Learning information generating device, learning device, detection device, and learning information generating method

Abstract

A learning information generation device is provided that automatically collects learning information and reduces the load of acquiring learning information. The learning information generation device according to the present disclosure includes an image receiving unit (1) that receives monitoring images of a monitoring area, a signal acquiring unit (2) that acquires a reference signal indicating that the monitoring area is in a reference state, and an image acquiring unit (3) that, in response to the signal acquiring unit acquiring the reference signal, acquires, as a learning image, from the monitoring images received by the image receiving unit, the monitoring image at the time when the signal acquiring unit acquired the reference signal.

Description

[Technical Field]

[0001] The present disclosure relates to a training information generation device, a learning device, a detection device, and a training information generation method. [Background technology]

[0002] Patent Document 1 discloses an intruder detection system that can detect whether an abnormal state has occurred even when the background image changes. [Prior art documents] [Patent documents]

[0003] [Patent Document 1] Japanese Patent Publication No. 2023-168733 Summary of the Invention [Problem to be solved by the invention]

[0004] The intruder detection system described in Patent Document 1 has multiple trained models corresponding to small areas, seasons, and time periods into which surveillance video is divided, and therefore has the problem of requiring a huge amount of training data.

[0005] Therefore, an object of the present disclosure is to reduce the burden of acquiring learning information. [Means for solving the problem]

[0006] The learning information generation device according to the present disclosure includes: This is an area that includes the station platform. an image receiving unit that receives a monitoring image of a monitoring area; A signal indicating the closed status of the platform doors installed on the platform, The monitoring area No people or objects can enter a signal acquisition unit that acquires a reference signal indicating that the state is in the reference state; as In response to the acquisition, an image acquisition unit acquires, as a learning image, the monitoring image when the signal acquisition unit acquires the reference signal from among the monitoring images received by the image receiving unit. [Effects of the Invention]

[0007] According to the present disclosure, the load of acquiring learning information can be reduced by automatically collecting learning information using a reference signal indicating that the monitoring area is in a reference state as a trigger for acquiring learning information. [Brief explanation of the drawings]

[0008] [Figure 1] 1 is a functional configuration diagram of a learning information generation device and a learning device according to Embodiment 1. FIG. [Figure 2] 2 is an example of a monitoring area according to the first embodiment. [Figure 3] 1 is an example of input and output of a learning model including an autoencoder according to the first embodiment. [Figure 4] 1 is a functional configuration diagram of a detection device according to a first embodiment. [Figure 5] 1 is a hardware configuration diagram of a learning information generation device, a learning device, and a detection device according to a first embodiment. [Figure 6] 4 is a flowchart showing an example of processing performed by the learning information generation device and the learning device according to the first embodiment. [Figure 7] 4 is a flowchart showing an example of processing performed by the detection device according to the first embodiment. [Figure 8] 10 is a modified example of the learning information generating device and learning device according to the first embodiment. [Figure 9] 10 is a modified example of the detection device according to the first embodiment. [Figure 10] 10 shows a modified example of a monitored area and an opening / closing device according to the first embodiment. [Figure 11] FIG. 10 is a functional configuration diagram of a learning information generating device and a learning device according to a second embodiment. [Figure 12] 10 is an example of a mask region according to the second embodiment. [Figure 13] FIG. 10 is a functional configuration diagram of a detection device according to a second embodiment. [Figure 14]10 is a hardware configuration diagram of a training information generation device, a learning device, and a detection device according to a second embodiment. FIG. [Figure 15] 10 is a flowchart showing an example of processing performed by the learning information generation device and the learning device according to the second embodiment. [Figure 16] 10 is a flowchart illustrating an example of processing performed by the detection device according to the second embodiment. [Figure 17] 10 is a modified example of a mask region according to the second embodiment. [Figure 18] FIG. 11 is a functional configuration diagram of a learning information generation device and a learning device according to a third embodiment. [Figure 19] 10 is an example of a mask region according to the third embodiment. [Figure 20] FIG. 11 is a hardware configuration diagram of a training information generation device, a learning device, and a detection device according to a third embodiment. [Figure 21] 11 is a flowchart showing an example of processing performed by the learning information generation device and the learning device according to the third embodiment. DETAILED DESCRIPTION OF THE INVENTION

[0009] Embodiment 1 The training information generating device 100, the training device 101, and the detection device 102 will be described with reference to FIGS. In the embodiments and drawings, the same or corresponding elements are denoted by the same reference numerals. The description of elements denoted by the same reference numerals as those already described will be omitted or simplified as appropriate. Arrows in the drawings primarily indicate the flow of data or the flow of processing.

[0010] ***Outline of the First Embodiment*** <<<Study time>>> <<Functions of the Learning Information Generation Device 100>> First, the functional outline of the learning information generation device 100 according to this embodiment will be described. The learning information generation device 100 has a function of acquiring a reference signal indicating that the monitoring area 40 is in the reference state in order to acquire an image of the monitoring area 40 in the reference state as the learning image 51. When the monitoring area 40 is the area between the platform door 30a installed on the station platform 41a and the tracks 41b, the learning information generation device 100 has a function of acquiring a signal including the open / closed state from the platform door 30a in order to acquire an image of the platform door 30a closed (a state in which no people or objects are entering) as learning information. When the learning information generation device 100 receives a signal indicating that the platform doors 30a are closed (hereinafter referred to as the closed state), it acquires the camera image as the learning image 51. This makes it possible to collect an image in a state where no people or objects are present. It is also possible to reduce the load of collecting learning information.

[0011] <<Configuration of Learning Information Generation Device 100>> In order to perform the above functions, the learning information generation device 100 includes an image receiving unit 1, a signal acquiring unit 2, and an image acquiring unit 3. The image receiving unit 1 acquires an image output by a monitoring device 20 such as a camera. The signal acquisition unit 2 acquires a status signal 60 (opening / closing signal) output by an opening / closing device 30 such as a platform door 30a. When the opening / closing device 30 is in the reference state (closed state), the signal acquisition unit 2 outputs an acquisition signal 61 to the image acquisition unit 3. The image acquisition unit 3 acquires, as a learning image 51, an image when the acquisition signal 61 is input from the signal acquisition unit 2.

[0012] <<Configuration of learning device 101>> The learning device 101 includes a learning information generation device 100 and a learning unit 4 . The learning unit 4 trains a learning model 5 (obstacle detection artificial intelligence) using learning images 51. Note that artificial intelligence will be referred to as AI hereinafter. Learning model 5 (obstacle detection AI) is an artificial intelligence that inputs an image and trains it to minimize the difference between the input image and the output image.

[0013] <<<When analyzing>>> <<Functions of the detection device 102>> The detection device 102 of this embodiment will now be described functionally. During analysis, the surveillance image 50 is input into the learning model 5 (obstacle detection AI), and the output image is compared with the original video in the analysis unit 11. If the difference region is abnormal to the threshold, the difference region is detected as an obstacle 43. This disclosure does not mention how to control platform doors or trains based on the detection result 71.

[0014] <<Configuration of the detection device 102>> To perform the above functions, the detection device 102 includes an inference unit 10 and an analysis unit 11 that use a learning model 5 (obstacle detection AI). The inference unit 10 performs processing for detecting an obstacle 43 using the learning model 5 (obstacle detection AI). Learning model 5 (obstacle detection AI) is an AI that is trained by inputting images so that the difference between the input image and the output image is small. During training, images that do not show obstacles 43 are learned, so if an obstacle 43 is shown in the input image during analysis, a difference will occur between the input image and the output image. The analysis unit 11 calculates the difference between the input image and the output image, and if there is a difference, outputs the detection result 71 as an obstacle 43 .

[0015] <<<Terminology>>> <<Definition of Monitoring Area 40>> The monitoring area 40 is an area in which an unspecified number of moving objects move. The monitoring area 40 refers to an area where people or moving objects come and go (pass or get on and off).

[0016] <<Definition of Switchgear 30>> The opening and closing device 30 is a movement prohibition device that prohibits a moving object from moving within the monitoring area 40. The opening and closing device 30 is a traffic prohibition device that prohibits people or moving objects from passing through.

[0017] <<Definition of Obstacle 43>> The obstacle 43 refers to an object that may impede the movement of moving objects in the monitoring area 40. Obstacles 43 include lost items, forgotten items, suspicious items, waste, and the like.

[0018] <<Definition of reference condition>> The reference state refers to a state in which no moving object exists in the monitoring area 40. The reference state refers to a state in which no moving object is moving in the monitoring area 40.

[0019] ***Configuration Description*** The configuration of a learning device 101 equipped with the learning information generation device 100 will be described. <<<Functional Configuration of Learning Information Generation Device 100>>> First, the learning information generation device 100 will be described with reference to FIG. The learning information generation device 100 includes an image receiving unit 1, a signal acquiring unit 2, and an image acquiring unit 3. The learning information generation device 100 includes a signal storage unit (not shown). Information for identifying a reference signal is stored in advance in the signal storage unit as reference signal identification information. The signal storage unit may store reference signal identification information corresponding to a plurality of reference signals.

[0020] <<Image receiving unit 1>> The image receiving unit 1 receives a monitoring image 50 of a monitoring area 40 . The image receiving unit 1 receives a monitoring image 50 from a monitoring device 20 . The image receiving unit 1 receives, as a monitoring image 50 of the monitoring area 40, an image of the area around the opening and closing device 30 that blocks the movement of people or moving objects in the monitoring area 40. In this embodiment, the image receiving unit 1 receives the monitoring image 50 of the monitoring area 40, which is an area including a station platform 41a where platform doors 30a are installed.

[0021] The image receiving unit 1 receives the monitoring image 50 at a predetermined time interval. The predetermined time interval can be set arbitrarily. The image receiving unit 1 may receive the monitoring image 50 within a predetermined time period. The predetermined time period is a time period during which monitoring is deemed necessary, and can be set arbitrarily.

[0022] The monitoring area 40 in this embodiment will be specifically described based on the arrangement of the platform 41a, platform doors 30a, monitoring device 20, and tracks 41b shown in FIG. The monitoring area 40 is a three-dimensional area between the platform 41a and the platform 41a, sandwiched between the platform doors 30a and the tracks 41b. The monitoring device 20 is installed at the platform door 30a so as to be able to photograph the monitoring area 40.

[0023] The monitoring device 20 captures a monitoring image 50 that includes the monitoring area 40. The monitoring device 20 is a two-dimensional optical video camera that captures moving images. The monitoring device 20 is installed on the platform door 30a so that the angle of view and the shooting angle are fixed.

[0024] The monitoring image 50 is an image or video captured by the monitoring device 20. The monitoring image 50 is an image that includes the monitoring area 40 . The monitoring image 50 may be an image including the monitoring area 40, a background 42 including the tracks 41b, and the platform doors 30a. In this embodiment, the monitoring image 50 is assumed to be an image including the platform 41a, the platform doors 30a, and the background 42.

[0025] <<Signal Acquisition Unit 2>> As shown in FIG. 1, the signal acquiring unit 2 acquires a reference signal indicating that the monitoring area 40 is in a reference state. The reference state refers to a state in which the monitoring area 40 is normal. In this embodiment, the reference state is a state in which no people or objects have entered the monitoring area 40. In this embodiment, the closed state of the platform doors 30a is regarded as the reference state in which no people or objects have entered the monitoring area 40.

[0026] The signal acquiring unit 2 acquires a signal indicating the closed state of the opening and closing device 30 as a reference signal. In this embodiment, the signal acquisition unit 2 acquires, as a reference signal, a signal indicating the closed state of the platform doors 30a installed on the platform 41a. The signal acquisition unit 2 includes a state signal acquisition unit 2a and a signal determination unit 2b.

[0027] The status signal acquisition unit 2 a acquires a status signal 60 that indicates the status of the monitoring area 40 . The status signal 60 includes a reference signal indicating that the monitoring area 40 is in a reference state. The status signal 60 may be a signal indicating that a person or object has entered the monitoring area 40. A specific example in this embodiment is a signal indicating that the platform doors 30a are open (hereinafter, referred to as the open state). The status signal 60 may be a signal indicating an abnormality in the monitoring area 40. A specific example in this embodiment is a signal that is transmitted when an emergency button provided on the platform door 30a is pressed.

[0028] The signal determination unit 2b determines whether the status signal 60 acquired by the status signal acquisition unit 2a is a reference signal. The signal determination unit 2b determines whether the status signal 60 acquired by the status signal acquisition unit 2a is a reference signal based on whether the status signal 60 includes the reference signal identification information stored in the signal storage unit. When the signal determination unit 2b determines that the status signal 60 is a reference signal, it transmits an acquisition signal 61 to the image acquisition unit 3. The acquisition signal 61 is a signal that prompts the image acquisition unit 3 to acquire the learning image 51 .

[0029] <<Image Acquisition Unit 3>> The image acquisition unit 3 acquires the learning image 51 in response to the signal acquisition unit 2 acquiring the reference signal. Specifically, the image acquisition unit 3 receives an acquisition signal 61 from the signal determination unit 2b. Upon receiving the acquisition signal 61, the image acquisition unit 3 acquires, as a learning image 51, the monitoring image 50 received by the image receiving unit 1 when the signal acquisition unit 2 acquired the reference signal. The image acquisition unit 3 acquires the learning image 51 when the signal determination unit 2b determines that the signal acquisition unit 2 has acquired the reference signal.

[0030] <<<Functional configuration of learning device 101>>> Next, the configuration of the learning device 101 will be described with reference to FIG. The learning device 101 includes a learning information generation device 100 and a learning unit 4 .

[0031] <<Study Section 4>> The learning unit 4 trains the learning model 5 (obstacle detection AI) based on the learning images 51 acquired by the learning information generation device 100. That is, the learning unit 4 uses the learning image 51 to make the learning model 5 learn the reference state of the monitoring area 40, and creates a learned learning model 5.

[0032] The learning model 5 is a model for inferring whether or not the monitoring area 40 is in a reference state. The learning model 5 is a model that includes an autoencoder 5a.

[0033] When an input image is input, the autoencoder 5a extracts the feature amount of the input image, and outputs the output result 70 in which the input image is reconstructed based on the feature amount. The autoencoder 5a learns so that the output result 70 approaches the input image.

[0034] The autoencoder 5a used in this embodiment will be outlined below with reference to (a) and (b) of FIG. The autoencoder 5a is preferably a convolutional autoencoder using a convolutional neural network (hereinafter referred to as CNN). As shown in FIG. 3(a), the learning unit 4 causes the autoencoder 5a to learn using a plurality of normal images (learning images 51 in a state where there are no obstacles). The autoencoder 5a receives and outputs normal images. The autoencoder 5a includes an encoder and a decoder. The encoder extracts features from the input image by encoding it, and the extracted features are then sent to the decoder. The decoder decodes the extracted feature amount to generate an output result 70, which is an output image. The autoencoder 5a learns so that the difference between the input image and the output result 70 becomes smaller. As shown in (b) of Figure 3, when an autoencoder 5a that has learned normal images has an object (obstacle 43) in the input image, the object disappears or changes shape in the output result 70. The difference between the input image of the autoencoder 5a and the output result 70 becomes large, making it possible to detect the presence or absence of an obstacle.

[0035] The learning of the autoencoder 5a in the learning unit 4 will be described with reference to FIG. The learning unit 4 inputs a learning image 51 to the encoder as an input image. The learning image 51 is assumed to be an image including a platform 41a, platform doors 30a, and a background . The encoder extracts features by encoding the learning images 51. The decoder decodes the feature amount extracted by the encoder, and outputs the output result 70 that restores the learning image 51. The autoencoder 5a learns the reference state in the monitoring image 50 by learning so that the difference between the learning image 51, which is the input image, and the output result 70 becomes small.

[0036] <<<Functional configuration of the detection device 102>>> Next, the configuration of the detection device 102 using the learning model 5 learned by the learning device 101 will be described with reference to FIG. The detection device 102 comprises an image receiving unit 1 , an inference unit 10 and an analysis unit 11 . It is desirable that the detection device 102 detects the obstacle 43 when the platform doors 30a are in the closed state. For this reason, the detection device 102 in this embodiment includes an image acquisition unit 3 and a signal acquisition unit 2. The detection device 102 includes a signal storage unit (not shown). The signal storage unit corresponds to the signal storage unit included in the learning information generation device 100, and operates in the same manner as the signal storage unit included in the learning information generation device 100.

[0037] <<Image receiving unit 1>> The image receiving unit 1 corresponds to the image receiving unit 1 included in the learning information generation device 100. The monitoring image 50 received by the image receiving unit 1 is preferably an image taken in the same monitoring area 40 as the learning image 51 used to train the learning model 5. The monitoring image 50 received by the image receiving unit 1 is preferably an image captured at the same angle of view and shooting angle as the learning image 51 used to train the learning model 5.

[0038] <<Signal Acquisition Unit 2>> The signal acquiring unit 2 corresponds to the signal acquiring unit 2 included in the learning information generating device 100 and operates in the same manner as the signal acquiring unit 2 included in the learning information generating device 100 . <<Image Acquisition Unit 3>> The image acquisition unit 3 receives the acquisition signal 61 from the signal acquisition unit 2 and acquires, from among the monitoring images 50 received by the image receiving unit 1, the monitoring image 50 when the signal acquisition unit 2 acquired the reference signal.

[0039] <<Inference part 10>> The inference unit 10 inputs the surveillance image 50 acquired by the image acquisition unit 3 into the learning model 5 including the autoencoder 5a that has been trained by the learning device 101, and outputs the output result 70.

[0040] 3(b), the input and output of the autoencoder 5a using the monitoring image 50 in the inference unit 10 will be described. As an example, the monitoring image 50 in which an obstacle 43 is included in the monitoring area 40 will be used for the description. The inference unit 10 inputs the monitoring image 50 to the encoder as an input image. The encoder extracts features by encoding the surveillance image 50 . The decoder decodes the feature amount extracted by the encoder, and outputs the output result 70 that is the restored surveillance image 50. However, since the encoder has learned to extract the features required to restore the training image 51, the encoder cannot fully extract the features required to restore the surveillance image 50. In other words, it is not possible to extract the features of the obstacle 43. Furthermore, the decoder restores the surveillance image 50 from the incompletely extracted feature amount, and therefore outputs the output result 70 that does not reflect the features of the obstacle 43. In other words, a difference occurs between the surveillance image 50, which is the input image, and the output result 70, which is the output image.

[0041] <<Analysis Department 11>> The analysis unit 11 analyzes whether or not the monitoring area 40 is in the reference state based on the output result 70 output by the inference unit 10. The analysis unit 11 detects the difference between the output result 70 output by the inference unit 10 and the monitoring image 50 input to the inference unit 10 . The analysis unit 11 analyzes whether the monitoring area 40 is in a reference state based on the detected difference. Specifically, if the detected difference is equal to or greater than a threshold, it is determined that an obstacle 43 exists in the monitoring area 40. The analysis unit 11 outputs the analysis result as a detection result 71.

[0042] <<<Hardware Configuration of the Training Information Generation Device 100, the Training Device 101, and the Detection Device 102>>> The hardware configurations of the training information generation device 100, the training device 101, and the detection device 102 will be described with reference to FIG. The learning information generating device 100 , the learning device 101 and the detecting device 102 are connected to the monitoring device 20 and the switching device 30 . Each of the training information generation device 100 , the learning device 101 , and the detection device 102 includes a processor 90 , a storage device 91 , a communication device 92 , and an input / output interface 93 . The communication device 92 and the input / output interface 93 acquire the monitoring image 50 from the monitoring device 20 and the status signal 60 from the opening / closing device 30 . The storage device 91 stores programs for realizing the functions of the image receiving unit 1, signal acquiring unit 2, image acquiring unit 3, learning unit 4, inference unit 10, and analysis unit 11, and a learning model 5. The processor 90 is a CPU (Central Processing Unit) that executes arithmetic processing based on a program stored in a storage device 91 .

[0043] ***Explanation of Operation*** <<<Learning Information Generation Device 100, Learning Device 101>>> An example of processing performed by the learning information generation device 100 and the learning device 101 will be described with reference to the flowchart shown in FIG.

[0044] <<Step S101>> The image receiving unit 1 receives a monitoring image 50 . The image receiving unit 1 receives a monitoring image 50 of a monitoring area 40 from a monitoring device 20 .

[0045] <<Step S102>> The signal acquisition unit 2 acquires the status signal 60 . Alternatively, the signal acquiring unit 2 acquires a reference signal indicating that the monitoring area 40 is in a reference state. Specifically, the status signal acquisition unit 2 a acquires a status signal 60 indicating the status of the monitoring area 40 from the opening and closing device 30 . <<Step S103>> Then, the signal determination unit 2b determines whether the status signal 60 acquired by the status signal acquisition unit 2a is a reference signal. In this embodiment, it determines whether the status signal 60 is a signal indicating that the platform doors 30a are in a closed state. If the signal determination unit 2b determines that the status signal 60 is the reference signal, it transmits an acquisition signal 61 to the image acquisition unit 3 and proceeds to step S104. If it determines that the status signal 60 is not the reference signal, it returns to step S101.

[0046] <<Step S104>> The image acquisition unit 3 acquires the learning image 51. When the signal acquiring unit 2 acquires the reference signal, the image acquiring unit 3 acquires, as a learning image 51, the monitoring image 50 at the time the reference signal was acquired from among the monitoring images 50. The image acquisition unit 3 acquires the learning image 51 by receiving the acquisition signal 61 transmitted from the signal determination unit 2b.

[0047] <<Step S105>> The learning unit 4 uses the learning images 51 to train the learning model 5 . The learning unit 4 learns a learning model for inferring whether or not the monitoring area 40 is in a reference state based on the learning image 51 acquired by the image acquisition unit 3 in step S104. Then, the process returns to step S101.

[0048] <<<Detection device 102>>> An example of processing performed by the detection device 102 will be described with reference to the flowchart shown in FIG.

[0049] <<Steps S101 to S103>> In step S101, the image receiving unit 1 performs the same operation as in step S101 in the learning information generation device 100 shown in FIG. In step S102, the signal acquiring unit 2 performs the same operation as in step S102 in the learning information generating device 100 shown in FIG. In step S103, the signal determination unit 2b performs the same operation as in step S103 in the learning information generation device 100 shown in FIG.

[0050] <<Step S104>> In step S104, the image acquisition unit 3 receives the acquisition signal 61 transmitted from the signal acquisition unit 2 in step S103. Also, in step S104, the image acquiring unit 3 acquires, from among the monitoring images 50 received by the image receiving unit 1, the monitoring image 50 when the signal acquiring unit 2 acquired the reference signal. After the processing in step S104, the process proceeds to step S111.

[0051] <<Step S111>> The inference unit 10 inputs the monitoring image 50 received by the image receiving unit 1 to the learning model 5 that the learning device 101 has trained based on the learning image 51. The learning model 5 outputs an output result 70. The inference unit 10 outputs the output result 70 from the learning model 5 only when the state is closed. The inference unit 10 does not operate in any state other than the closed state.

[0052] <<Step S112>> The analysis unit 11 performs analysis based on the output result 70. The analysis unit 11 calculates the difference between the monitoring image 50 input to the learning model 5 by the inference unit 10 and the output result 70 output by the learning model 5, and analyzes whether the monitoring area 40 is in a reference state.

[0053] <<Step S113>> The analysis unit 11 outputs a detection result 71, which is the result of analyzing whether or not the monitoring area 40 is in a reference state. The detection result 71 is output to a management system for the platform doors 30a, a management system for the train 44, a station management system, etc.

[0054] ***Features of the First Embodiment*** The learning information generation device 100 according to the first embodiment does not set the monitored area 40 as a place where entry is prohibited. The learning information generation device 100 according to the first embodiment sets a monitoring area 40 as a place where the movement of moving objects such as people, robots, etc. is intermittent. The learning information generation device 100 according to the first embodiment does not constantly learn the image of the monitored area 40 . The learning information generation device 100 according to the first embodiment learns only images of the monitoring area 40 where no moving object is moving. The detection device 102 of the first embodiment does not detect an intrusion of a suspicious person into a place where intrusion is prohibited. The detection device 102 of the first embodiment detects an obstacle 43 that impedes the movement of a following moving object in a monitoring area 40 where the movement of moving objects is intermittent. The detection device 102 of the first embodiment does not detect the obstacle 43 all the time. The detection device 102 of the first embodiment does not detect the obstacle 43 when the moving object is moving. The detection device 102 of the first embodiment detects the obstacle 43 only when the moving object is not moving.

[0055] ***Effects of the First Embodiment*** As described above, the learning information generation device 100 of embodiment 1 can reduce the load of acquiring learning information by automatically collecting learning information using a reference signal indicating that the monitoring area 40 is in a reference state as a trigger for acquiring learning information.

[0056] Furthermore, a three-dimensional sensor may be mounted on the platform door 30a to detect obstacles in the area on the platform 41a between the platform door 30a and the tracks 41b. Specifically, the image receiving unit 1 receives the image output by the three-dimensional sensor as the monitoring image 50. However, 3D sensors are expensive. In this disclosure, costs can be reduced by replacing the 3D sensor with a camera (2D optical video camera).

[0057] The learning device 101 of the first embodiment can reduce the learning load on the learning model 5 because it is only necessary for the learning model 5 to learn the learning images 51 in which the monitored area 40 is in the reference state.

[0058] ***Supplement to the first embodiment*** A modification of the first embodiment will be described below.

[0059] <<<Modifications Regarding Monitoring Device 20>>> <<When the monitoring device 20 is a camera that takes still images>> In the present embodiment, the monitoring device 20 is a video camera that continuously captures images, but the present invention is not limited to this. The monitoring device 20 may be a camera that takes still images. In this case, as shown in FIG. 8, the signal acquisition unit 2 of the learning information generation device 100 further includes a photography instruction unit 2c.

[0060] When the signal acquiring unit 2 acquires the reference signal, the image capturing instruction unit 2c instructs the monitoring device 20 that captures the monitoring image 50 to capture the monitoring image 50. Specifically, the photographing instruction unit 2c receives the determination result that the status signal 60 is a reference signal from the signal determination unit 2b. Then, the photographing instruction section 2c transmits a photographing instruction signal 62 to the monitoring device 20 to instruct the monitoring device 20 to photograph the monitoring image 50. Upon receiving the photographing instruction signal 62 from the photographing instruction section 2c, the monitoring device 20 photographs a monitoring image 50 in the monitoring area 40 in the reference state.

[0061] The image receiving unit 1 receives a monitoring image 50 captured in response to an instruction from the image capturing instruction unit 2c. The image acquisition unit 3 acquires, as a learning image 51, a monitoring image 50 that has been received by the image receiving unit 1 and taken in response to an instruction from the photography instruction unit 2c. This eliminates the need for the learning information generation device 100 to select learning images 51 from the monitoring images 50, thereby further reducing the load of acquiring learning information. The process in which the signal determination unit 2b transmits the acquisition signal 61 to the image acquisition unit 3 may or may not be performed.

[0062] Next, the learning device 101 and the detection device 102 when the monitoring device 20 is a camera that takes still images will be described. The learning unit 4 of the learning device 101 causes the learning model 5 to learn the learning images 51 captured from the image acquisition unit 3 in response to instructions from the image capture instruction unit 2c.

[0063] As shown in FIG. 9, the signal acquisition unit 2 of the detection device 102 further includes an image capture instruction unit 2c. The photographing instruction unit 2c corresponds to the photographing instruction unit 2c included in the learning information generation device 100. When the signal acquiring unit 2 acquires the reference signal, the photographing instruction unit 2c transmits a photographing instruction signal 62 to the monitoring device 20 that captures the monitoring image 50, instructing the monitoring device 20 to capture the monitoring image 50. The image receiving unit 1 receives a monitoring image 50 captured in response to an instruction from the image capturing instruction unit 2c. The image acquisition unit 3 acquires from the image receiving unit 1 a monitoring image 50 that has been taken in response to an instruction from the photography instruction unit 2c.

[0064] The inference unit 10 inputs the surveillance image 50 taken at the instruction of the shooting instruction unit 2c into the learning model 5 learned from the learning image 51 taken at the instruction of the shooting instruction unit 2c, and outputs the output result 70. The analysis unit 11 analyzes whether the monitoring area 40 is in a reference state based on the output result 70 output by the inference unit 10, and outputs the result as a detection result 71. This eliminates the need for the detection device 102 to select the monitoring image 50, thereby further reducing the load of acquiring the monitoring image 50.

[0065] <<When the monitoring device 20 is not a 2D optical camera>> In the present embodiment, the monitoring device 20 is an optical camera, but the present invention is not limited to this. The monitoring device 20 may be a 3D sensor or an infrared camera, and different types of monitoring devices 20 may be used in combination.

[0066] <<When the monitoring device 20 is installed on the platform 41a other than the platform door 30a>> In the present embodiment, the case where the monitoring device 20 is installed at the platform door 30a provided on the platform 41a of the station has been described, but the present invention is not limited to this. The monitoring device 20 may be installed on equipment, a roof, a pillar, etc. on the platform 41a so as to be able to photograph the platform 41a and the three-dimensional area on the platform 41a sandwiched between the platform door 30a and the tracks 41b. The monitoring device 20 may be mounted on a moving object such as a vehicle, a robot, or a drone. It is desirable that the monitoring device 20 captures images with a constant angle of view and a constant shooting angle, but the angle of view and the shooting angle may vary. In this case, it is desirable to correct the angle of view and the shooting angle of the captured monitoring image 50 so that they are uniform.

[0067] <<<Modifications Regarding Image Acquisition Unit 3>>> <<When the image acquisition unit 3 determines whether the learning image 51 has been acquired>> In this embodiment, when the signal acquisition unit 2 acquires a reference signal, the image acquisition unit 3 acquires the monitoring image 50 at the time the signal acquisition unit 2 acquired the reference signal as the learning image 51, but this is not limited to this. The image acquiring unit 3 may perform a process of determining whether to acquire the learning images 51 when the signal acquiring unit 2 acquires the reference signal. For example, a threshold value for how many learning images 51 are necessary may be set, and if the number of learning images 51 acquired by the image acquiring unit 3 exceeds the threshold value, the image acquiring unit 3 may determine not to acquire the learning images 51.

[0068] <<<Modifications Regarding the Reference Signal Acquired by the Signal Acquirer 2>>> <<When the reference signal is not a signal indicating the closed state of the platform door 30a>> In this embodiment, the reference signal is a signal indicating the closed state of the platform doors 30a, but the present invention is not limited to this. The signal acquisition unit 2 may acquire, as the reference signal, a signal indicating the state from when the train 44 enters the platform 41a, stops, until the doors of the train 44 open.

[0069] <<When the monitoring area 40 and the opening / closing device 30 are other than the platform door 30a>> In the present embodiment, a case has been described in which the image receiving unit 1 receives the monitoring image 50 in which the area near the platform door 30a is the monitoring area 40, but the present invention is not limited to this. The image receiving unit 1 may receive, as a monitoring image 50, an image taken of the area around the opening and closing device 30 other than the platform doors 30a as a monitoring area 40. As shown in FIG. 10, the opening and closing device 30 is, for example, a barrier, an elevator, a ticket gate, a security gate, an automatic door, an automatic revolving door, a vehicle, or the like. When the opening and closing device 30 is a crossing gate and the monitoring area 40 is a railroad crossing, the signal acquiring unit 2 acquires a signal indicating the closed state of the crossing gate as the reference signal. When the opening / closing device 30 is an elevator and the monitoring area 40 is inside the car, the signal acquisition unit 2 acquires, as the reference signal, a signal indicating that a button inside the car has not been pressed for a predetermined time or a signal indicating that there is no one in the car.When the monitoring area 40 is an elevator hall, the signal acquisition unit 2 acquires, as the reference signal, a signal indicating that a hall button has not been pressed for a predetermined time or a signal indicating that there is no one in the elevator hall. When the opening / closing device 30 is a ticket gate and the monitoring area 40 is an entrance of the ticket gate, the signal acquiring unit 2 acquires, as the reference signal, a signal indicating that the door has been in a closed state for a predetermined period of time. When the opening and closing device 30 is a security gate and the monitored area 40 is an entrance of the security gate, the signal acquiring unit 2 acquires, as the reference signal, a signal indicating that the gate has been in a closed state for a predetermined period of time. When the opening and closing device 30 is an automatic door and the monitoring area 40 is the area around the automatic door, the signal acquiring unit 2 acquires, as the reference signal, a signal indicating that the door has not been opened or closed for a predetermined period of time. When the opening and closing device 30 is an automatic revolving door and the monitoring area 40 is inside the automatic revolving door, the signal acquiring unit 2 acquires, as the reference signal, a signal indicating that the door has not rotated for a predetermined period of time. When the opening / closing device 30 is a vehicle and the monitoring area 40 is the interior of the vehicle, the signal acquiring unit 2 acquires, as the reference signal, a signal indicating the closed state of the boarding / alighting door.

[0070] <<<Modifications of Learning Model 5>>> <<Model used in learning model 5>> In this embodiment, the learning model 5 is a model including an autoencoder 5a using CNN, but is not limited to this. The autoencoder 5a included in the learning model 5 does not have to be a convolutional autoencoder. The learning model 5 may use a model other than the autoencoder 5a, or may use another model in combination with the autoencoder 5a. The learning model 5 may be a model that utilizes background subtraction, a technique for comparing a prepared background image with an input image to find something that is not in the background. The learning model 5 may be a model that utilizes a technology for detecting an object based on optical flow that occurs when there is a moving object in an image.

[0071] <<<Modifications of the Detection Device 102>>> <<Monitoring image 50 acquired by inference unit 10>> In the detection device 102 of this embodiment, when a moving object such as a person or a robot is moving, the obstacle 43 is not detected, but the present invention is not limited to this. In the detection device 102 in this embodiment, the image acquisition unit 3 acquires the monitoring image 50 in the reference state where the platform doors 30a are closed from the image receiving unit 1, but the present invention is not limited to this. In the detection device 102 of this embodiment, the inference unit 10 inputs the surveillance image 50 acquired by the image acquisition unit 3 into the trained learning model 5, but the present invention is not limited to this.

[0072] The detection device 102 may detect the obstacle 43 even when the mobile object is moving. The detection device 102 does not necessarily have to include the image acquisition unit 3 . The inference unit 10 may input the monitoring image 50 received by the image receiving unit 1 into the learning model 5 even when the platform doors 30a are in a closed state. In the process performed by the detection device 102 shown in FIG. 7, step S104, which is the process performed by the image acquisition unit 3, may be omitted.

[0073] However, if a monitoring image 50 in which the platform doors 30a are not closed is input into the learning model 5, the accuracy of obstacle detection may decrease. This is because if the inference unit 10 inputs a surveillance image 50 when the platform doors 30a are in an open state into the learning model 5, the detection device 102 may detect a moving object entering or exiting the platform doors 30a as an obstacle 43.

[0074] Therefore, the learning model 5 in this modified example further has a model that, when a moving object is included in the input surveillance image 50, infers whether the moving object is entering or exiting the platform door 30a when it is in the open state. Learning model 5 is a model that not only learns to infer whether the monitoring area 40 is in a reference state, but also learns that moving objects entering and exiting the platform doors 30a when they are in an open state are not obstacles 43 (no abnormalities). The signal storage unit stores in advance status signal identification information for identifying the status signal 60. The reference signal identification information is included in the status signal identification information.

[0075] The detection device 102 in this modified example will be described. The operations of the image receiving unit 1 and the status signal acquiring unit 2a are the same as those of the image receiving unit 1 and the status signal acquiring unit 2a of the detection device 102 described in the first embodiment. The signal determination unit 2b identifies the status signal 60 acquired by the status signal acquisition unit 2a based on the status signal identification information stored in the signal storage unit. When the signal determination unit 2b identifies the status signal 60 as a reference signal, or when it identifies the status signal 60 as a signal indicating the open state of the platform door 30a, the signal determination unit 2b transmits the identification result to the inference unit 10. The inference unit 10 inputs the monitoring image 50 into the learning model 5 when the signal acquisition unit 2 acquires a reference signal or when it acquires a signal indicating that the platform doors 30a are open. The inference unit 10 inputs the monitoring image 50 received by the image receiving unit 1 and the classification result by the signal determining unit 2b into the learning model 5. When a moving object is included in the input surveillance image 50, the learning model 5 infers whether the moving object is entering or exiting the platform doors 30a in the open state. The learning model 5 refers to the identification result by the signal determination unit 2b when inferring whether a moving object is entering or exiting the platform doors 30a in the open state. If the learning model 5 infers that the moving object included in the surveillance image 50 is a moving object entering or exiting the platform door 30a in the open state (no abnormality), the inference unit 10 outputs the inference result that there is no abnormality in addition to the output result 70. The analysis unit 11 detects the difference between the output result 70 output by the inference unit 10 and the monitoring image 50 input to the inference unit 10, and detects an obstacle 43 based on the difference. If the detected difference is equal to or greater than the threshold, and if the learning model 5 infers that there is no abnormality, the analysis unit 11 determines that there is no obstacle.

[0076] In this modification, the detection device 102 does not need to include the image acquisition unit 3 and the signal acquisition unit 2. When a moving object is included in the input surveillance image 50, the learning model 5 infers whether the moving object is entering or exiting the platform door 30a when it is in the open state, without referring to the identification result by the signal determination unit 2b. In the processing performed by the detection device 102 shown in FIG. 7, steps S102, S103, and S104, which are the processing performed by the image acquisition unit 3 and the signal acquisition unit 2, may be omitted.

[0077] <<Detection result 71 of analysis unit 11>> In the present embodiment, the analysis unit 11 outputs as the detection result 71 whether or not the monitoring area 40 is in the reference state (whether or not there is an obstacle 43), but the present invention is not limited to this. The analysis unit 11 may output, as a detection result 71, a difference region between the output result 70 output by the inference unit 10 and the monitoring image 50 used for the inference, as an index indicating the degree of abnormality. The analysis unit 11 may identify the identity of the obstacle 43 and output the identification result as the detection result 71.

[0078] <<<<Modifications of the Configurations of the Training Information Generation Device 100, the Learning Device 101, and the Detection Device 102>>> <<Configuration of Learning Information Generation Device 100>> In the present embodiment, the learning information generation device 100 is provided in the learning device 101, but the present invention is not limited to this. The learning information generation device 100 may not be included in the learning device 101. In this case, after the process of acquiring the learning images 51 in step S104 shown in the example of the process performed by the learning information generation device 100 and the learning device 101 in FIG. 6, the process returns to the process of receiving the monitoring images 50 in step S101. <<Configuration of the detection device 102>> In the present embodiment, the detection device 102 does not include the learning information generation device 100 and the learning device 101, but the present invention is not limited to this. The detection device 102 may be configured to include a training information generation device 100 and a training device 101.

[0079] Embodiment 2 The training information generation device 100, the training device 101, and the detection device 102 will be described with reference to FIGS. In the embodiments and drawings, the same or corresponding elements are denoted by the same reference numerals. The description of elements denoted by the same reference numerals as those already described will be omitted or simplified as appropriate. Arrows in the drawings primarily indicate the flow of data or the flow of processing.

[0080] ***Outline of the Second Embodiment*** First, the present embodiment will be briefly described. Since many obstacles are often on the ground, movement in the background is ignored by setting the learning information generation device 100 to mask everything except the ground area (platform 41a) included in the monitoring image 50 and the learning image 51. Furthermore, since the learning model 5 only needs to learn the area excluding the masked area 80, the acquisition of learning information and the learning processing load can be reduced.

[0081] <<<Study time>>> <<Functions of the Learning Information Generation Device 100>> The functional outline of the learning information generation device 100 of this embodiment will be described below. A mask area setting unit 6 selects a home area for an image acquired from a monitoring device 20 such as a camera, and sets the area outside the selected area as a mask area 80. The set mask area 80 is stored in a mask area storage unit 7, and a mask is applied to the learning image by a mask area application unit 8.

[0082] <<Configuration of Learning Information Generation Device 100>> In order to perform the above functions, the learning information generation device 100 further includes a mask area setting unit 6, a mask area storage unit 7, and a mask area application unit 8. The mask area setting unit 6 detects the area of ​​the platform 41a automatically using area detection or the like, or sets the area other than the area of ​​the platform 41a as a mask area 80 while the user is viewing the input image (monitoring image 50). The mask area storage unit 7 stores the mask area 80 that has been set. The mask area application unit 8 applies the set mask area 80 to the monitoring image 50 .

[0083] <<<When analyzing>>> <<Functions of the detection device 102>> The detection device 102 of this embodiment will now be described functionally. A mask area application unit 8 applies a mask area 80 stored in a mask area storage unit 7 to an image acquired from a monitoring device 20 such as a camera.

[0084] <<Configuration of the detection device 102>> To perform the above functions, the detection device 102 further comprises a mask region storage unit 7 and a mask region application unit 8. The mask region storage unit 7 and the mask region application unit 8 correspond to those included in the learning information generation device 100.

[0085] ***Configuration Description*** <<<Functional Configuration of Learning Information Generation Device 100>>> The learning information generation device 100 will be described with reference to FIG. The learning information generation device 100 differs from the first embodiment in that it further includes a mask area setting unit 6, a mask area storage unit 7, and a mask area application unit 8.

[0086] The image receiving unit 1 receives an image including a station platform 41a as a monitoring image 50. The image acquisition unit 3 acquires, as a learning image 51, an image including a station platform 41a from the monitoring image 50 received by the image receiving unit 1.

[0087] <<Mask Area Setting Unit 6 and Mask Area Storage Unit 7>> The mask area setting unit 6 acquires the monitoring image 50 including the platform 41a from the image receiving unit 1. The mask area setting unit 6 sets the mask area 80 using an area detection AI for automatically detecting an area, called Area Detection.

[0088] The mask region 80 is a region of the monitoring area 40 excluding the ground region, which is at least either the ground surface or the floor surface. In this embodiment, the mask area 80 is an area excluding the platform 41a.

[0089] The area detection AI is a learning model that is trained to detect the platform 41a included in the monitored image 50. The mask area setting unit 6 inputs the monitoring image 50 to the area detection AI, and detects the platform 41a. The mask area setting unit 6 sets the area excluding the platform 41a detected by the area detection AI as a mask area 80. The mask area storage unit 7 stores the mask area 80 set by the mask area setting unit 6.

[0090] <<Mask area application unit 8>> The mask area application unit 8 masks the learning image 51 with a mask area 80 . The mask processing in the mask area application unit 8 will be specifically described with reference to FIG. The mask region application unit 8 acquires the learning image 51 from the image acquisition unit 3 . The mask area application unit 8 reads out the mask area 80 from the mask area storage unit 7 . The mask area application unit 8 masks the learning image 51 in a mask area 80 that is an area excluding the home 41 a, and creates a masked learning image 52.

[0091] <<<Functional configuration of learning device 101>>> The learning device 101 will be described with reference to FIG. This embodiment differs from the first embodiment in that the learning unit 4 of the learning device 101 causes the learning model 5 to learn the masked learning images 52 . The learning unit 4 acquires the masked learning image 52 from the mask region application unit 8 . The learning unit 4 causes the learning model 5 to learn the acquired masked learning images 52.

[0092] <<<Functional configuration of the detection device 102>>> The detection device 102 will be described with reference to FIG. The detection device 102 differs from the first embodiment in that it further includes a mask region storage unit 7 and a mask region application unit 8.

[0093] <<Mask area storage unit 7>> The mask region storage unit 7 corresponds to the mask region storage unit 7 of the device for generating information for learning 100 and operates in the same manner as the mask region storage unit 7 of the device for generating information for learning 100 .

[0094] <<Mask area application unit 8>> The mask region application unit 8 corresponds to the mask region application unit 8 of the device 100 for generating information for learning. The mask area application unit 8 reads out the mask area 80 from the mask area storage unit 7 . The mask area application unit 8 masks the monitoring image 50 acquired from the image acquisition unit 3, to create a masked monitoring image 53 by masking a mask area 80 that is an area excluding the platform 41a.

[0095] <<Inference part 10>> The inference unit 10 inputs the masked surveillance image 53 into the learning model 5 that has been trained by the learning device 101, and outputs the output result 70.

[0096] <<Analysis Department 11>> The analysis unit 11 operates in the same manner as the analysis unit 11 of the detection device 102 in the first embodiment.

[0097] <<<Hardware Configuration of the Training Information Generation Device 100, the Training Device 101, and the Detection Device 102>>> The hardware configurations of the training information generation device 100, the training device 101, and the detection device 102 will be described with reference to FIG. Note that the explanation of the same parts as those in the first embodiment will be omitted. The memory device 91 stores programs for realizing the functions of the image receiving unit 1, signal acquiring unit 2, image acquiring unit 3, learning unit 4, inference unit 10, analysis unit 11, mask area setting unit 6, and mask area application unit 8, as well as a learning model 5 and a set mask area 80.

[0098] ***Explanation of Operation*** <<<Learning Information Generation Device 100, Learning Device 101>>> An example of processing performed by the learning information generation device 100 and the learning device 101 will be described with reference to the flowchart shown in FIG.

[0099] <<Step S101>> The image receiving unit 1 performs the same operation as step S101 in the learning information generation device 100 shown in FIG. After the processing of step S101, the process proceeds to step S201.

[0100] <<Step S201>> The mask area setting unit 6 determines whether the mask area 80 has been set (whether the mask area 80 has been stored in the mask area storage unit 7). If the mask area has already been set in the mask area setting unit 6, the process proceeds to step S102. If the mask area has not been set in the mask area setting unit 6, the process proceeds to step S202.

[0101] <<Step S202>> The mask area setting unit 6 acquires a monitoring image 50 including the platform 41a from the image receiving unit 1, A mask area 80 is set. The mask area 80 is not set constantly by the mask area setting unit 6, but is set only once. The mask area 80 set by the mask area setting unit 6 is stored in the mask area storage unit 7. After the processing of step S202, the process proceeds to step S102.

[0102] <<Step S102>> The state signal acquiring unit 2a of the signal acquiring unit 2 performs the same operation as step S102 in the learning information generating device 100 shown in FIG.

[0103] <<Step S103>> The signal determination unit 2b performs the same operation as step S103 in the learning information generation device 100 shown in FIG.

[0104] <<Step S104>> The image acquisition unit 3 performs the same operation as step S104 in the learning information generation device 100 shown in FIG. After the processing of step S104, the process proceeds to step S203.

[0105] <<Step S203>> The mask area application unit 8 masks the learning image 51. Specifically, the mask area application unit 8 reads out the mask area 80 from the mask area storage unit 7. The mask region application unit 8 masks the mask region 80 in the training image 51 to create a masked training image 52. After the processing of step S203, the process proceeds to step S105.

[0106] <<Step S105>> The learning unit 4 uses the masked learning images 52 to train the learning model 5 . Then, the process returns to step S101.

[0107] <<<Detection device 102>>> An example of processing performed by the detection device 102 will be described with reference to the flowchart shown in FIG.

[0108] <<Steps S101 to S103>> In step S101, the image receiving unit 1 performs the same operation as in step S101 in the detection device 102 shown in FIG. In step S102, the signal acquiring unit 2 performs the same operation as in step S102 in the detection device 102 shown in FIG. In step S103, the signal determination unit 2b performs the same operation as in step S103 in the detection device 102 shown in FIG.

[0109] <<Step S104>> In step S104, the image acquisition unit 3 receives the acquisition signal 61 from the signal acquisition unit 2 and acquires the monitoring image 50 received by the image receiving unit 1 when the signal acquisition unit 2 acquired the reference signal. After the processing in step S104, the process proceeds to step S203.

[0110] <<Step S203>> The mask area application unit 8 masks the monitoring image 50 . Specifically, the mask area application unit 8 reads out the mask area 80 from the mask area storage unit 7. The mask area application unit 8 masks the mask area 80 on the surveillance image 50 to create a masked surveillance image 53 . After the processing in step S203, the process proceeds to step S111.

[0111] <<Step S111>> The inference unit 10 inputs the masked surveillance image 53 created by the mask region application unit 8 to the learning model 5 that the learning device 101 has trained based on the masked training image 52. The learning model 5 outputs an output result 70.

[0112] <<Step S112>> The analysis unit 11 performs analysis based on the output result 70. The analysis unit 11 calculates the difference between the masked surveillance image 53 input into the learning model 5 by the inference unit 10 and the output result 70 output by the learning model 5, and analyzes whether the surveillance area 40 is in a reference state.

[0113] <<Step S113>> In step S113, the analysis unit 11 performs the same operation as in step S113 in the detection device 102 shown in FIG.

[0114] ***Effects of the Second Embodiment*** As described above, the learning information generation device 100 of embodiment 2 can ignore movement of the background 42 of the learning image 51 by setting the learning image 51 to mask the area other than the home 41a. Furthermore, the detection device 102 according to the second embodiment can ignore movement of the background 42 of the monitoring image 50 by masking the area of ​​the monitoring image 50 other than the home 41a. Therefore, in the detection device 102 of the second embodiment, it is possible to improve the accuracy of detecting an object (obstacle 43) placed on the platform 41a.

[0115] ***Supplement to Embodiment 2*** A modification of the second embodiment will be described below.

[0116] <<<Modification of the Mask Area Setting Unit 6 of the Learning Information Generation Device 100>>> <<When the monitoring area 40 and the opening / closing device 30 are other than the platform door 30a>> In the present embodiment, the case has been described in which the image receiving unit 1 receives a monitoring image 50 in which the area near the platform door 30a is set as the monitoring area 40. Therefore, the mask area setting unit 6 sets the area excluding the platform 41a as the mask area 80, but the present invention is not limited to this. The mask area setting unit 6 may set, as the mask area 80, an area excluding the ground area that is different from the platform 41a.

[0117] As shown in FIG. 17, the mask area 80 may be changed arbitrarily to suit the opening and closing device 30. When the switching device 30 is a crossing gate and the monitoring area 40 is a railroad crossing, the mask area 80 is set to be an area excluding the ground surface within the railroad crossing (the ground surface includes the railroad tracks). When the opening and closing device 30 is an elevator and the monitoring area 40 is the inside of the car, the mask area 80 is the area excluding the floor surface inside the car. When the monitoring area 40 is an elevator hall, the mask area 80 is the area excluding the floor surface of the elevator hall. When the opening / closing device 30 is a ticket gate and the monitoring area 40 is the entrance to the ticket gate, the mask area 80 is the area excluding the floor of the entrance to the ticket gate. When the ticket gate is installed outdoors, the mask area 80 is the area excluding the ground of the entrance to the ticket gate. If the opening / closing device 30 is a security gate and the monitoring area 40 is an entrance to the security gate, the mask area 80 is the area excluding the floor of the entrance to the security gate. If the security gate is installed outdoors, the mask area 80 is the area excluding the ground of the entrance to the security gate. If the opening and closing device 30 is an automatic door and the monitoring area 40 is the area around the automatic door installed inside a building, the mask area 80 is the area excluding the floor surface around the automatic door. If the monitoring area 40 is the area around an automatic door installed at the entrance of a building, the mask area 80 is the area excluding the floor surface and ground surface around the automatic door. If the opening and closing device 30 is an automatic revolving door and the monitoring area 40 is inside an automatic revolving door installed inside a building, the mask area 80 is the area excluding the floor surface inside the automatic revolving door. If the monitoring area 40 is inside an automatic revolving door installed at the entrance of a building, the mask area 80 is the area excluding the ground surface inside the automatic revolving door. When the opening / closing device 30 is a vehicle and the monitoring area 40 is the interior of the vehicle, the mask area 80 is set to be an area excluding the seats inside the vehicle. Alternatively, the mask area 80 may be an area excluding the floor surface inside the vehicle.

[0118] The image receiving unit 1 receives, as the monitoring image 50, an image including a ground area that is at least one of the ground and the floor. The image acquisition unit 3 acquires, as a learning image 51, an image including a ground area from the monitoring image 50 received by the image receiving unit 1.

[0119] The area detection AI used in the mask area setting unit 6 is a model trained to detect a ground area, which is at least one of the ground and floor surface, included in the monitoring image 50 according to the mask area 80 corresponding to the opening and closing device 30.

[0120] <<When the mask area setting unit 6 sets the mask area 80 without using the area detection AI>> In the present embodiment, the case where the mask area setting unit 6 sets the mask area 80 using the area detection AI has been described, but the present invention is not limited to this. The mask area setting unit 6 may set an area set by the user while viewing the monitoring image 50 as the mask area 80.

[0121] <<<Modifications of the Mask Region Application Unit 8 of the Detection Device 102>>> <<When the mask area application unit 8 acquires the surveillance image 50 from the image receiving unit 1>> In this embodiment, the mask area application unit 8 of the detection device 102 acquires the monitoring image 50 from the image acquisition unit 3 and creates the masked monitoring image 53, but the present invention is not limited to this. The mask region application unit 8 may receive the surveillance image 50 from the image receiving unit 1 and create a masked surveillance image 53 . In particular, if the detection device 102 does not include an image capture unit 3 , the mask region application unit 8 captures the surveillance image 50 from the image receiving unit 1 .

[0122] Embodiment 3 The training information generation device 100, the training device 101, and the detection device 102 will be described with reference to FIGS. In the embodiments and drawings, the same or corresponding elements are denoted by the same reference numerals. The description of elements denoted by the same reference numerals as those already described will be omitted or simplified as appropriate. Arrows in the drawings primarily indicate the flow of data or the flow of processing.

[0123] ***Outline of the Third Embodiment*** First, the present embodiment will be briefly described. The area where the train 44 exists is detected from the video (monitoring image 50) captured by the 2D optical video camera (monitoring device 20), and masking processing is performed. The method for detecting the area where the train 44 exists is to utilize the fact that the direction of travel of the train 44 is constant and mask the area where optical flows 45 occur in the same direction. Alternatively, the train 44 is detected by object detection and the detected area is masked.

[0124] <<<Study time>>> <<Functions of the Learning Information Generation Device 100>> The functional outline of the learning information generation device 100 of this embodiment will be described below. The motion vector calculation unit 9 calculates an optical flow 45 for a video (monitoring image 50) acquired from a 2D optical video camera (monitoring device 20). The motion vector calculation unit 9 extracts, as a mask area 80, an area where optical flows 45 occur collectively in the same direction as the area where the train 44 is running.

[0125] <<Configuration of Learning Information Generation Device 100>> To perform the above functions, the learning information generation device 100 further includes a motion vector calculation unit 9. The motion vector calculation unit 9 calculates motion vectors (optical flows 45) from the input image (monitoring image 50) and extracts areas of optical flows 45 that occur together in the same direction.

[0126] ***Configuration Description*** <<<Functional Configuration of Learning Information Generation Device 100>>> The learning information generation device 100 will be described with reference to FIG. The learning information generation device 100 differs from the first embodiment in that it further includes a mask region storage unit 7, a mask region application unit 8, and a motion vector calculation unit 9.

[0127] <<Image Receiving Unit 1 and Image Acquisition Unit 3>> The image receiving unit 1 receives an image including a train 44 as a monitoring image 50. The image acquisition unit 3 acquires, as a learning image 51, an image including a train 44 from the monitoring image 50 received by the image receiving unit 1.

[0128] The mask area 80 in this embodiment is an area in the monitoring image 50 where the train 44 is present. A train 44 runs on a track 41b outside the platform 41a. For train 44, it does not matter whether it is running or stopped. For train 44, the position when stopped does not matter. For train 44, any type of vehicle is allowed.

[0129] <<Motion Vector Calculation Unit 9>> The operation of the motion vector calculation unit 9 to detect the mask area 80 will be described. The motion vector calculation unit 9 calculates an optical flow 45 indicating the movement of a certain area from a plurality of consecutive monitoring images 50 received by the image receiving unit 1. The motion vector calculation unit 9 detects a mask area 80 based on the calculated optical flow 45 . Specifically, the motion vector calculation unit 9 determines whether the optical flows 45 generated collectively in the same direction are greater than a preset threshold value. When the optical flow 45 occurs to a degree greater than a threshold value, the motion vector calculation unit 9 regards the area where the optical flow 45 occurs as an area where the train 44 exists, and detects it as a mask area 80. The motion vector calculation unit 9 stores the detected mask area 80 in the mask area storage unit 7. If the optical flow 45 is smaller than the threshold value, the motion vector calculation unit 9 does not detect the mask area 80.

[0130] <<Mask area storage unit 7>> The mask area storage unit 7 stores the mask area 80 detected by the motion vector calculation unit 9.

[0131] <<Mask area application unit 8>> The mask area application unit 8 masks the learning image 51 with a mask area 80 . The mask processing in the mask area application unit 8 will be specifically described with reference to FIG. The mask region application unit 8 acquires the learning image 51 from the image acquisition unit 3 . The mask area application unit 8 reads out the mask area 80 from the mask area storage unit 7 . The mask area application unit 8 masks the learning image 51 with the mask area calculated by the motion vector calculation unit 9 to create a masked learning image 52.

[0132] <<<Functional configuration of learning device 101>>> The learning device 101 will be described with reference to FIG. The learning device 101 in the third embodiment operates in the same manner as the learning device 101 in the second embodiment. However, the masked learning image 52 inputted to the learning model 5 by the learning unit 4 is an image obtained by masking the mask area 80 detected by the motion vector calculation unit 9 from the learning image 51.

[0133] <<<Functional configuration of the detection device 102>>> The detection device 102 has the same configuration as the detection device 102 in the second embodiment. The detection device 102 operates in the same manner as the detection device 102 in the second embodiment. However, the mask area 80 stored in the mask area storage unit 7 is the mask area 80 detected by the motion vector calculation unit 9 of the device 100 for generating information for learning. In addition, the mask area application unit 8 masks the mask area 80, which is the area where the train 44 exists and which was detected by the motion vector calculation unit 9, from the surveillance image 50 acquired from the image acquisition unit 3, to create a masked surveillance image 53.

[0134] <<<Hardware Configuration of the Training Information Generation Device 100, the Training Device 101, and the Detection Device 102>>> The hardware configurations of the training information generation device 100, the training device 101, and the detection device 102 will be described with reference to FIG. Note that the explanation of the same parts as those in the first and second embodiments will be omitted. The memory device 91 stores programs for realizing the functions of the image receiving unit 1, signal acquisition unit 2, image acquisition unit 3, learning unit 4, inference unit 10, analysis unit 11, motion vector calculation unit 9, and mask area application unit 8, as well as a learning model 5 and a set mask area 80.

[0135] ***Explanation of Operation*** <<<Learning Information Generation Device 100, Learning Device 101>>> An example of processing performed by the learning information generation device 100 and the learning device 101 will be described with reference to the flowchart shown in FIG.

[0136] <<Step S101>> The image receiving unit 1 performs the same operation as step S101 in the learning information generation device 100 shown in Fig. 15. After the process of step S101, the process proceeds to step S201.

[0137] <<Step S201>> The motion vector calculation unit 9 determines whether the mask area 80 is set (whether the mask area 80 is stored in the mask area storage unit 7). If the mask area 80 has already been set in the motion vector calculation unit 9, the process proceeds to step S102. If the mask area 80 is not set in the motion vector calculation unit 9, the process proceeds to step S301.

[0138] <<Step S301>> The motion vector calculation unit 9 acquires a monitoring image 50 including the train 44 from the image receiving unit 1, A mask area 80 is set. The mask area 80 set by the motion vector calculation unit 9 is stored in the mask area storage unit 7. After the processing of step S301, the process proceeds to step S102.

[0139] <<Step S102>> The state signal acquiring unit 2a of the signal acquiring unit 2 performs the same operation as step S102 in the learning information generating device 100 shown in FIG.

[0140] <<Step S103>> The signal determination unit 2b performs the same operation as step S103 in the learning information generation device 100 shown in FIG.

[0141] <<Step S104>> The image acquisition unit 3 performs the same operation as step S104 in the learning information generation device 100 shown in FIG. After the processing of step S104, the process proceeds to step S203.

[0142] <<Step S203>> The mask area application unit 8 performs the same operation as step S203 in the learning information generation device 100 shown in FIG. However, the mask area application unit 8 masks the mask area 80 detected by the motion vector calculation unit 9 on the learning image 51 to create a masked learning image 52. After the processing of step S203, the process proceeds to step S105.

[0143] <<Step S105>> The learning unit 4 performs the same operation as step S105 in the learning information generation device 100 shown in FIG. Then, the process returns to step S101.

[0144] <<<Detection device 102>>> The process performed by the detection device 102 in the third embodiment is the same as the process performed by the detection device 102 in the second embodiment shown in FIG. However, in step S203, the mask area application unit 8 uses the mask area 80 detected by the motion vector calculation unit 9 as the mask for the monitoring image 50.

[0145] ***Effects of the Third Embodiment*** The train 44 outside the platform 41a included in the monitored image 50 is likely to vary greatly depending on its operating state (stopped or running), its position when stopped, and the type of train. As described above, the learning information generation device 100 according to the third embodiment can ignore movement of the background 42 of the learning image 51 by masking the area in which the train 44 is included in the learning image 51. Furthermore, since the motion vector calculation unit 9 detects the area where the train 44 exists based on the calculation of the optical flow 45, the load on detecting the mask area 80 can be reduced. The detection device 102 of the third embodiment can improve the accuracy of obstacle detection by excluding the train 44 included in the monitoring image 50 from obstacle detection targets.

[0146] ***Supplement to embodiment 3*** A modification of the third embodiment will be described below.

[0147] <<<Modifications Regarding the Motion Vector Calculation Unit 9>>> <<Detection Operation of Mask Area 80 by Motion Vector Calculation Unit 9>> In this embodiment, the motion vector calculation unit 9 detects the mask area 80, which is the area where the train 44 exists, by calculating the optical flow 45 from multiple consecutive monitoring images 50, but this is not limited to this. The motion vector calculation unit 9 may detect the train 44 by object detection and set the detected area as the mask area 80. Specifically, the motion vector calculation unit 9 includes an object detection AI. The object detection AI is a model that has been trained in advance to detect the train 44 included in the surveillance image 50. The motion vector calculation unit 9 inputs the monitoring image 50 to the object detection AI. The motion vector calculation unit 9 stores the area in the monitoring image 50 where the train 44 is present, which is detected by the object detection AI, as a mask area 80 in the mask area storage unit 7.

[0148] The motion vector calculation unit 9 may combine detection of the mask area 80 by calculating the optical flow 45 and detection of the mask area 80 by object detection AI. Specifically, the motion vector calculation unit 9 detects the mask area 80 by calculating the optical flow 45, and then detects the mask area 80 by the object detection AI. The motion vector calculation unit 9 calculates the difference between the mask area 80 calculated using the optical flow 45 and the mask area 80 calculated using the object detection AI.

[0149] If the difference between the two mask areas 80 is equal to or less than a preset threshold, the motion vector calculation unit 9 determines that the two mask areas 80 represent the same area. Thereafter, the motion vector calculation unit 9 stores the mask area 80 obtained by calculating the optical flow 45 in the mask area storage unit 7. The motion vector calculation unit 9 may store the mask area 80 obtained by the area detection AI in the mask area storage unit 7.

[0150] If the difference between the two mask areas 80 is equal to or greater than a preset threshold, the motion vector calculation unit 9 determines that the two mask areas 80 represent different areas. That is, the motion vector calculation unit 9 determines that at least one of the mask area 80 calculated using the optical flow 45 and the mask area 80 calculated using the object detection AI is an erroneous mask area 80. In this case, the motion vector calculation unit 9 acquires a new monitoring image 50 and starts detecting the mask area 80 again.

[0151] Various aspects of the present disclosure are summarized below as appendices. (Appendix 1) an image receiving unit that receives a monitoring image of a monitoring area; a signal acquisition unit that acquires a reference signal indicating that the monitoring area is in a reference state; an image acquisition unit that acquires, as a learning image, a monitoring image at the time when the signal acquisition unit acquires the reference signal from among the monitoring images received by the image receiving unit in response to the acquisition of the reference signal by the signal acquisition unit; A learning information generation device comprising: (Appendix 2) the signal acquisition unit includes a status signal acquisition unit that acquires a status signal indicating a status of the monitoring area, and a signal determination unit that determines whether the status signal acquired by the status signal acquisition unit is the reference signal; The learning information generation device according to claim 1, wherein the image acquisition unit acquires the learning image when the signal determination unit determines that the signal acquisition unit has acquired the reference signal. (Appendix 3) the signal acquisition unit has an imaging instruction unit that, upon receipt of the reference signal by the signal acquisition unit, instructs a monitoring device that captures the monitoring image to capture the monitoring image; 3. The learning information generation device according to claim 1, wherein the image acquisition unit acquires the monitoring image captured in response to an instruction from the imaging instruction unit as the learning image. (Appendix 4) the image receiving unit receives, as the monitoring image, an image of the periphery of a switching device that blocks traffic, 4. The learning information generation device according to claim 1, wherein the signal acquisition unit acquires a signal indicating a closed state of the opening and closing device as the reference signal. (Appendix 5) the image receiving unit receives the monitoring image of the monitoring area, which is an area including a station platform; A learning information generation device described in any one of Appendix 1 to 4, wherein the signal acquisition unit acquires a signal indicating the closed state of a platform door installed on the platform as the reference signal. (Appendix 6) the image receiving unit receives the monitoring image of the monitoring area, which is an area including a station platform; A learning information generation device described in any one of Appendices 1 to 3, wherein the signal acquisition unit acquires as the reference signal a signal indicating the state of a train that has entered the platform and stopped until the train doors open. (Appendix 7) the image receiving unit receives an image including a ground area, which is at least one of a ground surface and a floor surface, as the monitoring image; the image acquisition unit acquires, as the learning image, an image including the ground area from the monitoring image received by the image receiving unit; A learning information generation device as described in any one of Appendices 1 to 6, comprising a mask area application unit that sets an area excluding the ground area of ​​the learning image as a mask area and masks the mask area from the learning image. (Appendix 8) the image receiving unit receives an image including a station platform as the monitoring image; the image acquisition unit acquires, as the learning image, an image including the platform from the monitoring image received by the image receiving unit; A learning information generation device as described in any one of Appendices 1 to 6, comprising a mask area application unit that sets an area of ​​the learning image excluding the home as a mask area and masks the mask area from the learning image. (Appendix 9) the image receiving unit receives an image including a train as the monitoring image; the image acquisition unit acquires, as the learning image, an image including the train from the monitoring image received by the image receiving unit; A learning information generation device as described in any one of Appendices 1 to 6, comprising a mask area application unit that sets the area in which the train exists in the learning image as a mask area and masks the mask area from the learning image. (Appendix 10) a motion vector calculation unit that calculates an optical flow indicating movement of a certain area from the plurality of consecutive monitoring images received by the image receiving unit, and detects the mask area based on the optical flow; 10. The learning information generation device according to claim 9, wherein the mask area application unit masks the mask area calculated by the motion vector calculation unit on the learning image. (Appendix 11) a learning information generation device that receives a monitoring image of a monitoring area and acquires, as a learning image, the monitoring image when a reference signal indicating that the monitoring area is in a reference state is acquired; a learning unit that learns a learning model for inferring whether the monitoring area is in the reference state based on the learning images acquired by the learning information generation device; A learning device comprising: (Appendix 12) a learning information generation device that receives a monitoring image of a monitoring area and acquires, as a learning image, the monitoring image when a reference signal indicating that the monitoring area is in a reference state is acquired; a learning unit that trains a learning model, including an autoencoder that outputs an output result for inferring whether the monitoring area is in a reference state from the monitoring image received by the learning information generation device, based on the learning image acquired by the learning information generation device; A learning device comprising: (Appendix 13) a learning information generation device that receives a monitoring image of a monitoring area and acquires, as a learning image, the monitoring image when a reference signal indicating that the monitoring area is in a reference state is acquired; a learning unit that, when an input image is input, extracts features of the input image and outputs an output result in which the input image is reconstructed based on the features, and that trains the learning model so that the output result approaches the input image based on the learning image acquired by the learning information generation device; A learning device comprising: (Appendix 14) an image receiving unit that receives a monitoring image of a monitoring area; an inference unit that inputs the monitoring image received by the image receiving unit into a learning model that has been trained using a learning image, which is a monitoring image obtained when a reference signal indicating that the monitoring area is in a reference state is acquired, and that outputs an output result for inferring whether the monitoring area is in a reference state when a monitoring image is input, and outputs the output result from the learning model; an analysis unit that analyzes whether the monitoring area is in the reference state based on the output result output by the inference unit; A detection device comprising: (Appendix 15) an image receiving unit that receives a monitoring image of a monitoring area; an inference unit that inputs the monitoring image received by the image receiving unit into a learning model that includes an autoencoder trained with a learning image that is a monitoring image when a reference signal indicating that the monitoring area is in a reference state is acquired, and outputs an output result for inferring from the monitoring image whether the monitoring area is in a reference state or not, and outputs the output result from the learning model; an analysis unit that analyzes whether the monitoring area is in the reference state based on the output result output by the inference unit; A detection device comprising: (Appendix 16) an image receiving unit that receives a monitoring image of a monitoring area; an inference unit that extracts features of a training image, which is a monitoring image when a reference signal indicating that the monitoring area is in a reference state, from the monitoring image received by the image receiving unit, inputs the extracted features into a learning model including an autoencoder that has been trained to output an output result that approximates the training image from the features, and outputs the output result from the learning model; and an analysis unit that analyzes whether the monitoring area is in the reference state based on a difference between the output result output by the inference unit and the monitoring image input to the inference unit; A detection device comprising: (Appendix 17) Receives surveillance images of the surveillance area, obtaining a reference signal indicating that the monitoring area is in a reference state; A learning information generating method, in which, upon acquisition of the reference signal, a monitoring image at the time of acquisition of the reference signal is acquired from among the monitoring images as a learning image. [Explanation of symbols]

[0152] 1 Image receiving unit 2. Signal acquisition unit 2a Status signal acquisition unit 2b Signal judgment section 2c Shooting instruction section 3. Image acquisition unit 4. Learning Department 5 Learning Model 5a Autoencoder 6 Mask area setting section 7 Mask area memory section 8 Mask area application section 9. Motion vector calculation unit 10 Reasoning part 11 Analysis Department 20 Monitoring equipment 30 Switchgear 30a Platform doors 40 Surveillance Area 41a Home 41b railroad track 42 Background 43 Obstacles 44 Train 45 Optical Flow 50 Surveillance Images 51 training images 52 masked training images 53 Masked Surveillance Images 60 Status Signal 61 Acquisition signal 62 Shooting command signal 70 Output results 71 Detection Results 80 Mask Area 90 processors 91 Storage device 92 Communication equipment 93 Input / Output Interface 100 Learning information generation device 101 Learning Device 102 Detection device

Claims

1. an image receiving unit that receives a surveillance image of a surveillance area that is an area including a station platform; A signal indicating the closed state of the platform doors installed on the platform is transmitted to the monitoring area when a person or object a signal acquisition unit that acquires a reference signal indicating that the state is a reference state without any intrusion of the When the signal acquisition unit acquires the reference signal, the image reception unit Among the monitoring images, the monitoring image when the signal acquisition unit acquires the reference signal is used as a learning image. an image acquisition unit that acquires an image; A learning information generation device comprising:

2. the signal acquisition unit is a status signal acquisition unit that acquires a status signal indicating a status of the monitoring area; a signal for determining whether the status signal acquired by the status signal acquisition unit is the reference signal; a code determination unit; The image acquisition unit determines whether the signal acquisition unit has acquired the reference signal. The learning information generation device according to claim 1 , wherein the learning image is acquired when

3. When the signal acquisition unit acquires the reference signal, the signal acquisition unit a photographing instruction unit for instructing a monitoring device that photographs an image to photograph the monitoring image; The image acquisition unit acquires the monitoring image captured in response to an instruction from the image capture instruction unit.

2. The learning information generating device according to claim 1, wherein the learning information generating device acquires the learning information as a learning image.

4. The signal acquisition unit acquires a state from when the train that has entered the platform stops until the doors of the train open.

2. The learning information generating device according to claim 1, further comprising a signal indicating a state of the learning information generating device as the reference signal. 。

5. The image acquisition unit selects an image including the platform from the monitoring image received by the image reception unit. an image is acquired as the learning image; The area of ​​the learning image excluding the home is set as a mask area, and the learning image is set as a mask area.

5. The method according to claim 1, further comprising a mask area application unit that masks the mask area. The learning information generation device described above.

6. the image receiving unit receives an image including a train as the monitoring image; The image acquisition unit selects an image including the train from the monitoring image received by the image reception unit. is acquired as the learning image, A region of the learning image where the train exists is set as a mask region, and The method according to any one of claims 1 to 4, further comprising a mask area application unit that masks the mask area. The learning information generating device described above.

7. A plurality of consecutive monitoring images received by the image receiving unit are used to select an area indicating movement of a certain region. Calculating an optical flow and detecting the mask region based on the optical flow a motion vector calculation unit for calculating a motion vector, The mask area application unit applies the motion vector calculated by the motion vector calculation unit to the learning image.

7. The learning information generating device according to claim 6, wherein the mask area is masked.

8. It receives surveillance images of the surveillance area, which is an area including the station platform, and The signal indicating the closed state of the platform door is sent to the monitoring area when the monitoring area is in a standard state with no people or objects entering. The monitoring image obtained when the reference signal indicating the state of the camera is acquired is used as the learning image. a learning information generation device; A learning model for inferring whether the monitoring area is in the reference state is a learning unit that performs learning based on the learning images acquired by the image information generating device; A learning device comprising:

9. It receives surveillance images of the surveillance area, which is an area including the station platform, and The signal indicating the closed state of the platform door is transmitted to the monitoring area when the monitoring area is in a standard state with no people or objects entering. The monitoring image obtained when the reference signal indicating the state of the camera is acquired is used as the learning image. a learning information generation device; The learning information generating device determines whether the monitoring area is in a reference state from the monitoring image received. A learning model including an autoencoder that outputs an output result for inferring whether the a learning unit that performs learning based on the learning images acquired by the image information generating device; A learning device comprising:

10. It receives surveillance images of the surveillance area, which is an area including the station platform, and The signal indicating the closed state of the platform door is transmitted to the monitoring area when the monitoring area is in a standard state with no people or objects entering. The monitoring image obtained when the reference signal indicating the state of the camera is acquired is used as the learning image. a learning information generation device; When an input image is input, a feature amount of the input image is extracted, and based on the feature amount, A learning model including an autoencoder that outputs an output result of reconstructing the input image. and a learning model that learns so that the output result approaches the input image is generated based on the learning information. a learning unit that performs learning based on the learning images acquired by the generation device; A learning device comprising:

11. an image receiving unit that receives a surveillance image of a surveillance area that is an area including a station platform; The image receiving unit receives the monitoring image and displays the closed state of the platform doors installed on the platform. A signal indicating the state of the monitoring area is transmitted to the monitoring area, which indicates that the monitoring area is in a reference state with no people or objects entering the area. This is a learning model trained using training images, which are monitoring images acquired as reference signals. When a monitoring image is input, the output result is used to infer whether the monitoring area is in the reference state or not. an inference unit that inputs the result into a learning model that outputs the result, and outputs the result from the learning model; Based on the output result output by the inference unit, the monitoring area is in the reference state. an analysis unit that analyzes whether A detection device comprising:

12. an image receiving unit that receives a surveillance image of a surveillance area that is an area including a station platform; The image receiving unit receives the monitoring image and displays the closed state of the platform doors installed on the platform. A signal indicating the state of the monitoring area is transmitted to the monitoring area, which indicates that the monitoring area is in a reference state with no people or objects entering the area. The autoencoder is trained using the training image, which is the monitoring image acquired as the reference signal. A learning model including a learning parameter, which determines whether the monitoring area is in a reference state from the monitoring image. The output results for inference are input to a learning model that outputs the output results, and the learning model outputs the output results. an inference unit that outputs the result; Based on the output result output by the inference unit, the monitoring area is in the reference state. an analysis unit that analyzes whether A detection device comprising:

13. an image receiving unit that receives a surveillance image of a surveillance area that is an area including a station platform; The image receiving unit receives the monitoring image and displays the closed state of the platform doors installed on the platform. A signal indicating the state of the monitoring area is transmitted to the monitoring area, which indicates that the monitoring area is in a reference state with no people or objects entering the area. The feature amount of the learning image, which is the monitoring image acquired as the reference signal, is extracted and the feature amount is An autoencoder trained to output results that approach the training image is an inference unit that inputs the output result from the learning model into a learning model including the input data; The difference between the output result output by the inference unit and the monitoring image input to the inference unit is an analysis unit that analyzes whether the monitoring area is in the reference state based on the result of the analysis; A detection device comprising:

14. The image receiving unit receives a monitoring image of a monitoring area that is an area including a station platform, A signal acquisition unit receives a signal indicating the closed state of the platform doors installed on the platform from the monitoring engine. A reference signal indicating that the rear is in a reference state with no people or objects intrusion is acquired. When the signal acquisition unit acquires the reference signal, the image acquisition unit acquires the monitoring image. Among these, the monitoring image when the reference signal is acquired is acquired as a learning image. Generation method.

15. A motion vector calculation unit calculates an optical flow indicating movement of a certain area from a plurality of consecutive monitoring images received by the image receiving unit, and detects the mask area based on the optical flow, The learning information generating device according to claim 6 , wherein the mask area application unit masks the learning image with the mask area calculated by the motion vector calculation unit.

Images