed between two parallel side walls of the groove, a sponge wiper plate 2 is fixed on the surface of the baffle, after the lifting plate is retracted into the mounting frame, the sponge wiper plate 2 is clamped between the display screen 2 and the baffle to generate elastic deformation, the height of the sponge wiper plate 2 is greater than the height of the display screen 1, a side groove is provided on the surface of the mounting frame, and a plurality of connecting angle plates are fixed on the surface of the side groove.

[0013] Preferably, an embedding groove is provided on one side of the lifting plate, the height of the embedding groove is smaller than the height of the lifting plate, a plurality of teeth are provided, and the plurality of teeth are arranged at equal distances and sizes and fixed inside the embedding groove, the mounting frame is a square tube structure, a reserved opening is provided on the surface of the mounting frame, the toothed disc extends into the reserved opening, a servo motor three is fixed on the surface of the protective frame, and a rotating shaft of the servo motor three is fixed on the surface of the toothed disc after passing through the side plate of the protective frame.

[0014] Preferably, a mounting plate is inserted into the groove of the protective frame, and the mounting plate includes a "U"-shaped plate and two "L"-shaped plates integrally formed at both ends of the "U"-shaped plate, and the two "L"-shaped plates are respectively buckled on the two side plates of the protective frame, and a hand-tightened screw is screwed on the surface of the "L"-shaped plate, and one end of the hand-tightened screw is screwed on the side plate of the protective frame. A socket is provided on the surface of the "U"-shaped plate, and a traction slide is inserted inside the socket. The traction slide is a "T"-shaped plate, and one end of the traction slide is fixed to the surface of the oil storage box, and the width of the plate surface at the other end of the traction slide is greater than the slot width of the socket.

[0015] Preferably, an oil-absorbing cotton plate is fixed to the inner wall of the oil storage box, a connecting cotton core is fixed to the surface of the oil-absorbing cotton plate, the connecting cotton core passes through two holes and is fixed to the surface of the oil-coated cotton plate, the oil-coated cotton plate is fixed to the surface of the oil storage box, an oil filling hole is provided on the surface of the oil storage box, a sealing head is plugged into the inside of the oil filling hole, the sealing head is in the shape of a "T"-shaped cylinder, a rubber ring is fixed on one end of the sealing head extending into the interior of the oil storage box, the outer diameter of the rubber ring is larger than the aperture of the oil filling hole, and a pick-out groove is provided on the plate body at the other end of the sealing head, the pick-out groove is in the shape of an annular groove.

[0016] A method for identifying abnormal behavior in a station based on multi-scale vision, the method comprising the following steps:

[0017] Construct a dataset of abnormal behaviors in railway stations;

[0018] Extract effective features from the input video frames and enhance the model's ability to identify abnormal behaviors by fusing features of different scales;

[0019] Use the extracted features to accurately identify abnormal behaviors in video frames;

[0020] Build an alarm system for bad behavior within the site.

[0021] Preferably, the specific operations of constructing the abnormal behavior data set in the railway station include:

[0022] Perform HSV transformation on the original video frame, and further scale and crop the image to ensure the consistency of the data size of the input model, and increase the diversity of the data through random cropping;

[0023] Use CutMix fusion technology to mix parts of different images to generate new training samples, thereby enhancing the model's ability to recognize targets;

[0024] Use the labelImg tool to accurately annotate abnormal behaviors in each video frame, including behavior type and target location.

[0025] Preferably, the specific operations of extracting effective features from the input video frames and enhancing the model's ability to recognize abnormal behaviors by fusing features of different scales include:

[0026] Use pre-trained ResNet-18 networks to extract low-level features of images. These models have learned rich feature representations and are adapted to the task of identifying abnormal behavior in railway stations by fine-tuning during training.

[0027] The parameters of the pre-trained network are used as the initialization parameters of the feature extraction module, and the knowledge is transferred from the source task to the target task through transfer learning technology, which accelerates the convergence speed of the model and improves the recognition accuracy.

[0028] By extracting features at different levels of CNN and fusing them using a common attention mechanism, the shallow features and deep features are combined to enhance the model's ability to represent the target. Weights are applied to feature maps to strengthen attention to important areas and suppress interference from irrelevant areas.

[0029] PCA dimensionality reduction technology is used to reduce the dimension of high-dimensional features generated after fusion, while retaining key information;

[0030] The long short-term memory network (LSTM) is used to process the video frame sequence after feature extraction to capture the dynamic changes of behavior in the time dimension.

[0031] Preferably, the specific operation of using the extracted features to accurately identify abnormal behaviors in the video frames includes:

[0032] Adopt a multi-task learning strategy to combine abnormal behavior recognition with other tasks, share some model parameters, and optimize the loss functions of multiple tasks at the same time;

[0033] After extracting the spatiotemporal features, the fully connected layer is used as a classifier to classify different behavior categories, and the Adam optimizer is selected to optimize the recognition performance of the model;

[0034] During the training process, strategies such as cross-validation, regularization techniques, and learning rate adjustment are used to optimize the performance of the model and prevent overfitting.

[0035] Compared with the prior art, the present invention has the following beneficial effects:

[0036] The system and method for identifying abnormal behaviors in stations based on multi-scale vision proposed in the present invention can effectively extract rich feature information from video data and enhance the model's recognition ability for targets of different scales by combining a pre-trained deep learning feature extractor and multi-scale feature fusion technology. By utilizing multi-scale feature fusion and combining a multi-task learning strategy, the system can accurately identify a variety of abnormal behaviors and has a strong ability to detect minor abnormal behaviors, while reducing the false alarm rate and avoiding incorrect recognition of normal behaviors. The system and device design take into account compatibility with existing monitoring and security management systems, and the plug-in design allows it to be easily integrated into the existing railway station security architecture, facilitating rapid deployment and application. With the help of high-performance hardware configuration and optimized algorithm design, the system can realize real-time video processing and behavior recognition. This real-time performance ensures that the system can respond to and handle abnormal situations in a timely manner, providing immediate support for the safety management of railway stations. The device can detect a variety of abnormal conditions in railway stations, Such as safety hazards, illegal activities, emergencies and uncivilized behaviors, and the model parameters can be learned and updated online, and the recognition model can be continuously optimized and improved according to the newly collected data. This continuous learning and updating capability maintains the long-term effectiveness of the system and adapts to the ever-changing security needs; the system supports personalized management, allowing security managers to customize the content and level of alarms according to different abnormal behaviors and scenario requirements. For minor uncivilized behaviors, the system can send gentle reminders; for serious security threats, the system will issue an emergency alarm and immediately notify relevant personnel to deal with it. This personalized management and flexible alarm function enables the system to respond to various situations more accurately, improving the pertinence and effectiveness of security management; further solving the existing problems of low accuracy and efficiency in identifying multiple categories of abnormal behaviors in complex scenarios; it can achieve accurate identification of abnormal behaviors from large-scale to small-scale, providing strong technical support for security monitoring within railway stations. BRIEF DESCRIPTION OF THE DRAWINGS

[0037] Figure 1 is a flow chart of the method of the present invention;

[0038] Figure 2 is a system block diagram of the present invention;

[0039] Figure 3 This is a schematic diagram of the structure of the surveillance camera of the present invention after installation;

[0040] Figure 4 For the present invention Figure 3 Top view of the structure;

[0041] Figure 5 for Figure 4 Structural cross-section at AA in the middle;

[0042] Figure 6 for Figure 5 A schematic diagram of the structure enlargement in the middle;

[0043] Figure 7 for Figure 6 A magnified schematic diagram of the structure at B in the middle;

[0044] Figure 8 for Figure 4 Structural cross-section at the middle BB;

[0045] Fig. 9 for Figure 8 A magnified schematic diagram of the structure at C in the middle;

[0046] Fig.10 for Figure 8 A magnified schematic diagram of the structure at D in the middle;

[0047] Fig.11 This is a schematic diagram of the connection structure between the mounting plate and the oil storage box of the present invention;

[0048] Fig.12 This is a schematic diagram of the structure of the mounting plate of the present invention;

[0049] Fig.13 This is a schematic diagram of the structure of the oil storage box of the present invention;

[0050] Fig.14 This is a schematic diagram of the head structure of the present invention;

[0051] Fig.15 This is a schematic diagram of the connection structure of the cotton core and the oil-coated cotton board of the present invention;

[0052] Fig.16 This is a schematic diagram of the lifting plate structure of the present invention;

[0053] Fig.17 This is a schematic diagram of the connection structure between the monitoring camera and the rotating seat of the present invention;

[0054] Fig.18 This is a schematic diagram of the cleaning plate structure of the present invention;

[0055] Fig.19 It is a schematic diagram of the structure of the rotating seat of the present invention;

[0056] Fig. 20 This is a schematic diagram of the installation frame structure of the present invention;

[0057] Fig.21 This is a schematic diagram of the structure of the monitoring camera of the present invention after being tilted.

[0058] In the figure: surveillance camera 1, storage slot 2, rotating seat 3, servo motor 1 4, assembly slot 5, cleaning plate 6, perforation 1 7, plug rod 8, bristles 9, sponge wiper 10, elastic traction sheet 11, card strip 12, card slot 13, pull ring 14, loading platform 15, servo motor 2 16, missing slot 17, installation frame 18, lifting plate 19, side slot 20, connecting angle plate 21, embedded slot 22, teeth 23, toothed disc 2 4. Reserved opening 25, protective frame 26, servo motor three 27, mounting plate 28, hand screw 29, socket 30, traction slide plate 31, oil storage box 32, micro electric push rod 33, oil absorbing cotton board 34, connecting cotton core 35, perforation two 36, oiling cotton board 37, oil filling hole 38, head 39, rubber ring 40, picking groove 41, display screen one 42, display screen two 43, baffle 44, sponge wiper two 45. DETAILED DESCRIPTION

[0059] In order to make the purpose and technical solution of the present invention clearly and completely described, and the advantages more clearly understood, the embodiments of the present invention are further described in detail with reference to the accompanying drawings. It should be understood that the specific embodiments described herein are part of the embodiments of the present invention, rather than all of the embodiments, and are only used to explain the embodiments of the present invention, and are not used to limit the embodiments of the present invention. All other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.

[0060] Embodiment 1

[0061] See also Figure 2 ,The present invention provides a technical solution: a system for identifying abnormal behavior in a station based on multi-scale vision, the system is composed of an image acquisition and preprocessing module, a feature extraction and fusion module, a behavior recognition module and an alarm module;

[0062] The image acquisition and preprocessing module is responsible for collecting video data in the railway station from the monitoring camera 1 or other image acquisition devices, including a digital-to-analog converter, an image acquisition card, and a data transmission interface component, which is used to convert the real-time scene into a digital image signal for processing; the video signal of the real-time scene is captured by the monitoring camera and other devices, and these analog video signals are then sent to the digital-to-analog converter to convert them into digital image signals. The converted digital signal is further transmitted to the image acquisition card, which is responsible for processing the signal and performing preprocessing operations such as denoising and format conversion. The processed digital image signal is finally sent to the subsequent image processing or model analysis module through the data transmission interface component to identify and analyze abnormal behavior.

[0063] The feature extraction and fusion module is connected to the image acquisition and preprocessing module and the behavior recognition module, receives the collected signals, and performs feature extraction and fusion on the preprocessed video information to facilitate classification by the behavior recognition module; it is composed of an edge computing graphics processing unit, a cache unit, and a data wireless transmission unit; the feature extraction and fusion module receives the preprocessed video information via the data wireless transmission unit. This information is sent to the edge computing graphics processing unit, which is responsible for deep feature extraction of the video information, including but not limited to extracting features such as texture, shape, and motion of the image. The extracted features are then temporarily stored by the cache unit for further feature fusion processing.

[0064] The behavior recognition module is connected to the feature extraction and fusion and alarm modules, accepts multi-scale abnormal behavior features in the station, and accelerates the response to emergency situations by synchronizing the information to other relevant managers in the station; it consists of a central graphics processing unit, a cache unit, a data wireless transmission unit, and a data transmission interface; the abnormal behavior information identified by the central graphics processing unit is then temporarily stored by the cache unit and synchronized to the equipment of relevant managers in the station through the data wireless transmission unit, accelerating the response to emergency situations. In addition, the data transmission interface component is responsible for sending the recognition results to the alarm module so that an alarm can be issued in time when abnormal behavior is detected to ensure safety in the station.

[0065] The alarm module is connected to the behavior recognition module, stores the collected abnormal situation images and abnormal situation categories, and issues a sound alarm; it is composed of a control chip, a storage controller, a data wireless transmission unit, an alarm, and an interface circuit; the received and identified abnormal situation information is sent to the control chip, which is responsible for analyzing and processing the received abnormal information. The processed abnormal information is then saved by the storage controller for subsequent data analysis and record query. At the same time, the control chip will instruct the alarm to issue a sound alarm to remind the station personnel to take timely countermeasures. The data wireless transmission unit is responsible for synchronizing the abnormal situation information to other related equipment or management personnel in the station to ensure the rapid transmission and response of information. The interface circuit component is responsible for connecting and coordinating the work of various internal components to ensure the normal operation and efficient response of the alarm module.

[0066] Embodiment 2

[0067] See attached Figures 3 to 21 On the basis of the first embodiment, in order to realize that the monitoring camera 1 satisfies the traditional monitoring function, the position of the monitoring camera 1 is adjusted to meet the monitoring effect of the monitoring camera 1 as a drone perspective; the monitoring camera 1 is rotatably connected to the rotating seat 3, the rotating seat 3 is rotatably connected above the stage 15, and a cleaning plate 6 is slidably provided inside the rotating seat 3. The stage 15 is fixed on the surface of the lifting plate 19, and the lifting plate 19 is movably inserted in the installation frame 18. One side of the lifting plate 19 is provided with teeth 23, and the teeth 23 are meshed with a toothed disc 24. The toothed disc 24 penetrates the side plate of the installation frame 18, and the toothed disc 24 is rotatably connected to the protective frame 26. One end of the protective frame 26 is movably inserted with an oil storage box 32, and the interior of the oil storage box 32 is filled with lubricating oil. The surface of the oil storage box 32 is provided with a second perforation 36, and the lubricating oil seeps out from the second perforation 36 and is smeared on the toothed disc 24. A display screen 1 42 is fixed on the surface of the installation frame 18, and a display screen 2 43 is fixed on the surface of the lifting plate 19;

[0068] The mounting frame 18 is mounted on one side of the ticket checking platform, that is, after the mounting frame 18 is placed against one side of the ticket checking platform, the connecting angle plate 21 is fixed to the ticket checking platform by bolts, and at this time, the monitoring camera 1 is suspended on the ticket checking turntable for monitoring and shooting the surrounding area of ​​the ticket checking platform; the display screen 1 42 is used to display the information of the train being checked, or the display screen 2 43 plays the warning information of the station in a loop, and the display screen 2 43 is used to display the name information of the lost persons in the station, and the latter displays the train number information of the ticket checking platform being changed; after the lifting plate 19 is driven to be lifted by the gear plate 24, the position of the display screen 2 43 is raised, and at this time, the information displayed on the display screen 2 43 is easily viewed by people at a long distance, and at this time, the height of the monitoring camera 1 is raised, and the lens of the monitoring camera 1 can be adjusted horizontally and vertically, thereby expanding the monitoring angle of the monitoring camera 1, and the height of the monitoring camera 1 is large, so it can be used as a drone monitoring angle, which is conducive to capturing image data from more angles.

[0069] A storage slot 2 is provided on the top surface of the rotating seat 3, and connecting shafts are fixed on both sides of the monitoring camera 1. The two connecting shafts respectively penetrate the two sides of the storage slot 2. A servo motor 4 is fixed on the surface of the rotating seat 3, and the driving shaft of the servo motor 4 is fixed on the surface of a connecting shaft. An assembly slot 5 is provided on the bottom surface of the storage slot 2, and a plug rod 8 is fixed on the surface of the assembly slot 5. The plug rod 8 is inserted in the through hole 7, and the through hole 7 is provided on the surface of the cleaning plate 6. The cleaning plate 6 has a "T"-shaped plate structure. Brush bristles 9 and a sponge wiper 10 are fixed on the top surface of the cleaning plate 6. Two elastic traction pieces 11 are fixed to the end, and the two elastic traction pieces 11 are slidably connected in the assembly groove 5. A clamping strip 12 is fixed on the surface of the elastic traction piece 11, and the clamping strip 12 is inserted in the clamping slot 13. The clamping slot 13 is arranged on the side wall of the assembly groove 5, and the height of the clamping slot 13 is less than the height of the assembly groove 5. A pull ring 14 is fixed on the surface of the elastic traction piece 11, and the pull ring 14 and the clamping strip 12 are symmetrically distributed about the elastic traction piece 11; a servo motor 16 is fixed to the bottom surface of the stage 15, and the rotating shaft of the servo motor 16 passes through the stage 15 and is fixed to the bottom surface of the rotating seat 3;

[0070] After the switch of the servo motor 16 is triggered, the shaft of the servo motor 16 rotates and drives the rotating seat 3 to rotate, so that the monitoring camera 1 installed on the rotating seat 3 can be adjusted 360 degrees in the horizontal direction. After the switch of the servo motor 14 is triggered, the monitoring camera 1 is driven to swing in the vertical direction to adjust the vertical monitoring direction of the monitoring camera 1, that is, to expand the monitoring angle of the monitoring camera 1. When the lens surface of the monitoring camera 1 is blocked by foreign matter, the switch of the servo motor 14 is triggered to drive the monitoring camera 1 to swing downward. In this process, the bristles 9 and the sponge wiper 1 0 Sweep away foreign matter on the surface of the lens of the surveillance camera 1, and there is no need for the staff to climb to a high place to clean the lens of the surveillance camera 1; when the cleaning plate 6 needs to be replaced or cleaned, two fingers hold the two pull rings 14 respectively. After the two fingers are closed, the pull rings 14 pull the corresponding elastic traction sheet 11 to produce elastic deformation, and the elastic traction sheet 11 pulls the card strip 12 to be pulled out of the card slot 13. The width of the card strip 12 is smaller than the width of the card slot 13, so that the card strip 12 can be pulled out of the card slot 13 when the elastic traction sheet 11 is deformed. At this time, there is no fixation between the cleaning plate 6 and the assembling slot 5, and the cleaning plate 6 can be pulled out of the assembling slot 5.

[0071] A mounting groove 22 is provided on one side of the lifting plate 19, the height of the mounting groove 22 is less than the height of the lifting plate 19, a plurality of teeth 23 are provided, and the plurality of teeth 23 are arranged at equal distances and in equal sizes and fixed inside the mounting groove 22, the mounting frame 18 is a square tube structure, a reserved opening 25 is provided on the surface of the mounting frame 18, a toothed disc 24 extends into the reserved opening 25, a servo motor 3 27 is fixed on the surface of the protective frame 26, a rotating shaft of the servo motor 3 27 passes through the side plate of the protective frame 26 and is fixed on the surface of the toothed disc 24; a mounting plate 28 is inserted into the notch of the protective frame 26, and the mounting plate 28 includes A "U"-shaped plate and two "L"-shaped plates integrally formed at both ends of the "U"-shaped plate, the two "L"-shaped plates are respectively buckled on the two side plates of the protective frame 26, the surface of the "L"-shaped plate is screwed with a hand screw 29, one end of the hand screw 29 is screwed on the side plate of the protective frame 26, the surface of the "U"-shaped plate is provided with a socket 30, the inside of the socket 30 is plugged with a traction slide 31, the traction slide 31 is a "T"-shaped plate, one end of the traction slide 31 is fixed to the surface of the oil storage box 32, and the other end of the traction slide 31 has a plate width greater than the slot width of the socket 30; The inner wall of the box 32 is fixed with an oil-absorbing cotton plate 34, and a connecting cotton core 35 is fixed on the surface of the oil-absorbing cotton plate 34. The connecting cotton core 35 passes through the second perforation 36 and is fixed on the surface of the oil-applied cotton plate 37. The oil-applied cotton plate 37 is fixed on the surface of the oil storage box 32. An oil filling hole 38 is provided on the surface of the oil storage box 32. A sealing head 39 is inserted into the inside of the oil filling hole 38. The sealing head 39 is a "T"-shaped cylinder. A rubber ring 40 is fixed on one end of the sealing head 39 extending into the inside of the oil storage box 32. The outer diameter of the rubber ring 40 is larger than the aperture of the oil filling hole 38. A pick-up hole 38 is provided on the plate body of the other end of the sealing head 39. Groove 41, the picking groove 41 is an annular groove, the stage 15 is slidably connected in the notch 17, the notch 17 is opened on the top surface of the installation frame 18, a baffle 44 is fixed between two parallel side walls of the notch 17, a sponge wiper plate 2 45 is fixed on the surface of the baffle 44, after the lifting plate 19 is retracted into the installation frame 18, the sponge wiper plate 2 45 is clamped between the display screen 2 43 and the baffle 44 to generate elastic deformation, the height of the sponge wiper plate 2 45 is greater than the height of the display screen 1 42, a side groove 20 is opened on the surface of the installation frame 18, and a plurality of connecting angle plates 21 are fixed on the surface of the side groove 20;

[0072] After the switch of the servo motor 3 27 is triggered, the servo motor 3 27 drives the toothed disc 24 to rotate. After the toothed disc 24 engages with the teeth 23, it drives the lifting plate 19 to slide along the mounting frame 18, thereby adjusting the position of the monitoring camera 1. When the lifting plate 19 is retracted into the mounting frame 18, the height of the monitoring camera 1 is more suitable for shooting the front of the ticket inspection platform, and the display screen 2 43 is retracted into the mounting frame 18 at this time. When the display screen 2 43 needs to be used, the lifting plate 19 is driven to move upward by the toothed disc 24. In this process, the sponge wiper 2 45 wipes off the dirt on the surface of the display screen 2 43. It should be noted that the servo motor 1 4, the servo motor 2 16 and the servo motor 3 27 are all programmable bidirectional drive motors.

[0073] When lubrication between the toothed disc 24 and the teeth 23 is required, the switch of the micro-electric push rod 33 is triggered, and the micro-electric push rod 33 pushes the oil storage box 32 to move toward the toothed disc 24. After the oil-absorbing cotton plate 34 absorbs the lubricating oil inside the oil storage box 32 to be saturated, the lubricating oil penetrates the oil-coated cotton plate 37 through the connecting cotton core 35. The oil-coated cotton plate 37 applies the lubricating oil to the surface of the toothed disc 24, and then the oil storage box 32 is pulled and reset, and the toothed disc 24 is driven to rotate again. The toothed disc 24 transfers the lubricating oil to the surface of the teeth 23. This process is repeated many times to achieve lubrication of the toothed disc 24 and the teeth 23. When replenishing the lubricating oil inside the oil storage box 32, the finger is pulled into the extraction groove 41 to pull the head 39 outward. After the head 39 is removed from the oil filling hole 38, the lubricating oil is injected into the oil storage box 32 through the oil filling hole 38.

[0074] Embodiment 3

[0075] See attached Figure 1 Based on the second embodiment, a method for identifying abnormal behavior in a station based on multi-scale vision is proposed, and the method includes the following steps:

[0076] S1: Constructing a dataset of abnormal behaviors in railway stations

[0077] S1.1: Perform HSV transformation on the original video frame, and further scale and crop the image to ensure the consistency of the data size of the input model, and increase the diversity of the data through random cropping. Specifically, obtain the color components R, G, B of the original image, and calculate the maximum value of the color component Max and minimum value Min , calculate the hue H, saturation S and brightness V according to formulas 1), 2) and 3):

[0078] 1)1)

[0079] 2)

[0080] 3)

[0081] S1.2: Use CutMix fusion technology to mix parts of different images to generate new training samples, thereby enhancing the model's ability to recognize targets. Randomly select a rectangular area from the , let the coordinates of this rectangular area be , the width is , the height is . Select image Cut out the corresponding rectangular area in the image and paste it into the image Generate a new image at the corresponding position of Finally, the labels of the two images are mixed according to the area ratio of the rectangular region. Let the area ratio of the rectangular region be ,in and are the width and height of the image respectively. Then, the new image Tags It can be expressed as:

[0082]

[0083] S1.3: Use the labelImg tool to accurately label abnormal behaviors in each video frame, including behavior type and target location.

[0084] S2: Feature extraction and fusion module. The feature extraction and fusion module is the core part of the multi-scale railway station abnormal behavior recognition system. Its main purpose is to extract effective features from the input video frames and enhance the model's recognition ability of abnormal behavior by fusing features of different scales.

[0085] S2.1: Use the pre-trained ResNet-18 network to extract low-level features of the image. These models have learned rich feature representations and are adapted to the task of abnormal behavior recognition in railway stations by fine-tuning during training. Suppose the input image is , the pre-trained ResNet-18 can be expressed as a function , the output is the feature , then the whole process can be expressed as: ,in θ Represents the parameter combination of the model.

[0086] S2.2: Use the parameters of the pre-trained network as the initialization parameters of the feature extraction module, and transfer knowledge from the source task to the target task through transfer learning technology. This can accelerate the convergence speed of the model and improve the recognition accuracy.

[0087] S2.3: By extracting features at different levels of CNN and fusing these features using a common attention mechanism, the shallow features and deep features are combined to enhance the model's ability to represent the target. By applying weights to the feature map, attention can be focused on important areas and interference from irrelevant areas can be suppressed. Assume there are two types of input information and , use vector dot product to calculate their similarity , and then calculate the attention weight and :

[0088]

[0089] Finally, the collaborative representation of the two inputs is obtained by weighted summation :

[0090]

[0091] S2.4: PCA dimensionality reduction technology is used to reduce the dimension of the high-dimensional features generated after fusion while retaining key information.

[0092] S2.5: Since video data has temporal continuity, a long short-term memory network (LSTM) is used to process the video frame sequence after feature extraction to capture the dynamic changes of behavior in the temporal dimension.

[0093] S3: The purpose of the behavior recognition module is to use the extracted features to accurately identify abnormal behaviors in video frames.

[0094] S3.1: Adopt the strategy of multi-task learning to combine abnormal behavior recognition with other tasks (such as person detection, pedestrian re-identification, posture estimation, etc.), share some parameters of the model, and optimize the loss functions of multiple tasks at the same time, so that other auxiliary information can also be obtained to improve the accuracy and robustness of recognition. In multi-task recognition, suppose there is K subtasks, and the loss function corresponding to each subtask is L k , the weight is w k , then the total loss function L total for:

[0095]

[0096] S3.2: After extracting the spatiotemporal features, use the fully connected layer as a classifier to classify different behavior categories. And select the Adam optimizer to optimize the recognition performance of the model. During training, the optimization goal is to minimize the total loss function, that is,

[0097]

[0098] S3.3: During the training process, cross-validation, regularization techniques, learning rate adjustment and other strategies are used to optimize the performance of the model and prevent overfitting.

[0099] S4: Construct the bad behavior alarm device in the station, which mainly includes image acquisition and preprocessing module, feature extraction and fusion module, behavior recognition module and alarm module. The connection method of each component is as follows: Figure 2 shown.

[0100] Although embodiments of the present invention have been shown and described, it will be appreciated by those skilled in the art that various changes, modifications, substitutions and variations may be made to the embodiments without departing from the principles and spirit of the present invention, and that the scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. A multi-scale vision-based abnormal behavior recognition system within a station, characterized by: The system consists of an image acquisition and preprocessing module, a feature extraction and fusion module, a behavior recognition module and an alarm module; The image acquisition and preprocessing module is responsible for acquiring video data in a railway station from a monitoring camera (1) or other image acquisition equipment, and comprises a digital-to-analog converter, an image acquisition card, and a data transmission interface component, and is used to convert a real-time scene into a digital image signal for processing; the monitoring camera (1) is rotatably connected to a rotating seat (3), the rotating seat (3) is rotatably connected to the top of a loading platform (15), a cleaning plate (6) is slidably provided inside the rotating seat (3), the loading platform (15) is fixed to the surface of a lifting plate (19), the lifting plate (19) is movably plugged into a mounting frame (18), and the lifting plate (19) is rotatably connected to a mounting frame (18). A tooth (23) is provided on one side of the mounting frame (19), the tooth (23) and the toothed disc (24) are meshed, the toothed disc (24) penetrates the side plate of the mounting frame (18), and the toothed disc (24) is rotatably connected to the protective frame (26), an oil storage box (32) is movably inserted at one end of the protective frame (26), the interior of the oil storage box (32) is filled with lubricating oil, a second perforation (36) is provided on the surface of the oil storage box (32), the lubricating oil seeps out from the second perforation (36) and is smeared on the toothed disc (24), a display screen (42) is fixed on the surface of the mounting frame (18), and a second display screen (43) is fixed on the surface of the lifting plate (19); The feature extraction and fusion module is connected to the image acquisition and preprocessing module and the behavior recognition module, receives the collected signals, extracts and fuses the features of the preprocessed video information, and facilitates the classification of the behavior recognition module; it is composed of an edge computing graphics processing unit, a cache unit, and a data wireless transmission unit; The behavior recognition module is connected with the feature extraction and fusion and alarm modules, receives multi-scale abnormal behavior features within the station, and accelerates the response to emergency situations by synchronizing the information to relevant managers in other stations; it consists of a central graphics processing unit, a cache unit, a data wireless transmission unit, and a data transmission interface; The alarm module is connected to the behavior recognition module, stores the collected abnormal situation images and abnormal situation categories, and issues a sound alarm; It is composed of a control chip, a storage controller, a data wireless transmission unit, an alarm and an interface circuit; Specific implementation methods include: Construct a dataset of abnormal behaviors in railway stations; including: Perform HSV transformation on the original video frame, and further scale and crop the image to ensure the consistency of the data size of the input model, and increase the diversity of the data through random cropping; Use CutMix fusion technology to mix parts of different images to generate new training samples, thereby enhancing the model's ability to recognize targets; Use the labelImg tool to accurately label abnormal behaviors in each video frame, including behavior type and target location; Extract effective features from the input video frames and enhance the model's ability to identify abnormal behaviors by fusing features of different scales; including: Use the pre-trained ResNet-18 network to extract low-level features of the image. The model has learned rich feature representations and is fine-tuned during training to adapt to the task of identifying abnormal behavior in railway stations. The parameters of the pre-trained network are used as the initialization parameters of the feature extraction module, and the knowledge is transferred from the source task to the target task through transfer learning technology, which accelerates the convergence speed of the model and improves the recognition accuracy. By extracting features at different levels of CNN and fusing them using a common attention mechanism, the shallow features and deep features are combined to enhance the model's ability to represent the target. Weights are applied to feature maps to strengthen attention to important areas and suppress interference from irrelevant areas. PCA dimensionality reduction technology is used to reduce the dimension of high-dimensional features generated after fusion, while retaining key information; Use the long short-term memory network (LSTM) to process the video frame sequence after feature extraction to capture the dynamic changes of behavior in the time dimension; Use extracted features to accurately identify abnormal behaviors in video frames; including: Adopt a multi-task learning strategy to combine abnormal behavior recognition with other tasks, share some model parameters, and optimize the loss functions of multiple tasks at the same time; After extracting the spatiotemporal features, the fully connected layer is used as a classifier to classify different behavior categories, and the Adam optimizer is selected to optimize the recognition performance of the model; During the training process, cross-validation, regularization techniques, and learning rate adjustment strategies are used to optimize the performance of the model and prevent overfitting. Build an alarm system for bad behavior within the site.

2. The system for identifying abnormal behavior in a station based on multi-scale vision according to claim 1, characterized in that: The top surface of the rotating seat (3) is provided with a storage groove (2), and connecting shafts are fixed on both sides of the monitoring camera (1), and the two connecting shafts respectively penetrate the two sides of the storage groove (2). A servo motor (4) is fixed on the surface of the rotating seat (3), and the driving shaft of the servo motor (4) is fixed on the surface of a connecting shaft. The bottom surface of the storage groove (2) is provided with an assembly groove (5), and a plug rod (8) is fixed on the surface of the assembly groove (5), and the plug rod (8) is inserted into a through hole (7). The through hole (7) is provided on the surface of the cleaning plate (6), and the cleaning plate (6) is a "T"-shaped plate structure. The top surface is fixed with bristles (9) and a sponge wiping plate (10); one end of the cleaning plate (6) is fixed with two elastic traction plates (11); the two elastic traction plates (11) are both slidably connected in the assembly groove (5); a clamping strip (12) is fixed on the surface of the elastic traction plate (11); the clamping strip (12) is inserted in the clamping slot (13); the clamping slot (13) is arranged on the side wall of the assembly groove (5); the height of the clamping slot (13) is less than the height of the assembly groove (5); a pull ring (14) is fixed on the surface of the elastic traction plate (11); the pull ring (14) and the clamping strip (12) are symmetrically distributed about the elastic traction plate (11).

3. The system for identifying abnormal behavior in a station based on multi-scale vision according to claim 1, characterized in that: A servo motor 2 (16) is fixed to the bottom surface of the loading platform (15). The rotating shaft of the servo motor 2 (16) passes through the loading platform (15) and is fixed to the bottom surface of the rotating seat (3). The loading platform (15) is slidably connected in the notch (17). The notch (17) is provided on the top surface of the mounting frame (18). A baffle (44) is fixed between two parallel side walls of the notch (17). A sponge wiper plate 2 (45) is fixed to the surface of the baffle (44). After the lifting plate (19) is retracted into the mounting frame (18), the sponge wiper plate 2 (45) is clamped between the display screen 2 (43) and the baffle (44) to generate elastic deformation. The height of the sponge wiper plate 2 (45) is greater than the height of the display screen 1 (42). A side groove (20) is provided on the surface of the mounting frame (18). A plurality of connecting angle plates (21) are fixed to the surface of the side groove (20).

4. The system for identifying abnormal behavior in a station based on multi-scale vision according to claim 1, characterized in that: A mounting groove (22) is provided on one side of the lifting plate (19), the height of the mounting groove (22) is less than the height of the lifting plate (19), a plurality of teeth (23) are provided, and the plurality of teeth (23) are arranged at equal distances and in equal sizes and fixed inside the mounting groove (22), the mounting frame (18) is in a square tube structure, a reserved opening (25) is provided on the surface of the mounting frame (18), the toothed disc (24) extends into the reserved opening (25), a servo motor three (27) is fixed on the surface of the protective frame (26), and a rotating shaft of the servo motor three (27) passes through a side plate of the protective frame (26) and is fixed on the surface of the toothed disc (24).

5. The system for identifying abnormal behavior in a station based on multi-scale vision according to claim 1, characterized in that: A mounting plate (28) is inserted into the notch of the protection frame (26). The mounting plate (28) comprises a "U"-shaped plate and two "L"-shaped plates integrally formed at both ends of the "U"-shaped plate. The two "L"-shaped plates are respectively buckled on two side plates of the protection frame (26). A hand screw (29) is screwed on the surface of the "L"-shaped plate. One end of the hand screw (29) is screwed on the side plate of the protection frame (26). A socket (30) is provided on the surface of the "U"-shaped plate. A traction slide plate (31) is inserted into the interior of the socket (30). The traction slide plate (31) is a "T"-shaped plate. One end of the traction slide plate (31) is fixed to the surface of the oil storage box (32). The width of the plate surface of the other end of the traction slide plate (31) is greater than the notch width of the socket (30).

6. The system for identifying abnormal behavior in a station based on multi-scale vision according to claim 1, characterized in that: An oil-absorbing cotton plate (34) is fixed to the inner wall of the oil storage box (32), a connecting cotton core (35) is fixed to the surface of the oil-absorbing cotton plate (34), the connecting cotton core (35) passes through the second perforation (36) and is fixed to the surface of the oil-applying cotton plate (37), the oil-applying cotton plate (37) is fixed to the surface of the oil storage box (32), an oil filling hole (38) is provided on the surface of the oil storage box (32), a sealing head (39) is inserted into the inside of the oil filling hole (38), the sealing head (39) is in the shape of a "T"-shaped cylinder, a rubber ring (40) is fixedly sleeved on one end of the sealing head (39) extending into the inside of the oil storage box (32), the outer diameter of the rubber ring (40) is larger than the diameter of the oil filling hole (38), and a pick-out groove (41) is provided on the plate body at the other end of the sealing head (39), the pick-out groove (41) is in the shape of an annular groove.

Images