KR20200017601A - Analysis of illegal activities and monitoring based on recognition using unmanned aerial vehicle and artificial intelligence deep running that can monitor illegal activities in the field farm - Google Patents

Abstract

The present invention relates to an unmanned aerial vehicle capable of monitoring illegal practice in a field farm and a method for monitoring illegal practice by using the same, which, when a vessel committing illegal practice exists in a field farm, can monitor and warn the corresponding vessel to make the same stop the illegal practice, and can allow the vessel that committed the illegal practice to perform follow-up measures in various methods. To this end, provided is the method for monitoring illegal practice by using an unmanned aerial vehicle capable of monitoring illegal practice in a field farm, which is characterized by comprising: a step in which after the power of an unmanned aerial vehicle is turned on (S100), and the unmanned aerial vehicle is operated (S110), image information is received through a camera of the unmanned aerial vehicle (S120); a step in which in the case that an event occurs during the operation of the unmanned aerial vehicle, the same is moved to a corresponding location by using GPS information when information of a suspected vessel is received according to the event (S140) before an image is filmed (S150), such that the illegal practice-suspected vessel is filmed when the event is detected through a camera unit (140) of the unmanned aerial vehicle (100) (S150), wherein the filmed image is stored by the unmanned aerial vehicle, and the same is transmitted to the outside through a communication unit of the unmanned aerial vehicle (S160); and a step in which a voice is transmitted through an unmanned aerial vehicle controller (300) which controls the unmanned aerial vehicle, or a warning or a warning sound predetermined in a warning broadcasting unit of the unmanned aerial vehicle is outputted (S170).

Description

Analysis of illegal activities and monitoring based on recognition using unmanned aerial vehicle and artificial intelligence deep running that can monitor illegal activities in the field farm}

The present invention relates to an unmanned aerial vehicle capable of monitoring illegal activities of a field farm, and more particularly, a field farm illegal activity monitor capable of monitoring a field farm illegal activity that can monitor and follow-up illegal activities occurring in a field farm. The present invention relates to a surveillance method based on the analysis and recognition of illegal activities using possible drones and artificial intelligence deep learning.

In general, a drone is a type of unmanned aerial vehicle that is flying by radio waves without a person riding.

Such drones are used in various fields, and in particular, fire suppression, rescue and pesticide spraying.

Conventional techniques related to drones include fire suppression drones capable of firing fire extinguishers proposed in Korean Patent Laid-Open Publication No. 10-2017-0025386.

The patent has a fire suppression drone capable of firing a fire extinguisher, fire tracking unit for tracking the source of fire based on the thermal imaging data of the thermal imaging camera and the thermal sensor data of the built-in thermal imaging camera; A throwing unit capable of continuously firing a plurality of fire extinguishers, and adjusting a firing angle and a firing distance of the fire extinguisher based on a tracking result of the fire tracking unit; Fire suppression drones have been proposed, including a plurality of propellers and a moving unit for adjusting the movement direction.

Another conventional technique is a marine rescue drone proposed in the Republic of Korea Patent Publication No. 10-2017-0013100 and a marine rescue system using the same.

Patents include drone bodies; A drone GPS module accommodated inside the drone body and generating drone position data by detecting a position of the drone body; A flight propeller provided to be able to fly outside the drone body; A tube holder formed at a lower portion of the drone body to detachably hold a lifesaving tube; And an emergency call signal receiving unit which is accommodated inside the drone body and receives an emergency call signal including the distress position data from an external maritime control center, and compares the drone position data with the distress position data to move upwards of the distress. A marine rescue drone has been proposed that includes a controller including a flight control unit for controlling the flight propeller and a throwing control unit for controlling the tube holder to throw the lifesaving tube toward the distress after movement.

Another conventional technique is a drone with a pesticide container proposed in the Republic of Korea Utility Model Registration No. 20-0479365.

The utility model includes a drone body; A drone GPS module accommodated in the drone body and generating drone position data by detecting a position of the drone body; A flight propeller provided to be able to fly outside the drone body; A tube holder formed at a lower portion of the drone body and detachably holding a lifesaving tube; And an emergency call signal receiving unit which is accommodated inside the drone body and receives an emergency call signal including distress position data from an external maritime control center, and compares the drone position data with the distress position data and moves upwards of the distress. A marine rescue drone has been proposed that includes a controller including a flight control unit to control the flight propeller and a throwing control unit to control the tube holder to throw the lifesaving tube toward the distress after movement.

The above-mentioned conventional technologies are all used for fire suppression, lifesaving, and pesticide spraying using drones, so there is a problem that there is a limit to use for the purpose of managing the offshore field farms as in the present invention. something to do.

Prior Document 1: Republic of Korea Patent Publication No. 10-2017-0025386-Fire suppression drones that can fire a fire extinguisher Prior Document 2: Korean Unexamined Patent Publication No. 10-2017-0013100-Marine drone for marine rescue and marine rescue system using same

Therefore, the present invention is to solve all the disadvantages and problems of the prior art as described above, if there is a vessel or a mockery illegal in the field farms to stop illegal activities through monitoring and warning of the vessel or mock And unmanned aerial vehicles and artificial intelligence deep learning capable of monitoring illegal activities of field farms to enable analysis, recognition, and monitoring of illegal activities using artificial intelligence deep learning for vessels or migrating torts. It aims to provide a monitoring method based on the analysis and recognition of tort.

In order to achieve the above object, the unmanned aerial vehicle capable of monitoring the illegal activity of the field farm in the present invention receives a control signal for controlling the unmanned aerial vehicle 100 from the unmanned aerial vehicle controller 300 and detects the vessel from the marine radar 700. A communication unit 110 receiving a radar signal and transmitting the image data and the position information photographed by the unmanned aerial vehicle 100 to the unmanned aerial vehicle controller 300 and a smart device 500 of an official of a field farm or an unmanned aerial vehicle operator; A sensor unit 120 for sensing a wind direction, a wind speed, and a current altitude according to the moving direction of the unmanned aerial vehicle 100, measuring a rotational movement state, and detecting a tilt; Flight unit 130 for the flight of the unmanned aerial vehicle 100; A camera unit 140 for photographing the surroundings while the unmanned aerial vehicle 100 moves; Image storage unit 160 for storing the image taken by the camera unit 1400; GPS receiver 170 for receiving the current position of the unmanned aerial vehicle 100; Ship from the image taken while moving the unmanned aerial vehicle 100 Illegal activity identification unit 180 for identifying a warning warning output unit for outputting a warning broadcast or warning sound, or the voice or warning broadcast transmitted from the unmanned vehicle adjuster 300, the communication unit 110; The sensor 120 includes the flight unit 130, the camera unit 140, the image storage unit 160, the GPS receiver 170, the illegal activity identification unit 180, and the warning broadcaster 190. By controlling the vehicle 100 to control the flying unit 130 according to the control data transmitted from the unmanned vehicle adjuster 300 through the communication unit 110, from the image taken by the camera unit 140 If the illegal activity identification unit 180 identifies the vessel through the communication unit 110 To transmit to the smart device 500 of the vehicle regulator 300, the field farm official or the unmanned aerial vehicle operator, and outputs the voice through the warning broadcast unit 190 for the voice signal from the data transmitted from the drone regulator 300 And calculates the location information received from the GPS receiver 170 and the driving area set by the unmanned vehicle adjuster 300 to the unmanned vehicle adjuster 300 through the communication unit 110 when it is out of a preset range. Notifying and, if the flight time and returning place of the unmanned aerial vehicle 100 is calculated by calculating the remaining battery, the control unit 220 for controlling the flight time to the corresponding place to transmit to the unmanned aerial vehicle controller 300; Characterized in that configured to include.

In addition, the monitoring method according to the analysis and recognition of illegal activities in the field farm using the artificial intelligence deep learning of the present invention to achieve the above object is the unmanned vehicle is turned on (S100), while the unmanned aerial vehicle is operated (S110), Receiving image information through a camera (S120); If an event occurs during the operation of the unmanned aerial vehicle, if suspicious vessel information is received according to the event, the GPS information is used to move to the corresponding position (S140), and the image is taken (S150), and the event is the unmanned aerial vehicle 100. If detected through the camera unit 140 of the illegal activity suspected ship (S150), the captured image is stored in the unmanned vehicle, and transmitting it to the outside through the communication unit of the unmanned aerial vehicle ( S160); And transmitting a voice through the unmanned vehicle adjuster 300 for adjusting the unmanned vehicle, or outputting a preset warning broadcast or warning sound to the warning broadcaster of the unmanned aerial vehicle (S170).

Here, the operation of the unmanned aerial vehicle is a normal flight by an unmanned aerial vehicle manager or an official of a field farm, or a vessel radar radar receives a ship information on a field farm, or a vessel access signal from a buoy installed in a field farm. It is performed when received through a smart device (smartphone, smart pad) or a pre-scheduled receiver of the farm.

And in order to achieve the above object, the surveillance method according to the analysis and recognition of tort using artificial intelligence deep learning of unmanned aerial vehicle capable of monitoring illegal activity of the field farms of the present invention, the unmanned aerial vehicle is a diving of ships, intruders, seafood networks, intruders and Step (S400) data of the number of theft and buoy buoys of the field farm to be monitored and theft of the hand including the act of raising the hand over the buoy; The unmanned aerial vehicle generates a background area model by using the previous image of the field farm, and removes the image from the original image to identify the object area and the background area (S410); The unmanned aerial vehicle detects at least one of a shelf, an intruder, a buoy shape, a size, a field farm boundary, a position, a distance, a slope, a color, a text, a sign in real time (S420); The unmanned aerial vehicle detects a person by extracting a motion region from two frames of an incoming image in order (S430) (S440); The unmanned aerial vehicle removes a false image of the detected intruder region (S450); The unmanned aerial vehicle performs deep learning on the collected image (S470); And detecting, by the unmanned aerial vehicle, abnormal behavior by using a change of a screen or a scene change detection method of the image through the deep learning, and transmitting the collected image information to an administrator terminal (S480) (S490). Characterized in that made.

According to an embodiment of the present invention has the following effects.

First, if there is a ship that is illegal in the field farm, the illegal activity can be stopped by monitoring and warning the vessel.

Second, follow-up measures can be carried out in various ways, such as acquiring an image for arresting an illegal vessel through video recording and tracking of the illegal vessel.

Third, it is easy to monitor illegal activities on field farms by analyzing and recognizing illegal activities using AI deep learning and monitoring accordingly.

1 is a block diagram illustrating an unmanned aerial vehicle capable of monitoring illegal activities of field farms according to the present invention.
2 is a block diagram illustrating a regulator for adjusting an unmanned aerial vehicle capable of monitoring illegal farming activities in a field farm according to the present invention.
3 is a view for explaining the concept of illegal activity monitoring using the unmanned aerial vehicle capable of monitoring the illegal activity of the field farm according to the present invention.
4 and 5 is a view showing a mock and a diver in tort monitoring using an unmanned aerial vehicle capable of monitoring illegal farm activities in the field farm according to the present invention.
6 is a flowchart illustrating a method for monitoring illegal activity using an unmanned aerial vehicle capable of monitoring field farm illegal activity according to the present invention.
7 is a view for explaining a field farm illegal activity monitoring system according to the present invention.
8 is a flowchart illustrating a method for operating an unmanned aerial vehicle in a field farming illegal activity monitoring system according to the present invention.
9 is a flowchart illustrating a monitoring method according to the analysis and recognition of illegal activity using artificial intelligence deep learning of the field farm illegal activity monitoring system according to the present invention.

When described in detail with reference to the accompanying drawings a preferred embodiment of the present invention.

In addition, the terminology used in the present invention was selected as a general term that is widely used at present, but in certain cases, the term is arbitrarily selected by the applicant, and in this case, since the meaning is described in detail in the corresponding part of the present invention, a simple term is used. It is to be understood that the present invention is to be understood as a meaning of terms rather than names. In addition, in describing the embodiments, descriptions of technical contents that are well known in the technical field to which the present invention belongs and are not directly related to the present invention will be omitted. This is to more clearly communicate without obscure the subject matter of the present invention by omitting unnecessary description.

1 is a block diagram illustrating an unmanned aerial vehicle capable of monitoring illegal activities of field farms according to the present invention.

Field unmanned aerial vehicle 100 capable of monitoring illegal farming activities according to the present invention, as shown in Figure 1, the communication unit 110, the sensor unit 120. Flight unit 130, camera unit 140, image conversion unit 150, image storage unit 160, GPS receiver 170, illegal activity identification unit 180, warning broadcast unit 190, signal transmission unit ( 200, the travel distance calculator 210 and the controller 220.

The communicator 110 receives a control signal or an audio signal for controlling the unmanned aerial vehicle 100 from the unmanned aerial vehicle controller 300, and receives the image data and the position information photographed by the unmanned aerial vehicle 100. The field farm and transmits to the smart device 500 of the official or unmanned aerial vehicle operator. In addition, receiving the identification information of the buoy 600 or the approach signal of the vessel 640 from the buoy 600 to transmit the position information of the field farm buoy 600, and receives the radar signal from the marine radar 700, the drone adjuster Send to 300. Meanwhile, the communication unit 110 wirelessly transmits the driving information of the unmanned aerial vehicle to the external monitoring server 900 at a predetermined interval. The smart device 500 or the field farm buoy 600 and the radar 700 are shown in FIG. 3 to be described later.

The sensor unit 120 includes one or more of the wind direction / wind speed / altitude sensing unit 121, the geomagnetic sensor 122, the acceleration sensor 123, and the gyro sensor 124, and the wind direction / wind speed / altitude sensing unit 121. ) Senses the wind direction and wind speed and the current altitude according to the moving direction of the unmanned aerial vehicle 100. In this case, the wind direction / wind speed / altitude sensing unit 121 includes a wind direction sensor, a wind speed sensor, and an altitude sensor.

The geomagnetic sensor 130, the acceleration sensor 140 and the gyro sensor 150 measures the moving direction (progression direction), the rotational movement state of the unmanned aerial vehicle 100 and detects the tilt.

Flight unit 130 is composed of one or more of the rotary blade 131, the drive motor 132, the battery 133 and the power generation unit 134, the rotary blade 131 is driven by the drive motor 132 For the flight of the unmanned aerial vehicle 100, it is composed of a plurality, the drive motor 132 may also be composed of a plurality of to drive each of the rotary blades 131 independently. At this time, the unmanned aerial vehicle 100 of the present invention may be configured by an unmanned aerial vehicle, for example, a drone. In general, a drone is equipped with a camera, a sensor, a communication system, and the like, and also has various weights and sizes. Drones were first introduced for military use, but have recently been extended to high-altitude shooting and delivery. Recently, drones have been used for various purposes, such as spraying pesticides and measuring air quality. One of the characteristics of the drone is composed of a plurality of rotary blades 131 to enable vertical take-off and landing, and a plurality of drive motors 132 for driving each of the plurality of rotary blades 131 is also configured.

The battery 133 supplies power for driving the driving motor 132.

The driving engine 190 is for driving the driving motor 170, and receives the aviation oil, gasoline, diesel, etc. from the fuel tank 200 to drive the driving motor 170.

The power generation unit 134 charges the battery 133 by generating power according to the rotation force when the rotary blade 131 rotates.

Meanwhile, in the present invention, only the method using a battery is described, but a hybrid method or a method using only a driving engine may also be combined with a driving engine using gasoline as a fuel instead of a driving motor.

The camera unit 140 photographs the surroundings while the unmanned aerial vehicle 100 moves. The camera unit may be added with a zoom function and a tilt function.

The image converting unit 150 converts the image to compress and store the image photographed by the camera unit 140.

The image storage unit 160 stores the image converted by the image converter 150.

The GPS receiver 170 receives a current position of the unmanned aerial vehicle 100.

The illegality identification unit 180 identifies the chip particle and the vessel 640 such as the migrating 620 or the diver 630 in the image photographed during the movement of the unmanned aerial vehicle. For example, the vessel 640, the muzzle 620, or the diver 630 may be detected using a technique similar to that of detecting a face in an image. Among these identification technologies, the ship identification technology is an automatic identification device of AIS (automatic identification device) to support collision avoidance and maritime search and rescue activities between ships by automatically transmitting and receiving ship location information between ships and ships, ships and land control stations. system). In general, the port control system manages vessel traffic in connection with the vessel AIS. In order to enable efficient traffic management, ship recognition and display system interworking with AIS is required. In the present invention, an image-based vessel detection using background estimation with respect to a sea or port image input from the camera unit 140 may be used in conjunction with an AIS signal of a corresponding vessel detected. Techniques for detecting chip particles, such as the vessel 640, the migrating 620 or the diving shelf 630 will be described in more detail later.

The warning broadcast unit 190 outputs a warning broadcast or warning sound input in advance, or outputs a warning broadcast transmitted from the drone adjuster 300. Of course, it is possible to output a voice other than a warning broadcast.

The signal transmitter 200 may not be able to return to the drone due to abnormal weather or lack of battery or the flight unit 130, and falls in the sea or river, so that the drone 100 may be easily found. ) To emit light. For this purpose, the unmanned aerial vehicle 100 is made of a material floating in water, and the main parts are waterproofed.

The travel distance calculator 210 calculates the travel distance of the actual unmanned aerial vehicle 100 according to the change of the location information received through the GPS receiver 170.

The control unit 220 is a communication unit 110, the sensor unit 120. Flight unit 130, camera unit 140, image conversion unit 150, image storage unit 160, GPS receiver 170, illegal activity identification unit 180, warning broadcast unit 190, signal transmission unit ( Control the flight unit 130 according to the control data transmitted from the unmanned vehicle controller 300 through the communication unit 110 by controlling the driving distance calculation unit 210 and the traveling distance calculation unit 210, and the image captured by the camera unit 140. To convert the image to the image storage unit 160, the illegal activity identification unit 180 in the image taken through the camera unit 140 when identifying the vessel or the mugi drone through the communication unit 110 adjuster (300), the field nursery personnel or unmanned aerial vehicle operator to transmit to the smart device 500, and outputs the voice through the warning broadcast unit 190 for the voice signal from the data transmitted from the drone regulator 300, GPS Location information received from the receiving unit 170 and the operating spirit set in the unmanned vehicle adjuster 300 When the station is out of the preset range by calculating the station, it is notified to the drone adjuster 300 through the communication unit 110, and the distance calculated by the travel distance calculation unit 210 at each preset time is externally communicated through the communication unit 110. Control to transmit to the monitoring server 900. In addition, if the flight time and returning place of the unmanned aerial vehicle 100 is calculated by calculating the remaining amount of battery, the flight time to the corresponding place is calculated and transmitted to the drone adjuster 300.

2 is a block diagram illustrating a regulator for adjusting an unmanned aerial vehicle capable of monitoring illegal farm activities in a field farm according to the present invention.

The adjuster 300 for adjusting the unmanned aerial vehicle capable of monitoring illegal farm activities in the field farm according to the present invention is as shown in FIG. 2, the communication unit 310. Flight operation unit 320, display unit 330, camera control unit 340, voice input unit 350, alarm generating unit 360, ship tracking selector 370, driving area setting unit 380 and control unit 390 It consists of

Here, the communication unit 310 communicates with the unmanned aerial vehicle 100 to transmit various control signals or the image information from the unmanned aerial vehicle 100 or the identification information of the buoy from the field farm form buoy 600 and the radar signal from the marine radar 700. Receive the transmission.

The flight manipulation unit 320 generates a signal for controlling the flight unit 130 of the unmanned aerial vehicle 100.

The display unit 330 displays the current position of the unmanned aerial vehicle 100 or displays the image data transmitted from the unmanned aerial vehicle 100, and separately detects only the face and separately displays the face of the vessel in the image data. Can be displayed in the same manner as For example, when several ships are identified and transmitted at the same time, masking of each ship and displaying the ship area at the same time, and when a ship is called an object, separates and tracks an object. Generally, a face is detected by a security camera. And a technique for tracking the detected face.

The camera controller 340 controls the camera unit 140 of the unmanned aerial vehicle 100 and transmits a zoom or tilt control signal.

The voice input unit 350 may perform a warning broadcast or various guide broadcasts to be broadcast through the warning broadcast unit 190 of the unmanned aerial vehicle 100.

The alarm generator 360 notifies the alarm when ship identification data is transmitted from a signal transmitted through the communication unit 310 from the unmanned aerial vehicle 100 through the communication unit 310, and the flight time of the unmanned aerial vehicle 100 is increased. If it is close to the return time or if the remaining battery capacity is a preset capacity (e.g. 10 to 15%), output it as an alarm sound or LED light.

When the vessel tracking selector 370 is selected as a tracking vessel through the drone adjuster 300 for the vessel identified in the image identified by the illegal activity identification unit 180, the drone 100 automatically tracks the vessel. In this way, the operator who operates the drone adjuster 300 automatically tracks the selected vessel without the continuous operation of the drone operator. At this time, the camera unit 140 is also set to continuously photograph the selected vessel. In addition, even if illegal activities involving invaders such as muzzles and divers are to be traced to the vessel used by the chipper.

The driving area setting unit 380 sets the unmanned aerial vehicle 100 to travel in a preset section. In this regard, it is possible to use a buoy installed in the field farm or a method of inputting the coordinates of the field farm.

The control unit 390 is a communication unit 310. The flight control unit 320, the display unit 330, the camera controller 340, the voice input unit 350, the alarm generator 360, the ship tracking selector 370 and the driving area setting unit 380 are controlled.

3 is a view for explaining the concept of illegal activity monitoring using the unmanned aerial vehicle capable of monitoring the illegal activity of the field farm according to the present invention.

The illegal activity monitoring concept using the unmanned aerial vehicle capable of monitoring the illegal farming of field farms according to the present invention is as shown in FIG. 3. The unmanned aerial vehicle 100 of the plurality of first to nth field farms 410, 420, and 430 is used. It is possible to monitor vessels operating illegally while they are operating. In general, field farm personnel can receive the position information of the ship through the maritime radar 700, it may be difficult for the field farm officials to monitor the vessel in operation illegally throughout the day, so it is operating near the field farm from the maritime radar When the vessel information is transmitted, the vessel can move the unmanned aerial vehicle 100 to receive an image of the vessel and receive the image to the smart device 500 to determine whether the vessel is in illegal activity, and identify the vessel. You can perform immediate follow-up actions (eg maritime report) through the information (e.g. ship name) and video information. At this time, there is a buoy 600 in the field farm 400, and can receive the identification information from the buoy 600, the recent buoy solar self-powered mining unit for preventing accidents caused by passing through the field farm of the ship And a GPS sensor, and in some cases, a temperature sensor for monitoring water temperature change in a field farm, and a laser sensor including an approaching detection function. Therefore, when the Mug 620 or the diver 630 and the vessel 640 approaches the buoy 600, the smartphone of the field nursery notifies the murco 620 or the diver 630 and the vessel 640, Accordingly, it is also possible to move the unmanned aerial vehicle 100 to the corresponding buoy position to photograph. These buoys may be fixed through the buoy fixing device 610 or installed directly on the cage device.

4 and 5 is a view showing a mock and a diver in tort monitoring using an unmanned aerial vehicle capable of monitoring the illegal farming of field farms according to the present invention.

In the illegal activity monitoring using the unmanned aerial vehicle capable of monitoring the illegal activity of the field farm according to the present invention, the mercury is shown in Fig. It is also a term used by the back. The mercury wears a wetsuit, such as a spacesuit, as shown in FIG. 4, and is supplied with air through a hose connected to the surface of the water. They move and work in the water as if the astronauts were swimming. They play a wide variety of important roles, from coastal roads and breakwater construction, to submerged shipwrecks and to the salvage of bodies. Not only can you work in the water for a long time, but you can also talk through the hose.

In general, shallow water submersible, helmet-type submersible, circulating self-contained aquatic respirator, scuba diver, etc. are used according to the work purpose, but the helmet-type submersible as shown in FIG.

Recently, however, there is a case where a diver 630 using an oxygen tank is used as shown in FIG. 5, and the seafood or aquatic products collected by using the mercury or the diver equipment shown in FIGS. 4 and 5 are placed in the trap 621. Post to the award.

In the present invention, the mock 620 and the diver 630 by using the vessel 640 to approach the field farm, the mock 620 and the diver 630 is put down to move back to the pre-scheduled time after moving the mock ( The illegal act of loading the 620 and the divers 630 back to the ship 640 may also be controlled.

6 is a flowchart illustrating a method for monitoring illegal activity using an unmanned aerial vehicle capable of monitoring field farm illegal activity according to the present invention.

According to the present invention, the illegal activity monitoring method using the unmanned aerial vehicle capable of monitoring illegal farm activities in the field farm is turned on (S100), while the unmanned aerial vehicle is operated (S110), while the unmanned aerial vehicle is operated. Receive image information through (S120). At this time, the operation of the unmanned aerial vehicle may be a normal flight by an unmanned aerial vehicle manager or a field farm official, or may have received vessel information on a field farm from a maritime radar, and a vessel approach signal or a mercury, from a buoy installed on the field farm, This may be the case when an intruder such as a diver is received through a smart device (smartphone, smart pad) or a predetermined receiver.

If an event occurs during the operation of the unmanned aerial vehicle, if suspicious vessel or intruder information is received according to the event, it moves to a corresponding position using GPS information (S140) and photographs an image (S150).

On the other hand, if the event is detected through the camera unit 140 of the unmanned aerial vehicle 100 photographs a suspicious vessel or intruder (S150). In this case, the camera may shoot through the zoom function or the tilt, or may shoot a still image or a moving image.

The captured image is converted and stored in the unmanned aerial vehicle 100 and transmitted to the outside through the communication unit 110 (S160). Accordingly, the drone adjuster 300 or the smart device 500 may display and store the corresponding image. In this case, if the illegal operation is not suspected, such as a ship or a passing ship in advance, the operation of the unmanned aerial vehicle may be terminated.

However, if the vessel is suspected of illegal activity, mercury, divers may transmit a voice through the drone regulator 300, or may output a preset warning broadcast or warning sound (S170).

In addition, the ship tracking selection unit 370 of the drone adjuster 300 transmits a tracking command for the illegal ship (S180).

It automatically tracks and photographs the vessel. Of course, it is possible to manually shoot while tracking through the adjuster (S190).

On the other hand, if the remaining battery time is the return point, an alarm is generated in the drone adjuster 300, and accordingly stops shooting, and returns to the return location (S200).

On the other hand, it is determined whether the return is successful (S210), and if the return fails due to abnormal weather or lack of battery, the unmanned aerial vehicle 100 transmits a rescue signal through the signal transmitter 200 (S210).

7 is a view for explaining a field farm illegal activity monitoring system according to the present invention.

As shown in FIG. 7, the field farm illegal activity monitoring system according to the present invention includes a plurality of unmanned aerial vehicles 100 with respect to a plurality of first to n-th field farms 410, 420, and 430, and an unmanned aerial vehicle for controlling them. A plurality of buoys 600 for informing the installation area of each of the vehicle adjuster 300, the manager or the relationship of the smart device 500, and the plurality of first to nth field farms 410, 420, and 430, respectively Including a plurality of unmanned aerial vehicle 100 for a plurality of field farms, and a monitoring server 900 and a financial company server 1000 for efficiently providing the monitoring information to each smart device 500 of the plurality of parties It is composed.

The monitoring server 900 is a communication unit 910, database 920, operation information storage unit 930, management information storage unit 940, unmanned vehicle management request unit 950, cost calculation and billing unit 960 And a control unit 970.

On the other hand, the communication unit 910 of the monitoring server 900 receives the unmanned aerial vehicle distance information from the unmanned aerial vehicle 100 through the communication network 800, to the unmanned aerial vehicle manager or related personnel for the periodic management of the unmanned aerial vehicle 100 Send necessary information (regular checks, billing, etc.).

The database 920 includes buoy information (GPS location information) for each field farm to be monitored, unmanned vehicle basic information (model name, parts information and performance information, etc.) and unmanned vehicle contract information (management cost (including periodic performance check AS). ), Rental + administrative costs, etc.), unmanned aerial vehicle user information (name, phone number, mobile phone number, email address, smart device identification, purchase, unmanned aerial vehicle information, etc.), unmanned aerial vehicle contact information (name, telephone number, Mobile phone number, drone information under charge, etc.) are stored.

The driving information storage unit 930 stores the information when the distance information actually traveled by the unmanned aerial vehicle 100 is transmitted from the unmanned aerial vehicle 100 at a set period.

The management information storage unit 940 stores periodic inspection information for each unmanned aerial vehicle 100, parts repair or replacement information, and the like.

The unmanned aerial vehicle management request unit 950 requests a request for rescue from the unmanned aerial vehicle personnel upon periodic check or loss according to a contract set with the unmanned aerial vehicle user.

The cost calculation and billing unit 960 calculates management costs or rental costs for the unmanned aerial vehicle 100, parts repair or replacement costs, costs according to the driving distance, and insurance premiums for loss of the unmanned aerial vehicle.

This can be useful when the field farm operator wants to operate an unmanned aerial vehicle, if the initial purchase cost is burdened by the performance of the unmanned aerial vehicle, and can solve the AS problem including periodic inspection. The aircraft can be used, and as it is used in the field farm, both the user and the manager can feel free to use insurance when the drone is lost.

The controller 970 controls the communication unit 910, the database 920, the operation information storage unit 930, the management information storage unit 940, the unmanned aerial vehicle management request unit 950, and the cost calculation and billing unit 960. do.

Meanwhile, the financial company server 1000 may be a bank, a card company, or PayPal.

8 is a flowchart illustrating a method for operating the unmanned aerial vehicle in the field farming illegal activity monitoring system according to the present invention.

In the method for operating an unmanned aerial vehicle in the field farm illegal activity monitoring system according to the present invention, as shown in FIG. 8, the field farm is registered in the monitoring server 900 (S300). At this time, the GPS location information for each field farm and the field farm manager information (name, phone number, mobile number, etc.) are registered together.

Subsequently, a plurality of identification devices for each field farm, that is, buoy information and an unmanned aerial vehicle for monitoring illegal activities for each field farm, are registered (S310). At this time, the manager information on the unmanned aerial vehicle is also registered.

Subsequently, an event is generated, and if the event is a periodic check, a message requesting a periodic check is sent to an unmanned aerial vehicle terminal (mobile phone, etc.) (S330). And if the repair, replacement parts, etc. were performed according to the result of the regular inspection, the corresponding information is updated in the monitoring server 900.

On the other hand, if the event is driving-related data from the unmanned aerial vehicle, the driving distance is accumulated (S340). This distance accumulation is not only used for billing by distance with the base rate, but also related to AS of unmanned aerial vehicle parts (S350). In other words, in case of a lot of operation, the maintenance time can be advanced earlier. In the opposite case, the maintenance time can be delayed.

On the other hand, if the event is missing the unmanned aerial vehicle requests a check and search to the unmanned aerial vehicle terminal (S360). In the case of such a loss, the monitoring server 900 is to determine the search cost or insurance premium when lost.

9 is a flowchart illustrating a monitoring method according to the analysis and recognition of illegal activity using artificial intelligence deep learning of the field farm illegal activity monitoring system according to the present invention.

The monitoring method according to the analysis and recognition of tort using artificial intelligence deep learning of the field farm tort monitoring system according to the present invention is basically a ship, an intruder (Murray, diver) image, a seafood net image, a buoy image, an intruder's diving and Various theft activities including the act of raising the hand to a buoy are data-formed in advance in the illegal activity identification unit 180 (S400). Such data can be continuously updated according to field experience or experiment. In addition, the unmanned aerial vehicle dataifies the field farm boundary and the number of buoys at the boundary in advance.

The method for distinguishing invasive images in the field farm according to the present invention can be applied to the following techniques.

The control unit 220 creates a background area model by using the previous image in the boundary area of the image confirmation (candidate area) field farm and removes it from the original image, thereby removing the object area (ship, intruder (mug, diver), buoy, etc.) and the background area. Separate (S410).

The control unit 220 detects each action in real time by checking the shape or size of the ship, the intruder (murco), the buoy, the rules and irregularities of the boundary line, location, distance, slope, color, text, signs, etc. (S420) .

In the detection of the intruder (murco) region, the controller 220 first extracts a movement region from two frames of an incoming image (S430).

In the case of the intruder (Murray) area detection, the human case detects whether or not the person is through the learned Idaboost classifier using the histogram feature of the motion window search and the direction component in and out of the field farm (S440).

At this time, the movement is detected using the histogram (HoG) characteristic of the direction component and the Idaboost classifier as follows.

In order to extract the histogram (HoG) feature of the directional component, the grayscale transformation and histogram equalization are performed to reduce the noise of the image caused by the light source for the target image.

After that, the target image is divided into blocks of 3 pixels in width and 3 pixels in height, and then the direction of the gradient is obtained for the pixels inside each block.

For each of the divided blocks, a histogram is generated based on the slope, and the slope having the largest slope distribution is selected as the average slope of the block.

Linearly combining the directional and magnitude values of the generated blocks results in a HoG feature vector. This feature vector was trained using the Idaboost classifier and classified into two types, human and non-human, for the input target image.

Motion region extraction performs grayscale image conversion and histogram equalization as a preprocessing process for the second frame and the first frame received from the camera or webcam image.

After that, threshold filtering is performed to obtain a binary image, and then a difference image is obtained.

Next, after finding all boundary lines that follow the difference image, the center point of the boundary line is extracted.

Through this, the position of the center point of the portion moved above a certain threshold can be obtained.

These pixels are clustered and extracted into regions of a certain size.

The clustering process generates an adjacency matrix based on a certain pixel distance with respect to the movement center points.

A rectangular region is created by grouping adjacent matrixes in a group by depth-first graph search.

On the other hand, the motion tracking performs multi-level matching on the detected invader (Murray) area to track the movement of the intruder (Murkey) and removes the falsely detected intruder from the list by performing false image removal (S450).

On the other hand, there is a need for an improved similarity calculation process considering the overlap situation.

For this purpose, human domain tracking measures the similarity with the detected intruder. If the similarity is measured and found to be the same, the location information of the intruder is updated.

By connecting the intruder location information, it is possible to identify the movement line of the intruder.

However, if similarity matching is consecutively wrong, proper detection cannot be made.

The matching process is therefore very important as a trace.

The proposed algorithm performs multilevel matching to reduce false detections.

Step 1 of multi-level matching is an improved similarity calculation process considering the overlap situation

The methods used to calculate the improved similarity are the distance change rate and the color histogram similarity overlap processing.

The motion estimation using the Kalman filter is regarded as a matching failure when the final similarity coefficient S of the multi-level matching step 1 is less than a predetermined value, and the process proceeds to step 2.

In the second step, the Kalman filter is used to predict motion and to compare the image similarity for the location.

Multilevel matching is performed in three stages, and similarity is measured stepwise for the detection window region using the HoG feature, the Kalman filter prediction window region, and the window of the adjacent motion region using the final similarity coefficient S calculated above.

The detection window area using the HoG feature, which is the first step, is compared with the list of intruders already held in the area for the intruder detected by the detection of the human area. Update

In the proposed method, an experimentally determined threshold of 0.8 was applied.

Step 2 applies to areas that do not exceed the threshold in step 1.

First, a window is created around the predicted point of the linear Kalman filter.

The similarity between the generated window image and the previously stored intruder image is calculated and when the threshold value is exceeded, it is determined that the matching is successful.

In the step 3 using the adjacent motion area, step 2 is performed on the failed area, and a window is generated for the center point of the nearest motion area and the similarity between the image of the corresponding position and the previously stored image is calculated.

And false image removal is necessary.

Attempts to trace using the above algorithms, but there still remains a problem of false windows due to the wrong HoG feature detection region.

To eliminate these false windows, we apply a false positive elimination algorithm with Time To Live (TTL).

That is, after applying the TTL property to each window, the value is reduced depending on the number of image processing times. If 0, the window is deleted.

TTL was applied to the HoG detection area and Kalman filter prediction window.

The TTL can be reduced because it is very unlikely that the real human region will appear below the threshold in two-stage matching for a certain number of times in succession.

This TTL information is provided for each tag and is set to an initial value of 15.

Then perform a sliding window search.

Since the learning image of the person used in the proposed algorithm is 64 * 128 pixels in size and the target camera image is 640 * 480 pixels, it is necessary to perform a linear search on the input image.

In the case of linear search, it proceeds from one direction to the end of the image.

In addition, because of the camera's field of view, people appearing from the top of the camera may appear small, so a human area search is performed with a 64 * 128 pixel window on the seven images that scale the input image by 70-130%. .

Parallel processing is performed using POSIX threads in the process of obtaining and classifying histogram features of direction components on images.

Meanwhile, field farm deep learning will be described.

In the present invention, such deep learning uses automatic semantic annotation based on image content.

BoVW (Bag of Visula Words) based classification is showing good performance in content-based image retrieval.

It is a method mainly used for SIFT-> image recognition, which is a feature point extraction algorithm used to construct BoVW, and extracts feature points that are robust to changes in image color, position, size, and rotation.

The SIFT algorithm is a position coordinate (X, Y), scale and orien of the keypoint descriptor feature point to have 128 dimensions that add up to 4 * 4 arrays and vectors divided in 8 directions.

Compose a frame with information of the tation.

After extracting the feature points from the image, we can use the k-means algorithm to obtain local visual blocks, or BoVWs, which are k Visual Words.

For annotations, suitable annotations can be obtained by sorting in order of highest matching grade through the conditional probability P (Kj / Rj) between the local visual blocks set Ri and the annotation keyword Kj composed of the text set.

However, algorithms such as SIFT and SURF do not perform well on background images that lack the characteristic elements of the image, such as Natural Scene Retrieval.

Recently, in the field of image recognition, CNN (Convolutional Neural Network) method using deep learning technique overcomes the disadvantages of SIFT by using convolution image processing method and shows excellent performance in image recognition.

Based on SIFT, BoVW is a method of constructing local visual blocks, that is, BoVW, after extracting feature points from an image.It basically measures the change information of image pixel values using differentiation, extracts regions with large changes, and forms a feature vector. It is more effective on object images with distinct boundaries than on images.

On the other hand, the deep learning technique based on CNN can achieve good results in the background image by utilizing the association of layers because pattern recognition is integrated without a separate feature extraction algorithm.

Deep learning is a field of machine learning that learns by creating a large number of layers with neural network.

Recently, various studies have been conducted to automatically tag semantic information on images using CNN techniques.

Attempts are made to analyze the image in various ways using the CNN structure, and for this purpose, the CNN is trained using the previously collected purified image data set (S460).

Typically, refined image data sets are used for experiments in the field of computer vision, including PASCAL Visual Object Classes, MIT Indoor Scene, ImgaeNet, and Caltech 101.

In this study, we train a CNN with a dataset labeled on an image and then perform a classification experiment on the same dataset.

Experiments in the study show different results depending on the data sets used for training and classification.

Experimental results with PASCAL's image set showed an average image auto tagging accuracy of 72.6%, and an average of 59.6% accuracy with the MIT Indoor image set.

This is clearly a better result than the tagging method.

In the above experiment, even though the test images are randomly extracted from the training image set, they are artificially collected and refined data. Therefore, it is necessary to confirm what results will be obtained by targeting images stored by users on a real-time rapidly changing Internet sharing site. There was.

In the present invention, by implementing the CNN technique, the automatic tagging and classification experiments were performed on Instagram images that are currently in the spotlight as social image sites.

CNN consists of five Convolution, max-pooling, normalization layers and three fully-connected layers.

In the convolution layer, multiple convolution, max-pooling, and normalixation information overlaps the information of the image. This enables the image to be invariably learned. It is showing.

When connected to the input image Couvolution layer of N * N and there is a convolution filter of m * m, the size of the convolution layer is (N-m + 1) * (N-m +1)

Each convolution layer receives yij (l) input from the output of the previous layer.

In order to extract accurate and meaningful features, we train a convolution filter that is weighted through error backpropagation and gradual computation.

Full-connected will be performed at the top three layers.

Continuous training and features are extracted from the FC6 and FC7 tiers and finally the Softmax classification.

CNN excels at using convolution filters on images, allowing for fast calculations and the ability to consider not only certain objects but the entire image, including the background.

Recently, various studies have been developed for CNN-based image recognition and classification.

While using CNN's experimental image data set to demonstrate CNN's performance, CNN-based automatic is used for daily images uploaded directly by users using Instagram, a social media sharing site that has recently seen a surge in users. We tried to verify that tagging was effective.

As a result, it is possible to obtain tags with higher accuracy and detail than tags registered by users, as well as tags by SIFT-based BoVW.

As we enter the era of big data, research using deep learning, which uses huge amounts of data to learn and classify on its own, continues to perform well.

In the field of image recognition, CNN is currently evolving to keep up with human judgment.

Next, image dispersion / parallel processing techniques and algorithms will be described.

IBM developed Hadoop-based distributed / parallel processing system to develop a system for large-scale image retrieval.

The University of Virginia is developing the Hadoop Image Processing Interface for image-based map-reduce Tasks (HIPI) framework, a Hadoop MapReduce library that provides an API (Applicati On Program Interface) for distributed computing for image big data.

HIPI is based on the MapReduce framework and provides an open and extensible library for image processing and vision applications.

It supports image-based distributed / parallel processing without the user having to know the details of MapReduce framework.

The Google system performs parallel processing on 1,000 machines with 16 cores to learn 1B parameters for 10 million 200 * 200 images.

In addition, mobile device image analysis algorithms are being accelerated in real time. As the IoT environment arrives, the number of video devices for security and methods is expected to increase rapidly.

Video devices-> Low-power embedded systems that incorporate image sensors, processors, and wireless modems. Smartphones, IP cameras, and smart cameras are examples.

Because it has a built-in processor that performs worse than a PC, parallelization and acceleration are required for fast processing of high-demand applications such as video image analysis algorithms based on pixel-based computation.

For this purpose, SIMD-based video image analysis algorithm can be parallelized. The SIMD data processing method is based on background difference-based video image analysis algorithm-> pixel-based module. Increases too.

Therefore, parallelization and acceleration are required for real-time operation in a mobile environment.

Fast processing of data in a mobile environment can be met through parallel programming.

The data processing speed is accelerated by using a SIMD, which is a data parallel processing method, to process a plurality of data in a single instruction (S470).

We can use background difference based video image analysis algorithm,

Video-> RGB to Gray-> Background-> Foreground-> Morphology-> Labeling Seq conversion Generation Extraction Cperation

   -> Object

Tracking

Description: RGB to Gray conversion-> Roreground Extreaction

Object Tracking-> Background Generation

It compares the SISD processing a single data with a single command and the way of processing a plurality of data of the SIMD.

SISD processes a plurality of data individually to be processed by the same type of instruction, while SIMD vectorizes a plurality of data to process a plurality of data simultaneously in a single instruction.

Thus, SIMD-based data parallelism can be applied to modules with shading where adjacent pixels are not data dependent and must be processed with the same type of instruction.

In the SIMD-based video image analysis algorithm, the parallelization method sequentially accesses all the pixel data in the SISD and executes the same instruction repeatedly.However, in the case of the SIMD, the adjacent pixel data waiting for the same instruction is vectorized into a single instruction. A plurality of pixel data is processed simultaneously.

Therefore, using SIMD parallelism can reduce the number of CPU cycles required to process the same data as compared to using SISD, thereby accelerating the processing speed of the background difference-based video image analysis algorithm.

It also saves energy by allowing more data to be processed while using the same number of CPU cycles.

The ARM SIMD (NEON) model, which is supported for the Cortex-A series CPUs, is used to parallelize the data processing of the background difference-based video image analysis algorithm.

Background Two experiments were conducted to verify the performance improvement compared to the existing C-based SISD implementation due to the NEON parallelization of the difference-based video image analysis algorithm.

The first experiment is to verify the processing speed acceleration compared to the SISD implementation. The processing time of the modules with NEON parallelism is measured and compared.

The second experiment is to verify the energy saving compared to the implementation of SISD, and the amount of energy consumed when performing the background difference-based video image analysis algorithm is measured and compared with Monsoon power meter.

All experiments were conducted using Samsung's Galaxy S4 with a Cortex-A15 CPU, and the image resolution was considered CIF (320 * 240) D1 (720 * 480).

When using SISD and NEON, the video image analysis algorithm shows the processing time for each module by parallel module.

Overall is the result of total processing time including all modules without parallelism.

The processing speed of each module was accelerated by 3.36-26.96 times for the two resolutions, and the overall processing speed was accelerated by 5.34 times at the resolution D1.

This is because NEON parallelism, which can process multiple data simultaneously, reduces the number of CPU cycles required to process the same data.

Applicable for real time operation of low power mobile video devices.

In the present invention, parallelization is performed using SIMD, which is a data parallel processing method that processes a plurality of data in a single instruction for low power and real time operation of a background difference based video image analysis algorithm in a mobile environment.

Pixel-based independent operations such as color space conversion, background generation, foreground extraction, and morphology operations have been parallelized, and Android based has been implemented on mobile platforms.

Compared with the existing C-based SISD implementation, the overall processing speed of the video image analysis algorithm is accelerated by about 5.34 times and energy consumption is reduced by about 4.08 times for the D1 class video.

On the other hand, by analyzing the information of the image using the intelligent image analysis technology to detect abnormal behavior automatically and transmit the information to the administrator as follows.

There is a limit to preventing intrusions, so it is necessary to catch intrusions before theft.

It is necessary to find ships, intruders (fishes, divers), buoys, and estimate their location to detect intrusions in and around field boundaries.

It is difficult to detect theft of invaders in the field farms, so in order to detect theft, ships, intruders (murcoons), and buoys must be located within and outside the boundaries of the field farms. Detection using the quantity and the like (S480).

When theft occurs due to the invasion of field farms in the image, theft is detected by detecting the change of the whole screen or by using the scene change detection method of the image.

In order to steal, at night or day, a ship near a field farm can be used to dive an invader (mercury) into the sea and invade the field farm. Turn it over.

If theft is found, the buoys can be chopped off and tied to a string of seafood to try to destroy the evidence.

In this case, to steal in the field farm, the presence or absence of the number of buoys is generated, and the intruder's diving and raising of the hand on the buoy are detected.

Finally, using the schematic solution, the probability of theft is obtained for each frame and detected as theft by more than a certain probability. And according to the detection result, the unmanned aerial vehicle is more than a certain probability to the administrator's terminal (smartphone, etc.). The corresponding information is transmitted (S490). In such a case, tourists such as general skin scuba are not detected as thieves, so automatic information transmission is not reported, and the administrator also does not receive unnecessary information.

Approaches are hierarchical, descriptive methods of behavior recognition, such as quantity, location changes, and repetitive moving behaviors, to use diving and buoys as features of theft.

It is necessary to accelerate the development of deep learning technology using artificial neural networks for machine learning in artificial intelligence.

The development trend of deep learning based image processing technology will be explained.

The first visual cognitive technology refers to the technology used for object recognition, computer vision, behavior and situation understanding, image knowledge processing and video search technology.

The second spatial cognitive technique can be used to analyze distances and depths in 2D and 3D images, and to understand or recognize objects in space.

The third story compression and creation technique is used to understand the story from the video, to separate the object of object, to summarize the meaningful content from the video, and to automatically edit the video.

The machine learning technology can be used for statistical processing of objects in an image, clustering between objects, non-supervised learning for machine learning, and reinforcement learning for image processing.

The first is CNN, which is based on supervised learning.

Visual cognitive technology, spatial cognitive technology (distance and depth analysis in 2D and 3D images, understanding or recognizing objects in space), story compression and creation technology (understand storage from images, distinguish objects from objects, and mean from images Technology that summarizes the existing content and automatically edits the image), machine learning technology (statistical processing of objects in the image, clustering between objects, non-supervised learning reinforcement learning for image processing), etc. This is a method using pre-trained CNN.

It is applied in the form of detecting things in data sets other than ImageNet using CNN trained with ImageN data for classifying things in images.

When the detection method is applied to a video having a frame-based time axis, the feature of the image is extracted every frame of the video, and the averaged value is used as the feature value or the feature value of the image and optical flow is recognized for each frame. It is used in the method and averaged the resultant value.

Another method is to process 3D convolution, which interprets the video as three-dimensional data accumulated along the time axis instead of two-dimensional image data.

This method performs convolutional operations on the time base with kernels of constant time intervals (eg 5 frame intervals), because useful patterns between adjacent frames can appear anywhere in the beginning, middle, and end of a frame. Done.

The problem is that there are more parameters to learn and more data to learn.

Recently, in order to compensate for the lack of a large amount of data, a method of analyzing a video by applying temporal pooling to the time series of each feature value after converting the feature value of the CNN extracted for each frame into a time series form is proposed.

Second, there are Stacked autoencoders.

One way is to use Stacked autoencoders, which are stacked in layers of autoencoders that aim to equalize input and output data.

Third, there is RDNN / LSTM.

It is a cyclic deep learning method using RDNN (Recurrent Deep Neural Network) or Long-Short Term Memory (LSTM) for image processing including time axis like video.

This method combines learning and inference using the short-term and long-term memory model (LSTM), which is a kind of cyclic neural network, in which the temporal association between the recognition results of the feature patterns coming out is output.

Fourth, there is a combination of CNN and existing feature extraction methods.

Using multi-scale CNN, each pixel of the image is roughly identified as a part of an object, the determined pixels are clustered in units of super pixels, and the super pixels are connected to form a segment graph. Using the (multi-1evel cut) algorithm, the division combination that best matches the CNN's discriminant results is selected and used to analyze the entire scene image.

Another method is R-CNN (regions with CNN feature) that extracts candidate regions that are supposed to exist by using computer vision technology and then uses CNN for each candidate region to determine whether they are real objects and perform recognition. There is.

In addition, we combine the CNN structure with the learning method of the extreme learning machine to improve the learning speed and use it to improve lane detection.

Extreme learning machine is a learning method proposed by Tsinghua University. Because the error propagation method of backward propagation method has a slow learning speed, there is a learning method that calculates the optimal output weight by one operation using pseudo inverse.

Although the present invention has been described by way of example as described above, the present invention is not necessarily limited to these examples, and various modifications can be made without departing from the spirit of the present invention. Therefore, the examples disclosed in the present invention are not intended to limit the technical spirit of the present invention but to describe the present invention, and the scope of the technical spirit of the present invention is not limited by these examples. The protection scope of the present invention should be interpreted by the following claims, and all technical ideas within the scope equivalent thereto should be construed as being included in the scope of the present invention.

100: unmanned aerial vehicle 110: communication unit
120: sensor unit 130: flight unit
140: camera unit 150: image conversion unit
160: image storage unit 170: GPS receiver
180: illegal activity identification unit 190: warning broadcasting unit
200: signal transmission unit 210: driving distance calculation unit
220: control unit 300: drone regulator
310: communication unit 320: flight control unit
330 display unit 340 camera control unit
350: voice input unit 360: alarm generating unit
370: ship tracking selector 380: operation area setting unit
390: control unit
400, 410, 420, 430: Field farm 500, 510: Smart device
600: field farm buoy 610: buoy fixing device
620: Mug 630: diver
700: maritime radar 800: communication network
900: monitoring server 910: communication unit
920: database 930: driving information storage unit
940: operation information storage unit 950: unmanned aerial vehicle management request unit
960: Cost calculation and billing unit 970: Control unit
1000: financial company server

Claims (4)

Receives a control signal for controlling the unmanned aerial vehicle 100 from the unmanned aerial vehicle controller 300, receives a ship detection radar signal from the marine radar 700, and the image data and position information photographed by the unmanned aerial vehicle 100 A communication unit 110 for transmitting the unmanned aerial vehicle adjuster 300 and the field farm official or the unmanned aerial vehicle operator's smart device 500;
A sensor unit 120 for sensing a wind direction, a wind speed, and a current altitude according to the moving direction of the unmanned aerial vehicle 100, measuring a rotational movement state, and detecting a tilt;
Flight unit 130 for the flight of the unmanned aerial vehicle 100;
A camera unit 140 for photographing the surroundings while the unmanned aerial vehicle 100 moves;
An image storage unit 160 for storing an image photographed by the camera unit 1400;
A GPS receiver 170 for receiving a current position of the unmanned aerial vehicle 100;
An illegal activity identification unit 180 for identifying a vessel in an image captured while the unmanned aerial vehicle 100 moves;
A warning broadcast unit 190 for outputting a warning broadcast or warning sound previously input or outputting a voice or warning broadcast transmitted from the drone controller 300;
The communication unit 110, the sensor unit 120. By controlling the unmanned aerial vehicle 100 including a flight unit 130, a camera unit 140, an image storage unit 160, a GPS receiver 170, illegal activity identification unit 180, warning warning unit 190 Control the flight unit 130 according to the control data transmitted from the unmanned vehicle adjuster 300 through the communication unit 110, illegal activity identification unit 180 in the image captured by the camera unit 140 When the identification of the vessel is transmitted to the smart device 500 of the unmanned vehicle adjuster 300, field farm personnel or unmanned aerial vehicle operator through the communication unit 110, the voice from the data transmitted from the unmanned vehicle adjuster 300 The voice signal is output through the warning broadcast unit 190, and the location information received from the GPS receiver 170 and the driving area set by the drone adjuster 300 are calculated to deviate from a preset range. Through the communication unit 110 The drone adjuster 300 notifies this, calculates the remaining battery, calculates the flight time and the return time of the unmanned aerial vehicle 100, and calculates the available flight time to the place to the drone adjuster 300. Control unit 220 to control to control; so that the unmanned aerial vehicle capable of monitoring illegal activities of the field farm.
In the illegal activity monitoring method using an unmanned aerial vehicle capable of monitoring the illegal activity of field farms according to claim 1,
The unmanned aerial vehicle is powered on (S100), while the unmanned aerial vehicle is operated (S110), receiving image information through an unmanned aerial vehicle (S120);
If an event occurs during the operation of the unmanned aerial vehicle, when suspicious vessel information is received according to the event, the GPS information is moved to the corresponding position using the GPS information (S140), the image is taken (S150), and the event is the unmanned aerial vehicle 100. If detected through the camera unit 140 of the illegal acts suspected to shoot the ship (S150), the captured image is stored in the unmanned vehicle, and transmitting it to the outside through the communication unit of the unmanned aerial vehicle ( S160); And
Transmitting a voice through the unmanned aerial vehicle controller 300 for adjusting the unmanned aerial vehicle or outputting a preset warning broadcast or warning sound to the warning broadcasting unit of the unmanned aerial vehicle (S170); Method of monitoring illegal activities using unmanned aerial vehicle capable of monitoring illegal activities.
The method of claim 2,
The operation of the unmanned aerial vehicle is a normal flight by an unmanned aerial vehicle manager or a person in charge of a field farm, or when a vessel radar has received ship information on a field farm, or a vessel access signal is sent from a buoy installed in a field farm. A method for monitoring illegal activities using an unmanned aerial vehicle capable of monitoring illegal activities of field farms, which is performed when received through a related party's smart device (smartphone, smart pad) or a predetermined receiver.
A surveillance method according to the analysis and recognition of torts using artificial intelligence deep learning of an unmanned aerial vehicle capable of monitoring the illegal activities of field farms according to claim 1,
The unmanned aerial vehicle is configured to data the number of theft and buoy buoys of the field farm to be monitored and stealing activities including a ship, an intruder, a seafood network, an intruder's dive, and an act of raising a hand over a buoy (S400);
The unmanned aerial vehicle generates a background area model by using the previous image of the field farm, and removes the image from the original image to identify the object area and the background area (S410);
The unmanned aerial vehicle detects at least one of a shelf, an intruder, a buoy, a size, a field farm boundary, a location, a distance, a slope, a color, a text, a sign in real time (S420);
The unmanned aerial vehicle detects a person by extracting a motion region from two frames of an incoming image in order (S430) (S440);
The unmanned aerial vehicle removes a false image of the detected intruder region (S450);
The unmanned aerial vehicle performs deep learning on the collected image (S470); And
The unmanned aerial vehicle detects abnormal behavior by using a change of a screen or a scene change detection method of an image through the deep learning on the collected image information, and transmits the abnormal behavior to a manager terminal (S480) (S490). Monitoring method according to the analysis and recognition of illegal activities using artificial intelligence deep learning of unmanned aerial vehicles capable of monitoring illegal activities of field farms.

Images