CN211630273U - Intelligent image recognition device for railway environment - Google Patents

Abstract

The utility model provides an image intelligent recognition device for railway environment, including front end equipment, supervisory equipment and customer premises equipment. The monitoring equipment is divided into three levels from low to high, and comprises an I-type video access point, a region node and a core node, wherein each level can call lower-level image resources and has an independent image processing function. The existing railway video monitoring system is relied on, the hardware of the server is reconstructed, the existing defects of the railway video monitoring system are made up, and the automatic alarm of the intrusion behavior is realized.

Description

Intelligent image recognition device for railway environment

Technical Field

The utility model relates to a railway environment field specifically is an image intelligent recognition device for railway environment.

Background

In recent years, with the development and maturity of video monitoring technology, video monitoring has become an important safety protection means, and is widely applied to the field of railway basic safety protection, a railway comprehensive video monitoring system is built on a plurality of lines, and related personnel can clearly know the conditions of districts and all-line stations, intervals, bridges, roadbeds, railways, sand disasters, snow disasters, flood disasters and the like through display screen walls or computers in working sections and rapidly handle emergency situations. However, these video surveillance systems basically require a function of monitoring by a 24-hour professional. Because the working intensity of real-time monitoring is very high, workers may lose sensitivity to dangerous events due to fatigue and the like, and the situation of careless omission is caused. According to the statistics of the Federal railway administration, in the occurrence of railway traffic accidents every year, accidents caused by intrusion behaviors account for 72.3% of all accidents, and according to the statistics of the Shanghai railway administration, the proportion of the railway traffic accidents caused by the intrusion behaviors of vehicles and pedestrians in the total number of the accidents is higher, and the pedestrians walk on the rails and cross the railway accidents caused by the intrusion behaviors of the railways and the like, so that serious direct economic loss is caused.

The great increase of the train speed greatly shortens the avoiding time of pedestrians and animals which are required to illegally enter the railway boundary, and the train running at high speed has larger kinetic energy and inertia force. In case of accidental collision, the damage to the train and the crashed object is more serious.

Because the working strength of real-time video monitoring is high, workers may lose sensitivity to dangerous events due to fatigue and the like, and the condition of omission is caused, so that a large number of video source signals can be stored only in an automatic video recording mode in many times, and the video monitoring is a passive tool by means of a mode of checking afterwards. The utility model discloses compensate the existing not enough of railway video monitor system, can not need under the circumstances that the people participated in, can be through the processing to the sequence image by the computer, realize the autoalarm to the invasion action.

SUMMERY OF THE UTILITY MODEL

The utility model discloses a solve prior art's problem, provide an image intelligent recognition device for railway environment, rely on current railway video monitored control system, carry out the framework again to the hardware of server, compensate the existing not enough of railway video monitored control system, realize the autoalarm to the invasion action.

The utility model discloses mainly include front end equipment, supervisory equipment and customer premises equipment.

The front-end camera is arranged in key areas such as an interval tunnel portal, a highway iron overpass, a bridge evacuation channel, equipment rooms along the line, a dispatching office portal, a station throat area and the like, and is monitored day and night by adopting a high-definition infrared LED camera and a long-focus laser camera.

The whole intelligent monitoring system is divided into three levels from low to high, each level can call lower-level image resources, and the intelligent monitoring system has an independent image processing function.

The video processing server is deployed in each type I video access point (station segment) machine room, and the video processing server and corresponding video hardware are as same as the machine room as possible, so that the communication transmission workload is reduced. The method is mainly used for receiving image information captured by a front-end camera, and performing video analysis, alarm processing and image storage.

The alarm analysis server is deployed in each area node (road bureau/passenger special scheduling department), and is mainly used for receiving alarm information reported by the I-type video access point, distinguishing alarm levels, judging whether the core nodes need to be reported or not, storing all alarm image resources, calling all image resources of the I-type video access point, and performing manual intervention if necessary.

The data analysis server is deployed in a core node (railway department), and is mainly used for receiving alarm information reported by regional nodes, storing important alarm image resources, calling the alarm image resources, and performing data analysis and processing on the alarm information.

And the user side equipment receives and processes the alarm signal through the data analysis server of the core node.

The utility model has the advantages that:

1. the system is integrated, and the management is convenient:

the system supports direct access of various hardware devices and software systems, provides communication interfaces and communication protocols, and managers can check data of various subsystems on one platform, so that the interfaces of all the systems are uniform in style, graphical interfaces consistent with the site are adopted, and management difficulty is greatly reduced.

2. The alarm function is powerful:

the system provides a plurality of information notification means such as telephone, short message, field voice, picture pop-up, information release, broadcast and the like for the alarm information, and various alarm information can set an alarm scheduling list, a safety time period, a repeated alarm period (the alarm is not eliminated, the alarm is set to be sent repeatedly at a specified interval), an alarm upgrading period (the important alarm is not eliminated, and the alarm is automatically reported to the leader of the previous stage after a period of time).

3. The management mode is more various:

management can be remotely managed on a terminal through client software, and different login users have strict authority division. When an alarm occurs on a certain road section, the system can automatically pop up an electronic map of the alarm position, and simultaneously, the video and asset information related to the alarm point and the processing plan are displayed in a combined mode. When a plurality of alarms occur, the alarm pictures are sequentially and circularly jumped, and the device can move forward and backward in a single step. The system provides a history alarm inquiring function, can automatically retrieve the video in the alarm time period, and can graphically position the alarm occurring position.

Drawings

Fig. 1 is a schematic view of the topology structure of the present invention.

Fig. 2 is a schematic diagram of the implementation steps of the present invention.

Fig. 3 is a schematic diagram of a device connection scheme.

Detailed Description

The present invention will be further explained with reference to the accompanying drawings.

The utility model discloses topological structure is shown in fig. 1, mainly includes front end equipment, supervisory equipment and customer premises equipment.

The front-end camera is arranged in key areas such as an interval tunnel portal, a highway iron overpass, a bridge evacuation channel, equipment rooms along the line, a dispatching office portal, a station throat area and the like, and is monitored day and night by adopting a high-definition infrared LED camera and a long-focus laser camera.

The whole intelligent monitoring system is divided into three levels from low to high, each level can call lower-level image resources, and the intelligent monitoring system has an independent image processing function.

The video processing server is deployed in each type I video access point (station segment) machine room, and the video processing server and corresponding video hardware are as same as the machine room as possible, so that the communication transmission workload is reduced. The method is mainly used for receiving image information captured by a front-end camera, and performing video analysis, alarm processing and image storage.

The alarm analysis server is deployed in each area node (road bureau/passenger special scheduling department), and is mainly used for receiving alarm information reported by the I-type video access point, distinguishing alarm levels, judging whether the core nodes need to be reported or not, storing all alarm image resources, calling all image resources of the I-type video access point, and performing manual intervention if necessary.

The data analysis server is deployed in a core node (railway department), and is mainly used for receiving alarm information reported by regional nodes, storing important alarm image resources, calling the alarm image resources, and performing data analysis and processing on the alarm information.

And the user side equipment receives and processes the alarm signal through the data analysis server of the core node.

The utility model discloses a system be current system, the utility model discloses main improvement point lies in making up the distribution to its structure, realizes better equipment monitoring function through new level connected mode, adopts a concrete current system to explain its monitoring mode below.

The method comprises the steps of firstly preprocessing an original video to ensure normal work under the conditions of haze, overcast and rainy and evening, and specifically, algorithms such as demisting and Gaussian filtering can be used. The robustness in multiple scenes is also improved by preprocessing the original image.

For target detection, a method of training a deep network is adopted, a network structure adopts Densenet, and a training data set provided by a railway side is adopted for adjustment after the network is trained by utilizing data on ImageNet. The adjusted network can distinguish the invading human body in various postures. Or training by adopting other forms of networks such as R-FCN, Faster R-CNN and the like, and improving the accuracy of target detection by using the output results of a plurality of neural networks by using a boosting method. The parameters are adjusted through visualization of the neural network, and the neural network obtains more information through experience of people.

For target tracking, a DSOD detection algorithm which does not need pre-training is adopted, a backbone sub-network is used for feature extraction, and targets are detected and fused on feature maps of 6 scales. And predicting the target by utilizing the front-end sub-network, predicting the position of the target according to the motion vector of the object, and reducing the complexity of the algorithm. The algorithm can track the same object between frames, identify the intrusion of the object in a defined area and judge the intrusion direction and speed of the object. And the multi-camera combined research and judgment is carried out by a characteristic matching method.

The method comprises the steps of utilizing a three-dimensional environment reconstruction technology, such as a SLAM technology of robot positioning and the like, to automatically reproduce environment information in a long term, and then predicting and judging the behavior of an invasive object and early warning in advance by combining climate, temperature and human posture information provided by the Internet of things and a communication network.

The main functions are as follows:

(1) analysis rule set-up

In terms of intelligent analysis settings, the functions that an analysis system should possess include:

supporting to define a fortification area in a monitoring picture;

supporting the setting of the adopted analysis algorithm and parameters;

and different intrusion analysis functions for the fortification area are supported.

(2) Analysis function

The following analysis functions are provided:

and (3) intrusion detection: for an object (such as a person, an animal, a vehicle, a foreign object, and the like) in the video monitoring screen, it is determined whether the object appears in a predetermined area, and if so, the object is regarded as an intruding object and marked on the screen.

Train detection: and judging whether one (or more) trains exist in the objects in the video monitoring picture, and if so, labeling the objects in the picture.

Pedestrian detection: and judging whether one (or more) pedestrians exist in the objects in the video monitoring picture, and if so, labeling the objects in the picture.

And (3) line crossing detection: for an object (such as a person, an animal, a vehicle, or the like) in a video monitoring screen, it is determined whether the object crosses a pre-defined reticle along a predetermined direction, and if so, the object is regarded as an over-line object and marked on the screen.

Detecting a moving object: and judging whether the position of an object (such as a person, an animal, a vehicle and the like) in the video monitoring picture changes in different frame pictures, if so, taking the object as a moving target, and labeling the object in the picture.

And (3) target classification: marking the object which is taken as the target in the video monitoring picture, and judging the type (such as people, vehicles and foreign matters) of the object.

Jointly studying and judging multiple cameras: when a plurality of cameras exist simultaneously, the same intrusion behavior body is marked in a single picture.

(4) Alarm management

The alarm management function that should possess includes:

and (3) alarm filtering: the warning generated by the interference of train passing through the line, night light and the like is filtered.

And (4) alarm content storage: recording and storing of event images, event videos, event levels, event times, corresponding camera numbers and the like in the system should be supported.

And (4) alarm processing: supporting a background to manually confirm the event information; automatic pop-up of event information and related images should be supported; the method can support prompting by adopting modes of voice, prompting frame, icon color change and the like aiming at different monitoring logs.

Main technical index requirements

In the following performance requirements, the detection rate refers to the ratio of the positive report number of the tested video analysis system to the total number of the test samples; the false detection rate refers to the ratio of the number of false alarms of the tested video analysis system to the total number of tested samples.

(1) Intrusion detection

Finding an intrusion target of not less than 16 x 16 pixels;

the low-speed invasion of not less than 10 pixels/s can be detected;

the output delay time of the event signal should be less than 2 s;

detection rate: > 90%;

the false detection rate is as follows: < 10%.

(2) Train detection

Train targets can be found;

the output delay time of the event signal should be less than 2 s;

detection rate: > 90%;

the false detection rate is as follows: < 10%.

(3) Pedestrian detection

Pedestrian objects with heads not smaller than 16 x 16 pixels can be identified;

the output delay time of the event signal should be less than 2 s;

detection rate: > 80%;

the false detection rate is as follows: < 20%.

(4) Line crossing detection

Objects not smaller than 16 × 16 pixels can be found for detection;

the output delay time of the event signal should be less than 2 s;

detection rate: > 80%;

the false detection rate is as follows: < 20%.

(5) Moving object detection

A moving object not smaller than 16 × 16 pixels can be found;

low-speed motion not less than 10 pixels/s can be detected;

high-speed motion not higher than 200 pixels/s can be detected;

the output delay time of the alarm signal is less than 2 s;

detection rate: > 90%;

the false detection rate is as follows: < 20%.

System embodiments

1. Carrying out the step

The implementation of the system comprises five steps: the method comprises the steps of front-end camera installation position determination, intelligent video analysis system hardware installation, intelligent video analysis system platform installation, field detection and debugging and project acceptance inspection. The implementation steps are shown in FIG. 2:

determining the installation position of the camera: and surveying the site, and determining the camera arrangement position and the wiring mode to the monitoring center.

Installing intelligent video analysis system hardware: the monitoring center is built, the camera is erected, and the intelligent video analysis server is installed.

Installing intelligent video analysis system software: and an intelligent video analysis engine and a video analysis comprehensive platform are installed and set up.

Field detection and debugging: performing related debugging work on each device, including video channel detection and debugging, and intelligent video analysis of load detection and debugging of each system; and (4) performing test operation on all the subsystem functions by combining the field situation.

And (4) checking and accepting engineering: after the debugging is passed, entering a project acceptance link, and after acceptance is qualified, delivering for use; when the inspection and acceptance is unqualified, the design and construction units should repair the steel plate until the steel plate passes the inspection and acceptance.

2. Deployment hierarchy

By establishing a video monitoring and intelligent video analysis system in regions of interest such as an interval tunnel portal, a highway-span railway overpass, a bridge evacuation channel, equipment rooms along the line, a dispatching office portal, a station throat area and the like, data and results are analyzed with the help of an intelligent video analysis technology, real-time early warning and alarming functions are realized, the condition along the line of the railway can be mastered in real time through management, resources are reasonably adjusted and distributed, emergency measures are taken in time, danger is avoided, and the safety of the railway is further improved.

Meanwhile, each monitoring point respectively counts intelligent analysis results of the railways, and the intelligent analysis results are uniformly sent to the management platform, and the management platform classifies and processes the data to form big data. On the basis, applications aiming at different industries and users are developed, including but not limited to industries such as transportation, public security, tourism and the like, and related service providers and the like.

The whole intelligent monitoring system can be deployed in three levels:

class I video access point (station segment): and receiving image information captured by the front-end camera, and performing video analysis, alarm processing and image storage.

Regional nodes (road bureau/customer private bureau): receiving alarm information reported by the I-type video access point, distinguishing alarm levels, judging whether a core node needs to be reported or not, storing all alarm image resources, calling all image resources of the I-type video access point, and performing manual intervention if necessary.

Core node (railroad department): and receiving the report alarm information of the area nodes, storing important alarm image resources, calling the alarm image resources, and performing data analysis and processing on the alarm information.

Each level can call the lower image resource and has independent image processing function.

3. Monitoring system deployment

(1) Monitoring area

The project focuses on monitoring areas along the railway:

for the condition of railway enclosure (cement fence, protective net, drainage ditch, working door, solid wall, etc.), a high-definition infrared gun type camera is mainly adopted, each corner of a key area can be monitored, basically no blind spot is left, and the behaviors of damaging and invading the enclosure can be found;

for the condition of mainly covering the line and facilities along the line, an outdoor infrared gun type camera is adopted, a color fixed camera and an automatic aperture lens with low illumination and strong light inhibition functions are selected as much as possible, an invaded object can be clearly distinguished under the condition of light interference, and the behaviors of staying along the line, setting up obstacles on the line and approaching or damaging the facilities along the line can be found;

for cameras arranged at tunnel portals, bridge evacuation channels and machine rooms along the lines, infrared dome/gun type cameras are adopted, so that the characteristics of clothing and face of people can be preferably distinguished, and invasion, damage and barrier setting behaviors aiming at the fortification areas can be identified.

(2) Front end camera deployment along line

Infrared LED camera deployment

4-5 cameras which are installed on a roadbed section and contact net rods and are spaced at intervals of 200-300 meters are adopted.

Back-end video analytics hardware deployment

Deploying a video analysis server, as shown in fig. 3, using a streaming media server to forward the branches of the video signal to the server, and analyzing the simulated intrusion scene.

4. Algorithm resource assessment

(1) Test server configuration

A CPU: intel (R) Xeon (R) CPU E5-2620 v3 @ 2.40GHz 6 core

Memory: 64GB

A display card: NVIDIA TITANX video memory: 11GB

(2) Test of detection algorithm

The detection algorithm occupies 1GB of the video memory, the running speed is 10fps, and the server is expected to run 4-6 paths of videos.

(3) Suggesting server configurations

And measuring and calculating by taking the configuration of the test server as a reference, measuring and calculating 4-6 paths of videos running in each server, 1 path of videos per 200 meters, and configuring 1 server per kilometer.

5. Storage budget management

In the scheme, a mode of directly importing the video source from the streaming media coding server and storing and managing video analysis results in a centralized mode is adopted. In this mode, the special alarm image is stored in the disk arrays of the video area node and the video core node via the communication network. And the centralized management, retrieval and retrieval of video analysis results are realized. The stored data supports reliability protection of RAID0, 1, 5, 10 and the like, RAID10 is recommended to be used for storing alarm backup data, one hard disk is used for hot standby, and the other hard disk is used for data verification.

The storage capacity is calculated as follows:

and calculating the image storage capacity by using the 2Mbps single-path video image code stream:

2Mbps/8=250 KB/sec; the hourly capacity is 3600 seconds multiplied by 250 KB/second which is 900 MB/hour; the 24-hour storage capacity per image per day is 24 hours × 900 MB/hour is 21.6 GB/day.

The required storage capacity can be specifically calculated according to the number of specific monitoring points, and the storage equipment can be added at any time according to the needs and can be uniformly managed.

6. The management platform system comprises:

the railway intelligent video analysis system has video analysis and alarm functions, and can be compatible with the existing system and equipment from two layers, one is a hardware layer and the other is a subsystem layer.

From the hardware level, the system supports direct access of various hardware devices, including a front-end camera, a sensor, a collector and related devices. Only a communication interface and a communication protocol are provided, and the equipment integration of the existing manufacturer is supported.

From the subsystem level, the system supports the system of each factory of access, for example, each subsystem such as video system, access control system, access & exit system, broadcasting system, information issuing system, fire extinguishing system. Similarly, for systems from other vendors, the platform supports access openly.

7. Description of a network management functional module:

as required, the main functions to be implemented include: intrusion detection, train detection, pedestrian detection, line crossing detection, moving object detection, and the like. The system can realize the identification and analysis of the railway intrusion behavior and can give an alarm in time. The main functional modules comprise modules of video access, warning area setting, motion detection, motion tracking, behavior analysis, semantic understanding, alarming and the like. The specific functions of each part are as follows.

(1) Video access module

The video access module realizes the input function of video sequences. The input data in the system is a video file with AVI format, and the output is a video frame stream.

(2) Alert zone setting module

The system can support setting of the crossing defense area of the unilateral line segment/straight segment irregular polygon picture frame according to the actual situation of the railway monitoring environment.

(3) An intrusion detection module:

after the intrusion detection or the warning trip line detection (direction can be set) or the warning area is set, the system can automatically detect the picture condition, and when people or foreign matter intrusion prevention areas are found, the system can automatically alarm to remind monitoring personnel to pay attention to the field condition.

(4) A train detection module:

when a train appears in the rail system in the monitoring area, the system can automatically identify and prompt safety personnel to notice the field condition when the train enters an automatic picture frame.

(5) A pedestrian detection module:

some suspicious personnel wander around the area for a long time at or near the rail, the system automatically detects and actively gives an alarm to remind security personnel to pay attention to the field condition, and manual intervention is performed if necessary.

(6) The line crossing detection module:

the method comprises the steps that a manager manually draws a virtual line on a system video image according to needs, and when the system detects that a moving target crosses the virtual line, namely intersection exists between a pixel point set of a foreground target of the image and a pixel point set of a tripwire on the image, the system gives an alarm.

According to the practical application scene, the line crossing detection is divided into two types: one is bidirectional tripwire detection, namely, no matter how a target moves, the system can give an alarm immediately as long as the target crosses the wire, and the method is mainly used for preventing a key equipment deployment area; the other is one-way tripwire detection, when a moving target crosses a tripwire, the system firstly finds the moving target from the continuous N frames of images in the front, then analyzes the moving direction of the moving target, compares the moving direction with a user-defined tripwire crossing rule, finally judges whether the crossing rule is violated or not and whether an alarm is required or not, and is mainly used for preventing a forbidden area of a railway driving area.

(7) A moving object detection module:

the main functions of the moving object detection module are background modeling and motion detection, judgment of the types of the objects which invade and wander, and analysis and recognition of invasion behaviors. Classifying the objects by adopting a classification method based on shape and duty ratio when judging the types of the invaded objects; during behavior analysis, firstly, a behavior model based on point tracks is adopted to model the intrusion behavior of an object, then language description is recognized and given, and types are distinguished according to standards such as personnel, vehicles, animals, foreign matters and the like.

(8) An alarm module:

the alarm module has the main function of giving an alarm to abnormal events in various forms such as frame bouncing, screen flashing, voice, howling and the like according to the result of the behavior analysis module.

The utility model discloses the concrete application way is many, and the above-mentioned only is the preferred embodiment of the utility model, should point out, to ordinary skilled person in this technical field, under the prerequisite that does not deviate from the utility model discloses the principle, can also make a plurality of improvements, and these improvements also should be regarded as the utility model discloses a scope of protection.

Claims (2)

1. An image intelligent recognition device for railway environment, which is characterized in that: the system comprises front-end equipment, monitoring equipment and user-side equipment;

the front-end equipment is arranged at an inter-regional tunnel portal, a highway-span iron overpass, a bridge evacuation channel, equipment rooms along the line, the inside and outside, a dispatching office portal and a station throat area;

the monitoring equipment is divided into three levels from low to high, and comprises an I-type video access point, an area node and a core node, wherein each level can call lower-level image resources and has an independent image processing function;

the I-type video access point is provided with a video processing server, the video processing server is connected with the front-end equipment, receives image information captured by the front-end camera and performs video analysis, alarm processing and image storage;

the area nodes are provided with alarm analysis servers, the alarm analysis servers are connected with the video processing servers and receive alarm information reported by the type I video access points, alarm levels are distinguished, whether core nodes need to be reported or not is judged, and all alarm image resources are stored;

the core node is provided with a data analysis server, the data analysis server is connected with the alarm analysis server, receives alarm information reported by the area nodes, and stores important alarm image resources;

and the user side equipment receives and processes the alarm signal through the data analysis server of the core node.

2. The intelligent image recognition device for a railway environment of claim 1, wherein: the front-end equipment adopts a high-definition infrared LED camera and a long-focus laser camera.

Images