FR2944934A1 - Sites monitoring method for communication network, involves initially modifying video stream by integration of reference elements adapted to scenic contents of each image in order to identify causes of event - Google Patents

Abstract

The method involves detecting presence of an event triggering capture of a panoramic reference image by an image capturing unit along a direction of the event. The image is transmitted to a remote station through video streams. The video stream is initially modified by integration of reference elements adapted to scenic contents of each image in order to identify causes of the event. Displacement of cause of the event is automatically tracked by the image capturing unit. An independent claim is also included for a system for monitoring sites in a vision field of an image capturing unit connected to a remote station.

Description

TECHNICAL FIELD OF THE INVENTION [001] The invention relates to a method and a monitoring system. [002] The present invention relates to the field of site monitoring, and more particularly to that of televideosurveillance which consists of the exploitation of images of a site or several sites for surveillance purposes, since a checkpoint located at a site remote from the monitored sites. The monitored sites and the surveillance site are interconnected by a communication network. STATE OF THE PRIOR ART [003] Televideosurveillance is fundamentally different from video surveillance operated from a local station by one or more monitoring agents. [004] A local video surveillance station is generally organized around the following HMIs: - An image wall to display all the live cameras so that the agent can have a global view of the site and select a camera; - An alarm screen on which the cameras associated with the sensors in alarm are automatically displayed; - A supervision screen with the alarms on the water, the instructions to be applied and the site plans, and - A tracking screen allowing the agent to control the cameras live or reread records. [005] Most often, the monitoring agents only manage the site on which the checkpoint is located. As a result, the alarms are processed without delay and the removal of doubt video, operation of qualifying the alarm of a sensor by a camera, operates essentially by the live view of the alarm screen and the wall images. [006] Local video surveillance consists of a simple live viewing of an image wall is now obsolete. [007] Televideosurveillance poses three different problems: 1. The alarms are received offline, from a few ms (IP network) to several seconds (RTC or GSM Data). 2. The probability that a list of alarms waiting to be processed at the remote control station is more likely than in local management, due to the possibility of pooling the monitoring of multiple sites 3. The low speed IP networks (<1 Mbps) typically used to interconnect monitored sites with the central station limit the number of cameras that can be simultaneously viewed remotely with sufficient quality. [008] As a result of 1, the live connection to the alarmed camera can To be too late to qualify the event: for example, an intruder triggering the alarm by crossing a fence in front of a camera may be out of camera two seconds later, and the alarm can only be qualified based on the images recorded at the time of the alarm. [009] As a consequence of the 3, the operation of simultaneously viewing all the cameras is technically impossible with a sufficient level of quality so that the video analysis phase of the threatened site can be integrated into the process of securing the remote site. [010] Therefore, in the current state of the art, televideosurveillance is most often restricted to the video qualification of alarms and the dispatch of a field intervention team. [011] Such a televideosurveillance strategy creates two major security flaws: the evolution of the threat represented by the intruder, between the intrusion and the arrival of the intervention teams, is not mastered: the intruder has been able to abandon an explosive or compromise the safety of an equipment, and - the intervention team does not benefit from any information enabling it to locate the intruder: it will therefore waste time searching for it and, eventually, conclude to leave the site while the intruder is still there. [012] These security flaws are not eligible on most sensitive sites, security managers generally refuse to outsource CCTV. However, on many sensitive sites, it is possible to identify time slots during which the activity of the site being greatly reduced, televideosurveillance could advantageously replace the local guarding, subject to the demonstration of the effectiveness of the techniques put implemented. [013] The Scutum project consists in removing all technological barriers and in implementing an organization and techniques enabling it to offer credible and effective televideosurveillance services. [014] One of the important aspects of this project is the HMI of the operators in charge of the video analysis of the sites. The GUI alone crystallizes the credibility of the Scutum project. In the decision phase of a safety manager, his demonstration will be decisive. [015] This HMI should be conceived as a tool allowing an operator in charge of the surveillance of several sites, to immerse himself in a few seconds in a new site in alarm, to take quick control of it and to have automatic access. level of information necessary for the exercise of its mission. The ergonomics of this HMI must be particularly careful, because it is at the same time, guaranteeing the efficiency of the operator. SUMMARY OF THE INVENTION [016] The present invention aims to solve the problem related to the technical difficulties encountered for the deployment and reliability of video surveillance mechanisms. [017] To do this, one aspect of the invention relates to a method of monitoring a site located in the field of view of at least one image capture means connected to a remote station, the method comprising the following steps: detecting the presence of an event capable of triggering the capture of at least one image by the image capture means in the direction of said event; transmission of said at least one image to the remote station through a video stream, in which said video stream is initially modified by the integration of reference elements adapted to the scenic content of each of the images composing this video stream so that identify the cause and context of the event. [018] According to particular embodiments, the method comprises: a prior step of acquiring at least one reference panoramic image in the equirectangular format of the field of view of at least one image capture means; step of transmitting said at least one reference panoramic image to the remote station, - a series of calculation steps, from said at least one panoramic panoramic image, of panoramic images in rectilinear format each corresponding to an enlarged vision of said scene with respect to the vision of each image composing said video stream,

a series of steps of merging said calculated panoramic images and images composing said video stream in order to constitute an augmented reality video stream rendering, on the remote station, a field of view of said artificially expanded scene with respect to the initial video stream;

[019] According to particular embodiments, said reference elements relate to one of the information relating to one of the following elements:

- designation and identification of the cause of the event;

location of the cause of the event and / or - indication of the buildings and infrastructures of the site (s). [020] According to particular embodiments, the method comprises a step of automatically tracking the movement of the cause of said event by at least two image capturing means, characterized by the following steps: - tracking the displacement of the cause said event by the image capture means 1 - input detection of the cause of said event in an area of the field of view of said image capture means 1 or in an area of the field of view of a capture means image 2 - focusing of the image capture means 2 towards the cause of said event confirmation of entry of the cause of said event in an area of the field of view of said image capture means 2 and capture of at less an image of the cause of said event by said image-capturing means - followed by the displacement of the cause of said event by the image-capture means 2 [021] According to particular embodiments, the method includes a preliminary step of parameterizing at least one detection zone in the field of vision of at least one image sensor, characterized by the following steps: - association of an altitude matrix with said panoramic image allowing enter, for any point of the reference panoramic image, the altitude value at this point, marking on a layer superimposable on said panoramic image of said characteristic points of said detection zone, calculation of the altitude values in these areas; characteristic points of said detection zone and input of said values in the altitude matrix, 15 - complete manual parameterization on said layer of said detection zone, with the possible assistance of image analysis means. [022] The invention also relates to a system for monitoring one or more sites situated in a field of view of at least one image capture means, for carrying out the method, comprising at least one sensor associated with said at least one image capture means, capable of detecting an event likely to trigger the capture of at least one image by the image capture means in the direction of said event so as to transmit said at least one an image at the remote station through a video stream, characterized in that the system comprises processing means able to modify said video stream by integrating reference elements adapted to the scenic content of each of the images composing this video stream so as to identify the cause and context of said event, and by merging calculated panoramic images with the original video stream to artificially expand the field of view of the scene. [023] Advantageously, the system comprises a unit for automatically tracking the cause of said event by the different means of capturing images of the same site. [024] According to a particular embodiment, the image capture means is a thermal dome. BRIEF DESCRIPTION OF THE FIGURES [025] Other features and advantages of the invention will emerge on reading the description which follows, with reference to the appended figures, which illustrate: FIG. 1 is a view of a video surveillance buddy FIG. 2 is a view of a site under video surveillance, FIGS. 3, 4 and 5 show a view provided by a dome, FIGS. 6, 7 and 20 show a view. 8, relates to a schematic view of a human-machine interface (HMI) according to the invention, FIG. 9 is a view of a fixed-focus camera, FIGS. 10 and 11 show a view of a dome; FIG. 12 is a schematic view of the image recovery and merge into a single panoramic image; FIGS. 13 and 14 show a view of a panoramic image; FIG. 15 is a schematic view of a panoramic image and an embodiment according to the invention. FIG. 16 represents a schematic view of the dynamic positioning of a target according to the invention; FIG. 17 is a view of the absolute pointing of an area to be viewed according to the invention; FIG. is a schematic view of the system according to the invention, FIG. 19 is a schematic view of a target tracking according to the invention; FIGS. 21, 22, 23 and 29 show a schematic view of the panoramic images according to FIG. 24 is a diagrammatic view of a dome according to the invention; FIGS. 25, 26 and 27 show a schematic view of the transformation of the panoramic image into the plane according to the invention; FIG. 28 is a schematic view of the accuracy in the vicinity of the horizon according to the invention; FIG. 30 is a schematic view of the detection of an intruder, FIG. 31 is a substantially similar schematic view. in FIG. 2, FIGS. 32 to 45 show a schematic view of the operation 1 FIG. 46 is a diagrammatic view of a dome coupled to autotracking; FIGS. 47 and 48 show a view relating to autotracking and augmented reality; FIGS. 49 to 56 show a view of the cover of the system according to the invention, - figures 57 to 61 show a schematic view of the calibration of the system according to the invention, and - figures 62 to 69 show a schematic view. the target tracking mechanism by the system according to the invention. [026] For greater clarity, the identical or similar elements are marked with identical reference signs throughout the figures. DETAILED DESCRIPTION OF AN EMBODIMENT [027] In FIG. 3, the video opposite is provided by a dome with a 320x240 pixel thermal sensor. The image is extended in full screen by the operator (1920x1200 pixel screen 24 "), where it can clearly see an individual heading towards a portal The scene is set 250 m from the dome. [028] operator has identified a threat, but can not contextualize it: - Where is the intruder? - Which buildings are threatened? - Which domes or cameras can identify the individual? - Where is the speaker? [029] Instinctively, the operator would like to command the dome to zoom out to widen the field of view and search for information in the image: the goal of augmented reality is to provide that information by immersing it as much as possible in the site. [30] On command of the operator, the augmented reality engine of the video surveillance site will gradually widen the field of vision of the thermal dome by revealing neighboring buildings and infrastructures (roads ). He will also insert information useful for locating the target and stakeholders. [31] Figure 4 shows a virtual panoramic immersion. The thermal dome does not have a continuous zoom but a fixed focus lens of 50mm or 100mm. It can not produce live images of buildings appearing gradually in the screen when zoomed out: they are calculated in real time by the augmented reality engine from a panoramic photo mapping made a few minutes ago by the dome. Thus, on the screenshot opposite, the floating central image is real, the panoramic wallpaper is virtual. [032] By pushing the zoom out further, the operator can get a complete panoramic view of 360 ° horizontally and up to 180 ° vertically. [033] This technique, as illustrated in Figures 4 and 5, is applicable to thermal domes fixed focal and visible domes. Indeed, when the zoom of the dome is on a high value, the manual movements of the dome by the operator are delicate, even knowing the site well. Now, from the panoramic view of augmented reality, the operator can designate with the mouse an area of interest in the image in order to focus the dome. [034] The augmented reality engine will also enrich the stream with simple elements (texts, symbols ...): - designation and identification of targets (example: vehicle at 500m); - identification of buildings visible in the image closest to the target when the operator flies over them with the mouse (example: EDF, GEFCO); - precise location of the target in the environment (example: north portal); - location of important equipment to report to stakeholders (access control readers for example). It will finally insert in the video stream, two channels of real-time geolocation information: - the operator geolocation channel allowing the operator to view: • the location of the intruder on a detailed map of the area, • the video resources at its disposal to visualize the target, • possibly the property limits (example: green lines) • the important equipment for the speaker.

By clicking on the symbol of the dome or a camera, it can obtain a preview of the vision of this one (right click) or the commutation of the live video on this camera (left click). - the intervening geolocalisation channel (option) allowing to view the relative positions of the intruder (red square example) and the speakers on the site plan (example: green and blue symbols). [035] This list is not exhaustive, the principle being to enrich the information without destroying the visual feeling of immersion in the scene of the operator by too much information in his field of vision . [036] Figure 6 shows a view of the geolocation channel of a speaker, and Figure 7 a geolocation channel of the operator.

[037] The view shown in FIG. 8 is a modular view of the organization of the surveillance screen whose ergonomics can be modified according to the operator's choices.

[038] Thereafter, a panoramic mapping is made. This operation consists in photographing the landscape over the entire range of action of a thermal or visible dome and fusing these photos into a single panoramic image in equirectangular projection.

[039] In Figure 9, there is shown a thermal dome. A fixed focal length camera of 50mm has an angular aperture of the image of 14 ° by 10 °. In Fig. 10, a dome has an amplitude of 360 ° in elevation and -45 ° / + 90 ° in azimuth. Sweeping in all 14 ° site and in azimuth every 10 °, it takes 26x14 = 338 photos in total to map all the vision of the dome. At the rate of 6 images / s, it takes 60 s to perform the mapping. If a JPEG image is 8 KB, 2.6 MB memory is required to store the panoramic image in native resolution. At a transmission rate of 1 Mbps, it takes 21 seconds to transmit the panoramic image. [40] In Figure 11, restricting the mapping to 270 ° in site (we remove the 90 ° at the back of the dome) and 45 ° in azimuth (we start from the horizontal and we go down by 45 °) , it takes 20x5 = 100 photos, 16 s reasons for 6 images / s, 0.8 MB 30 memory, and 6s transmission at 1 Mbps. [41] From a visible dome performing the maximum zooming mapping, the angular aperture of the image is 54 ° by 42 °. For a total mapping5, you need 7x4 = 28 photos, ie 4s with 6 CIF / s images, ie 224 KB of memory and 2 s of transmission at 1 Mbps. For a partial mapping, you need 6x2 = 12 photos, ie 96 KB of memory and 1 s of transmission at 1 Mbps. [042] In Figure 12 is the straightening of images and merging into a single image or panoramic view. [043] The image must be straightened before being merged, in order to obtain the continuity of the pixels on a rectangular memory (see tools opensource panorama tools, PTAssembler ...). [044] The features also refer to a panoramic and point of interest (POI) image, as shown in Figure 13. To insert contextual information, the panoramic image is more natural to work than the site map because it corresponds to the viewing angle of the operator. For example, you can indicate points of interest: • Roads, crossroads • Buildings • Access (gate, gates ...) • Defended property limits (peripheral detection) • Important equipment for the workers (control readers 20 access). It is possible to indicate when inserting the POI (point of interest) on the panoramic image, if the information must be systematically inserted in the augmented image of the operator station or if it must appear only when the operator fly over the area concerned with his mouse. [045] The features also relate to a panoramic image and PTC (Points to Check). In the case of an intruder abandoning an object on the site, waiting for the safety of a site equipment, or having suspicious behavior at a specific location, the operator must be able to insert points to be controlled by the team. intervention. The potential insertion mode can be done by direct pointing in the image (example: right click, opening of an input window, entering a text) with automatic report on the panoramic image and on the plane of the image. site (operator channel and intervening channel). The HMI must also provide for the erasure of the PTCs. [046] Features also relate to a panoramic image and intelligent video detection. • Comparison of panoramic images n I n-1

If the panoramic image is refreshed at a high frequency (every 10 minutes for example) and is supplied to an intelligent video detection engine, the latter can suspect the appearance of an intruder by detecting the appearance of hot spots (white) from one image to another. It gives back information on the target (type, distance ...) to the augmented reality engine so that it inserts them into the HMI.

15 • Reference image

The Intelligent Video Detection Engine may need a reference image to help it analyze a scene where it does not detect or move. 20 Take the example of a dome moving towards a field sensor on alarm information. The intelligent video detection engine may have difficulty in confirming the presence of the intruder on the sole basis of motion detection. Indeed, if the intruder has stopped or moves slowly toward the dome, the number of pixels varying in the image may be zero or insignificant.

The detection engine can then use the latest panoramic image and compare it to the live image to help locate the target. This presence suspicion information may, for example, motivate a more prolonged observation of the area or be subject to the opinion of the operator. [047] The features also refer to a panoramic image and zone scan, as shown in Figure 14.5 To find a lost target, or to start a round on the fly, the operator may have to designate an area to be scanned by a dome from selection elements. The operator enters a simple segment or a more complex path, designates the sectors to be neutralized (road located outside the site, zone of authorized activity), indicates the speed of scan and names this round. When it launches the round, the dome will perform a slow and continuous landscape scan following the path and the intelligent video detection engine will assist the operator to locate moving targets in unneutralized areas. [048] The features also relate to a panoramic image and hypertracking engine, as illustrated in Figure 15. The hypertracking engine is a site device for automating the tracking of a target by the same camera (autotracking) and cameras in cameras. On the diagram opposite, thermal domes, visible and fixed cameras contribute to the security of an urban area. An individual crossing the Boulevard Georges Pompidou from left to right will be followed initially by the thermal dome installed on the building of PTT, then by the dome (yellow) of the stadium, whose range of action is represented on the plan of the site by the yellow circle. If the individual crosses the rue de Bayeux to the Rond-point de la Maladrerie, it will be followed initially by the PTT thermal dome, then taken again by the thermal dome located at the extreme right of the plan.

The action radiuses of the domes are indicated at the level of the ground plane during the parameterization phase, then reported in the panoramic images of the different domes for the operational implementation of the hypertracking. This report is automated by the system configurator. The technician performing the adjustments on the site will then adjust the rays of action on the panoramic image and add the conventions of direction of crossing. The final result is then transmitted to the hypertracking engine which translates it into instructions. [049] The invention comprises a target autotracking mechanism confirmed or designated by the operator. This operation is essential: - for the automatic tracking of an intruder between the moment it is detected and confirmed by a dome, and the moment when an operator will take the mission of monitoring the site live; during these few seconds (see tens of seconds), the site is not under operator control. - facilitate the work of the operator on a monitoring and analysis mission; from the HMI, the operator must be able to engage and disengage the hypertracking, designate a new target and recommit autotracking on this new target. [050] In the invention, a dynamic report of the position of the target on the site plan is performed. As shown in Figure 16, we prefer a coarse positioning (10 m accuracy) calculated from the position of the optical axis of the dome, possibly corrected by contour of the site when there are on it unevennesses important. These contour lines can be a matrix of level values attached to the panoramic image. The target is carried over to the intervening channel and the operator channel. [051] The invention also provides an absolute pointing of an area to be viewed by a dome, as illustrated in FIG. 17.

The operator must be able to indicate on the panoramic image, an area to be viewed live by a dome. It surrounds the area of interest with the mouse (rectangle) and the HMI controls the positioning of the optical axis of the dome on the center of the designated area, calculates the zoom level (virtual / real proportion) and increases the Dome image of the calculated panoramic image. [052] As illustrated in FIG. 18, the architecture of the system according to the invention comprises the following elements: the ground sensors: • detect the intrusions; • trace the information over a mesh radio network to a gateway connected in IP to a server. - On the server, a software interface (mesh SDK): • Acquires alarms (type, location); • transmits them to the hypertracking engine. - The hypertracking engine: command to the dome closest to focus on the target and to hang it; • forms a video clip and transmits it to the televideo monitoring center for qualification transmits to the augmented reality engine the video flow of the live dome; • Follows camera to camera and transmits video stream to augmented reality engine. - The augmented reality engine: • merges the real-time video with the recalculated panoramic image; • postpones the target on the site plan; • presents the merged information on the analyzer station of the remote surveillance center. The augmented reality module is composed of: a server module installed on a server of the remote site; • one or more client modules, installed on the stations of the analyzers operators. - The client module: • calculates the panoramic image; • display the merged content; • Raises the operator controls (mouse, keyboard) to the server module. - The server module: retrieves video streams via the hypertracking engine and merges the data; • pilot the domes for the acquisition of panoramic images, via the hypertracking engine; • goes back to the hypertracking engine the controls of the operator. [053] The hypertracking engine switches one video source at a time to the augmented reality engine. The video source composed of the succession of video sources relevant for tracking a target is called the alarm channel. [054] To switch from one camera to another, the principle is as follows: • Camera 1 follows the target; • a triggering event (crossing a boundary, for example) prepositioning one or more cameras to the target; • One or more cameras hang on the target and provide a JPEG image of the snapping; • the hypertracking engine chooses a camera according to a list of criteria; • He also follows the target with the chosen camera; • It switches the selected camera in the alarm channel when the camera 1 loses the target. [055] Cameras that have not been selected to track the target are the candidate channels. [056] In FIG. 19, the cameras participating in tracking the target are symbolized by the gray bands A (candidates), yellow B (alarm channel) or black C (lost target). Each horizontal band corresponds to a different camera. The triggering event is symbolized by the red line D. The snapping of a camera is materialized by the JPEG photo of the event.

This window is therefore a presentation of metadata provided by the hypertracking engine. It constitutes an element of the hypertracking strategy: it is presented by the augmented reality engine at the request of the analyzer operator, when the latter finds or suspects an error in the choices made by the hypertracking engine (tracking a bad target).

By clicking on a thumbnail, the operator can view the image in its native format. By clicking on a candidate channel, the operator can replay the corresponding video record or switch the live video source into the augmented reality engine. [057] As illustrated in FIG. 20, the position of the target is plotted on the operator geolocation channel. It is interesting to also refer to the cameras hanging on the target.

A right click on one of its cameras must allow to preview the camera (JPEG image, for example) without modifying the alarm channel. A confirmed left click must allow the candidate resource to be switched in the alarm channel. [058] The system according to the invention comprises connection and communication means adapted to the flows between the client module and the augmented reality server module. [059] The objective of the augmented reality module is to provide all the useful information to the operator for his analysis of the evolution of the threat and the steering of the intervention teams. The information flows must be optimized because the bandwidth available on the interconnection VPN network will be less than 1 Mbps and, most often, of the order of 500 kbps or 256 kbps. [060] The hypertracking engine helps to save bandwidth by choosing, for the operator, the camera best positioned to follow the target. [061] Some image weight and video bitrate data: • A 4CIF JPEG image weighs about 32 KB in B & W and 50 KB in color • A JPEG image in QCIF is about 8 KB in B & W and 12 KB in color • An MPEG4 stream in CIF at 6 frames / s occupies a bandwidth of 128 kbps (thermal camera). • An MPEG4 stream in 2CIF at 6 frames / s occupies a bandwidth of 256 kbps. • An MPEG4 stream at 4CIF at 6 frames / s occupies a bandwidth of 512 kbps.

In H264, the bandwidth requirements are the same, the quality being significantly better than in MPEG4 especially on the dimly lit scenes. [062] The augmented reality engine will increase the video flow centered on the target with the following streams, which are discrete streams: • panoramic image: for each dome, it is downloaded once by the client module, the first time the dome is switched in the alarm channel; on sites with low uploads, the reference panoramic image must be resampled to a lower resolution to accelerate the transfer to the augmented reality client module; • geolocation channel: the JPEG image of the site map is also downloaded once during the first connection to the site. The position of the target and the highlighting of available resources is discrete information that does not take up bandwidth. The update can take place every second; • candidate channel window: this window contains information about the choices made by the hypertracking engine.

Its composition requires: • the reception of discrete data that does not consume bandwidth and can be transmitted every second in a structured frame (XML); • the receipt of thumbnails of a dome or a camera (thumbnails JPEG of 8 KB, 1s to go up with 64 kbps of reserved bandwidth); • native resolution preview of the snapping JPEG image: a 32 KB image puts 4s to go up with 64 kbps of bandwidth reserved for the preview channel; • preview of a candidate channel in native resolution: it consists of the uploading of a JPEG image of the camera concerned at the request of the operator. [063] Table relating to the flow between the client module and the augmented reality server module: Bandwidth 1Mbps * 512 kbps 256 kbps alarm channel 512 * 256 * 128 * Preview 256 128 64 Geolocation 256 128 64 + candidate channels * CIF or * 4CIF * 2CIF 320x240 [064] The processing mechanism implemented by the invention offers a calculation power adapted to the system. [065] Server Power: On the server, the augmented reality server module will coexist with the hypertracking module and the intelligent video detection module. The intelligent video detection module consumes machine resource: currently a maximum of 8 video sources can be analyzed simultaneously on a quadcore server. The hypertracking module will consume a little resource for autotracking, almost nothing for camera tracking camera. The augmented reality server module will consume resources during the 15 phases of assembly and calibration of the panoramic photos. The calculation times, depending on the number of photos to be assembled, the number of domes and the mapping frequency, will determine the machine power to reserve. It is not excluded that a machine must be dedicated to augmented reality. [066] Power of the client station: The client must perform the MPEG4 or H264 decompression of the alarm channel of the real-time or recorded image. • 22 The client must also perform the extraction of panoramic vision centered on the real-time image (rectilinear projection) from the panoramic image in equirectangular projection. The computing power required depends on the desired degree of synchronization.

Unlike Augmented Reality models where 3D objects are calculated and added to the actual image 25 times per second, the augmented reality of our televideo surveillance model adds mostly virtual content to the outside of the actual image. The attention of the operator is supported in the center of his vision (fovea) on the live image. The immersion in the image is simulated by the reproduction of a panoramic vision calculated and adjusted on the live image. In order to visually overcome a perfect link between the live image and the panoramic image, it is necessary to visualize the image or real-time view as a floating window above the panoramic image, the two windows being able to be synchronous or not, depending on what the operator wants: • Asynchronous windows: the operator needs to be in the panoramic image while keeping an eye on the target in the live image: it circulates in the panoramic image, then request the resynchronization of the panoramic on the real time window • Synchronous windows: panoramic vision and real-time vision are perfectly synchronized. This is the principle of augmented reality. Ergonomic tests will have to determine what is acceptable as resynchronization delays of the 2 windows. We can keep the real-time floating window centered in the center of the HMI and recalculate the panoramic image accordingly. It is also possible to take into account the displacement of the dome by moving the floating window on the panoramic image, which avoids to recalculate the panoramic image. [067] The invention uses panoramic tools. As an example, for a panoramic image, photo collation and client viewer, tools such as Panorama Tools' open source programs and libraries, originally written by German Physics and Mathematics Professor Helmut Dersch, can be used. For an example of rendering on client workstation: http://www.animatif.com/panos/viaduc_mil.htm. [068] The calibration of panoramic images is necessary because of their use as reference images by the intelligent video analysis engine and as a trigger by the hypertracking engine, as illustrated in Figures 22 and 23. Various factors may lead to what 2 panoramic images taken under the same conditions by the same dome are not superimposed perfectly: support dome unregulated, wind bending the mast ... In the example opposite, the 2nd image is shifted a few degrees in Pan (x). For calibration, the autocorrelation calculation over the entire megapixel panoramic image would be too expensive in computing time. A faster method should be found, for example, by matching 2 or 3 16x16 pixel blocks characteristic of the reference panoramic image, with the same areas of the raw panoramic image; the calibration amounts to finding the shifts in x and y (Dx and Dy) of the transformation to be applied to the panoramic image so that it is superimposed on the reference image. [069] One of the advantages of the invention is to allow location geolocation of the site. [070] The domes, illustrated in FIG. 24, compatible with the concept of augmented reality must accept an absolute positioning control of their optical axis for their control, and an information request on the current PTZ values, for the synchronization. of the panoramic image on the live image. Like for example IP domes such as Sony RX550 type models. For such a dome, the Sony CGI commands are included in the following table: ° <pan pan position: "E020" to "1FEE" F) position>, R ~ tili Abso; utePanTilt position>, ~ speed> 'ptzf ptzf acted inptset "F808" 07F8 "tilt position: to speed: to 24 with set = position inq = retrieve position [O71] Commands and queries are sent in HTTP to the Sony IP dome: - Set the dome to IP address 192.168. 1.1 at 0000.0000 at the maximum speed: http://192.168.1.1/command/ptzf.chi?AbsolutePanTilt=0000,0000,24 Retrieve the current PTZ values: http://192.168.1.1/command/inquiry.cgi ? parameter = ptzf [O72] The invention offers the advantage of allowing the transformation of the panoramic image to the plane and vice versa. [O73] With regard to the panoramic image on the plane, as illustrated in FIG. 25: - Xa , Ya: absolute coordinates of the target in the common reference frame of the site (site plan) - X1, Y1: absolute coordinates of dome 1; R1: rotation of dome 1 reference relativity to the common reference system; - x1, yl: relative coordinates of the target in the reference of the dome 1; - x'1, y'1: relative coordinates of the target in the reference of the dome 1, after rotation R1 Xa = x'1 + X1 Ya = y'1 + Y1 p, t - in Figures 26 and 27- correspond to Absolute Pan and Tilt coordinates of the optical axis, h to the absolute dome installation altitude, n to

the absolute altitude at the point of intersection of the optical axis with the ground, and d at the distance of the target at the foot of the dome.

R1 corresponds to the rotation of the reference of the dome with respect to the absolute reference. • d = (h-n) tan (rr / 2-t)

• xl = d cos (p)

• yl = d sin (p)

X'1 = d cos (p + R1)

• y'1 = d sin (p + R1) Xa = (h-n) tan (rr / 2-t) cos (p + R1) + X1

Ya = (h-n) tan (rr / 2-t) sin (p + R1) + Y1

[074] Regarding, the plan to the panoramic image

• x'1 = XaùX1

Y'1 = Ya where Y1 • d = root (x'l 2 + y 'l 2)

• p + R1 = Acos (x'1 / d)

• rr / 2-t = Atan [d / (h-n)]

p = ùR1 + Acos [x'1 / root ((Xa-X1) 2+ (Ya-Y1) 2)]

t = rr / 2 where Atan [root ((Xa-X1) 2+ (Ya-Y1) 2) / (hn)] [075] In Figure 28 is illustrated the mechanism for obtaining a precision in the vicinity of the horizon.

The distance of the target as a function of the relative height dome / target (hn) and the value of the inclination with respect to the horizon in radian (t) is given by: • d = (hn) tan (rr / 2-t) The measurement of the distance as a function of the tilt measurement of the optical axis is obtained according to the process below. The derivative of the distance d with respect to the inclination t is: • var (d) / var (t) = - (hn) / cos2 (rr / 2-t) Near the horizon (t = 0) , the derivative of the distance d with respect to the inclination t tends to infinity: the calculation of the distance undergoes strong variations according to the error made on the measurement of the inclination of the optical axis var (t).

This value is set by autotracking which will position the optical axis at the foot of the target. This setting can not be more accurate than a pixel in the image. The tables below show that below 2 degrees, the error is greater than the relative height of the dome / target. thermal dome (50mm) vertical aperture (degree) 10 vertical resolution (pixel) 240 var (t) for 1 pixel (radian) 0.0007 relative height dome / target (m) 10 t in degree 1 2 3 4 5 var (d) for 1 pixel I 23.9 6.0_ 2.7 1.5 1.0 visible dome (Sony RX550 zoom out) vertical aperture (degree) 40.65 vertical resolution (pixel) 480 var (t) for 1 pixel (radian) 0.0015 relative height dome / target (m) 10 t in degree 1 2 3 4 5 var (d) for 1 pixel 48.5 12.1 5.4 3.0 1.9 Thermal dome and visible dome boards The distance measurement according to the measurement of the target relative height / dome is obtained according to the process below. 26 20

The derivative of the distance d relative to the relative altitude of the target n is:

• var (d) / var (h-n) = tan (rr / 2-t)

In the vicinity of the horizon (t = 0), the derivative of the distance d relative to the relative height dome / target (hn) tends to infinity: the calculation of the distance undergoes strong variations according to the error on measuring the relative height of the target var (hn). This value is set by the quantization value of the contour lines (accuracy at 1m, 2m or 5m).

[076] t in degree 1 2 3 4 5 10 15 20 25 var (d) for 1 m 57 29 19 14 11 6 4 3 2 var (d) for 2 m 115 57 38 29 23 11 7 5 4 var (d) ) for5 m 286 143 95 72 57 28 19_ 14_ 11 Table relating to geolocation [077] The table above gives the accuracy of the geolocation and depends on the accuracy of the measurement and the relative height. • For an accuracy of 1m on dome / target relative height measurement, below 5 °, the distance error is greater than 10m. • For an accuracy of 2m on dome / target relative height, the error is greater than 10m below 10 °. • For a 5m accuracy on dome / target relative height, the error is greater than 10m below 25 °. [078] We can therefore refer to the equirectangular panoramic projection of a dome, several bands centered on the horizon (t = 0). • Band +/- 2 °: within this band, the calculation of the distance gives a too important error. The position of the target is frozen on the plane until the next acceptable measurement given by the same dome or possibly by another dome. [079] As shown in Figure 29, for sites with elevations, these must be reported on the matrix associated with the pan for calculating the distance of the target. • Band + 1- 5 °: within this band, the dome / target relative height must be reported with an accuracy of 1m, for an error on the distance less than 10m; • Band + 1- 10 °: within this band, the dome / target relative height can be reported with an accuracy of 2m, for an error over the distance less than 10m; • Band + 1- 25 °: within this band, the relative height dome / target can be reported with an accuracy of 5m, for an error on the distance less than 10m.

Advantageously, the invention makes it possible to terminate guarding 24/24 for sensitive sites, because it meets the needs of televideosurveillance of sensitive sites during periods of reduced activity: • sites sensitive and of vital importance as industrial complexes, the airports, ports, railway stations, production sites, storage or distribution of energy, hospital infrastructure ...; • open and complex spaces such as a busy street, a public space, a shopping center, an airport hall, a harbor entrance, a place of interconnection between different modes of transport ...; [080] The HMI should be the showcase for a new technology designed to change the thinking that security system designers, security managers and remote control center operators make about televideosurveillance. unable to effectively monitor sensitive sites through low bit rate networks. [081] The invention makes it possible to ensure operational missions of intrusion monitoring and stakeholder control until the elimination of threats through stages of reception, qualification, alarm processing. [082] In connection with the invention, new working methods will redefine the limits of responsibility and skill levels of operators: it is easy to see the emergence of two types of operators within the same station: those in charge of the qualification of the alarms and those in charge of the analysis and the follow-up of the threatened sites. [083] In the invention, the hypertracking televideosurveillance is to follow a human or non-human target (vehicle) circulating inside a secure space by CCTV cameras. The cameras can be indifferently: - cameras with the PAL standard operating in the visible spectrum and giving images in color or black and white (cameras visible day / night); - Megapixel cameras visible day / night. - domes visible day / night; - thermal cameras. [084] In the case of a thermal or visible dome, hypertracking can work in conjunction with: autotracking, an operation of automatically driving the dome to maintain a target in its field of view; - manual tracking, identical operation, but performed by an operator; 25 - autotracking by virtual PTZ, operation of extracting a video from a Megapixel camera, a video centered on the target in lower resolution. [085] We will be particularly interested in the hypertracking of a single target, circulating in a space and at times when there is a low human activity, that is to say a small number of people or vehicles allowed to circulate and which must not be autotracking, hypertracking or autotracking by virtual PTZ. The target was previously automatically selected by the system or designated by an operator. In general, the fields of view of the domes and cameras can overlap or be disjoint: a moving target can, at a given moment, be autotracked by zero, one or more cameras. We will focus on the hypertracking of very large sites with a low density of domes and fixed cameras. Figure 30 shows an extended site with a central activity area. The intrusion is detected in perimetry by a network of thermal detectors. [086] The intruder is detected at the perimeter crossing by a thermal detector and confirmed by the video sensor of a thermal dome: it becomes the target. [087] It is autotracked by the thermal dome to the periphery of the zone of activity. The relay is then taken successively by visible domes, then by fixed cameras. [088] The cameras must ignore the persons and vehicles already present on the site before the intrusion. [089] The alarm channel is a single video source constituted by the linking of several video sources pursuing the same target. When there is recovery of the camera field, the alarm channel can be synthesized from different cameras for the same tracked target. The ideal alarm channel is one that optimizes certain criteria, for example: - size of the maximum target in the image; - visible camera rather than thermal, the day, the opposite, the night; - color camera rather than black and white, the day, the opposite, the night; - Megapixel camera rather than standard camera, the day, the opposite, the night; - dome rather than fixed camera; - minimize the number of camera changes in the channel; When there is no recovery of the field of the cameras, the target can be out of field of vision during periods more or less long. Domes can, for example, slowly scan an area to find the target. Hypertracking is then often a probabilistic operation, that is to say that in a given space time, several cameras can be candidates for the resumption of the tracking of a target out of the field of the cameras, without any guarantee that it is the right target. The hypertracking engine will make a choice in order to synthesize the alarm channel, this choice being able to be erroneous (recovery on a bad target) or irrelevant (choice of a video resource revealing itself less well placed than another with regard to the trajectory followed by the target). [090] The tree of multiple choices that could have achieved the hypertracking engine is the set of candidate channels. [091] Tele-video surveillance consists in exploiting the video resources of a remote site after receiving one or more alarms at the remote surveillance center. It takes place in three phases: 1. verification of the validity of the alarms (removal of doubt) and elimination of false alarms; 2. evaluation of the nature and evolution of the threat; 3. Information and management of field intervention teams. [092] Phase 1 is ideally handled by sending a video sequence of the triggering event to the central monitoring station and then analyzing it by the operator. This first operation is essential, but it differs from Phase 2, which consists of observing live site cameras. Typically, this phase can be delayed from several seconds to several minutes. [093] This poses two fundamental problems: How to find the intruder and what did the intruder do since his intrusion? [094] As illustrated in FIG. 31, in the absence of any hypertracking automatism, the operator supports the location of the target on the various video resources available on the site. The number of potentially affected cameras on a site is growing rapidly with the time taken since the intrusion. However, the operator is most often forced by the bandwidth of the interconnection link with the site giving him the opportunity to mount a limited number of simultaneous cameras. [095] The behavior of the intruder between the intrusion on the site and its live follow-up by an operator, is available only by consulting the recordings of the video resources of the site. The interest of hypertracking and the alarm channel can easily be seen, since in theory the live observation on the alarm channel prevents the operator from searching for the intruder on several cameras. In addition, reading the alarm channel record allows it to fill its monitoring gap between the intrusion and the live target tracking, as shown in Figure 32. [096] Hypertracking is currently a mobilizing topic many 20 resources in research and development, as evidenced by the document US7450735. [097] The variety of types and quality of video resources (visible, thermal, color, black and white, calibrated, uncalibrated cameras) makes it extremely complex to automate what would already require training for a supervisor who is perfectly familiar with the site. , the location of the available video resources, the current uses on the site, as well as the different operating modes of the previous intrusions. [098] In particular, in intelligent video processing, the recognition of one target (signature) among several is often restricted within the field of the same camera, with relatively low target shadowing times. The transmission of a signature between two cameras is today in the field of research and seems to be limited to cameras whose fields overlap. [099] When the fields of the cameras do not overlap, the simplest track to track multiple camera targets in cameras is to define a repository common to all cameras, to postpone the position of the targets on this repository and to make assumptions based on probabilistic optimization of the paths. In all cases, the proposed methods can be relatively expensive in computing time (image analysis, neural networks or Bayesians) so that the added value of hypertracking in terms of televideosurveillance can be neutralized by the hardware and software costs to implement to achieve it. The approach of the invention is an approach in terms of the result to be achieved, involving the human resource (the operator) in a semi automated process of reasonable force. By reasonable force we mean: - inexpensive in hardware and software resources; - easy to set up by the technicians performing the installation; - likely to be wrong, but giving the opportunity to the operator to make up for mistakes; - easy and fast to use by an operator; - to achieve the expected result: location of the target and vision of its trajectory within a time compatible with the arrival of the intervention; - compatible with the IP video surveillance software of the market. At the algorithmic power, the invention proposes to substitute a pragmatic approach based on: - and the professional design steps APSAD (R31 for remote monitoring, R81 for the alarm installation, R82 for the installation of CCTV an approach likely to fall within the scope of the draft Afnor interoperability standard External detection techniques (passive or active detectors) prove to be sources of numerous false alarms or conversely, when the detection thresholds are too hard, poor intrusion detectors Doppler microwave technologies have been virtually banned for their inability to control the detector's range of action, sometimes leading to justified detections, but for targets outside the monitored limits. Active infrared barriers trigger unexpectedly when vegetation cuts beams, or does not trigger when crossing the beams very quickly. Outdoor IR sensors are very sensitive to animals passing in the first meters of their field of detection. The dual technology detectors are inefficient when the movements are perfectly tangential (only the infrared triggers) or radiant (only the microwave triggers). Videosensor technologies have been ruled out in recent years due to the high rate of false alarms. They tend to come back in the news, under the name of intelligent video surveillance with the emergence of more robust algorithms and also in association with thermal cameras. However, the different operators encountered on this market do not guarantee an average false alarm rate lower than one alarm per camera per day. Often, it is necessary to model in 3D the field of vision and to choose wisely treatment filters in order to obtain good results. This parameterization work is not within the reach of all technicians, average detection profiles are precalibrated and the mechanisms of scene modeling simplified to the extreme. We can then arrive at cases of non-detection of real targets, without possible corrective action by the technician. In Sony products, there is an embedded intelligent detection, based on the segmentation of the target contour and the coherence of movements thereof on at least 15 successive images. The settings are simple, of type object size and detection threshold applicable to some detection windows defined in the image. Fast-oscillating objects are filtered, but not those in slow oscillation. The Sony approach does not require prior modeling of the scene viewed. The detection is also applicable to different 360 ° prepositions of domes with the possibility of adjusting the windows and detection thresholds preposition by preposition. Currently, Scutum uses passive infrared or dual technology type sensors, associated with Sony domes with intelligent analysis. The sensors are located in the field as close as possible to potential intrusions. They deliver pre-alarm information and force focus of the domes towards the event. If the intelligent detection embedded in one of the domes confirms the intrusion in the minutes following the pre-alarm information, a confirmed alarm is delivered. The event is filmed a few seconds and the resulting video clip is transmitted by remote monitoring for final analysis by the operator. With this technique, we obtain a false alarm rate of the order of a false alarm per week and per dome, without lowering the detection level of the site. The use of Sony domes is, however, limited to the visible range, which requires proper lighting of the monitored areas. [0108] The following decision-chain principle will be retained: - alarm triggering by a first device: pre-alarm; - Prepositioning a second equipment to the trigger zone of the first equipment; opening a space-time confirmation window on a second piece of equipment; - if detected: • alarm confirmed; • video clip or photo of the event. In order to understand the difficulty of achieving the hypertracking of a target designating an intruder moving in a sector where there is human activity, we will take the example illustrated in the following figure. A building is monitored externally by two angle domes, and indoors by fixed cameras. Paul and Sophie are employees. The shaded areas are supervised offices in which Paul and Sophie do not have to circulate. The cameras must follow in the alarm channel according to the following sequence: d1 G-cl -c2-c3-c4-c3-c5-d2G. - dlG pre-alert d2D, cl, c3 and dl D (see first case): cl confirms; - c pre-alert c2, c3, d2D and dlD or dlG: c2 confirms, but c3 can confirm via Paul or Sophie (see second case); - c2 pre-warning cl, c3 and d2D or d2G (see third case): c3 confirms; - c3 pre-alert cl, c2, c4, c5, dl G or d1 D, d2D: c4 confirms, but c5 or dl D can also confirm via Paul or Sophie; note that the intruder left c3 to re-enter (see 4th case); - c4 pre-alert d2G, c3, c5: c5 confirms but c3 can also confirm via Paul or Sophie - c5 pre-alert c3, c4, d2G and dlD: c3 confirms, but c4 can also confirm via Paul or Sophie; dl D can also trigger via Paul or Sophie coming out through the entrance (see 5th case); - c3 pre-alert cl, c2, c4, c5, dl G or dl D, d2D: c5 confirms, but c4 or dl D can also trigger via Paul or Sophie; - c5 pre-alert c4, c3, d2G and dl D: d2G confirms, but c3, c4 and dl D can trigger via Paul or Sophie. In the first case shown in FIG. 34, if dlG remains active, nothing can pre-initialise the dome to send it to dlD. Several solutions: - the dome dl can cycle between dlG and dlD: in this case, the proximity of the dome of the window gives the intruder the possibility to leave the field of vision of the dome before it returns to dlD ; - Pre-alerting by an opening contact on the exit window to dl D: but Sophie could leave the window open; - we add a camera in the room just before the exit. We place a sensor on the window towards dl D which pre-alerts dl D; - We replace dl with 2 fixed cameras In the second case shown in Figure 35, the camera cl pre-alert c2 and c3; It can be confirmed via Paul or Sophie. - If Paul confirms in c3, coming from c4 or c5, he could be eliminated because present on the stage before the intrusion in cl. - If Sophie triggers c3, the same reasoning is not applicable because her office is not under camera. Take into account the door through which it enters c3 is not relevant, because the intruder can also confirm in c3 using this same door. - This case of indecision can only be treated by eliminating Sophie by comparison with the intruder seen in cl: but be careful, Sophie may have a signature close to the intruder [0113] In the third case in Figure 36, the c2 pre-alarm camera c3, d2D and d2G. Several solutions: - the dome d2 can cycle between d2G and d2D: in this case, the proximity of the dome of the two windows gives the intruder the possibility to leave the field of view of the dome before it returns to the window Release ; - we set 2 sensors on c2: one on the window towards d2D and one on the window towards d2G. • the sensor to d2G pre-alert d2G; • the sensor to d2D pre-alert d2D. In the fourth case illustrated in FIG. 37, this case is of the most complex. The camera c3 pre-alerts cl, c2, c4 and c5, d1 D and d 1 G. The camera c4 confirms, but the problem is that the intruder has left c3 to re-enter. If the intruder has left for a very long time, he may have forgotten the signature of the intruder or else, when leaving c3, the intruder may have changed his appearance.

The camera c3 must therefore recover the intruder, which means that c3 must pre-alert itself. Paul or Sophie must not trigger c3. If Paul and Sophie are in c4 and c5 when the intruder leaves c3, they can be eliminated if they enter c3. We see, however, that if a third person enters the door, he can be mistaken for the intruder. By placing sensors on the doors of the hallway, the doors of c4 and c5 and the front door can be eliminated. In the fifth case illustrated in Figure 38, the camera c5 pre-warning c3 and dl D: the camera c3 confirms but also dl D via Paul coming out.

At night, dlD may have switched to black and white while indoor cameras are in color. The shooting of the dome is much more general than the camera. The recognition of Paul between c3 and dl D is complex (it is also complex by an operator). Recognition of the intruder between c5 and c3 is more likely, but again, the intruder is at the end of the aisle on c3.

The solution is to place a sensor on the windows of c5 and to pre-alert dl D only if this sensor is activated. The first case has shown that it may be necessary to modify the device in order not to lose an individual: to add a camera or to replace a dome by a fixed camera. The second case showed that recognition of the signature of the intruder may not be sufficient, when an unknown person entered the field of camera triggers. [0118] The third and fifth cases have shown the need to place sensors on windows leading outward when the intruder can be resumed by a dome or a fixed camera. The fourth case showed that a camera could pre-itself itself, that an intruder could change the signature by entering the field of the camera again, that it was possible to eliminate the active targets in the camera. a sector and that limiting the possible paths by sensors on the doors could restrict the assumptions, [0120] It is necessary to add the camera c6. The installation of sensors on the points of passage from one camera to another limits the number of possibilities as illustrated in FIG. 39. The tracking of the individuals already present in an area before an early warning is triggered makes it possible to eliminate their own movements to trigger a camera, which amounts to neutralizing the sectors around them. This can be extended to the movements of recognized individuals moving between several cameras with overlapping fields, but also, for cameras whose fields do not overlap, to detect their passage from one camera to another in order to continue to neutralize their movements. In an example shown in Figure 40, Sophie was neutralized because present in the building before the intruder enters. His movements are now tracked but do not trigger an alarm. Sophie leaves her office in C6 to go to that of Paul in C5, through the corridor in C3. When she comes out of her office, she activates c6-sensor 4. One arms c3-sensor 6 of the corridor that confirms. [0121] When she leaves the hall to go to Paul's office, she activates c3-sensor 4. One arms c5-sensor 3 from Paul's office that confirms. [0122] Now that we have eliminated Paul and Sophie, let's take the course of the intruder. In Figure 41, the intruder is autotracked by dl. Three sensors are positioned on the building's exits. Like a dynamic mask, the sensors follow the movements of the dome. It is possible to postpone the sensors on the panoramic image of the dome and detect the intrusion of the intruder into these sensors, then its disappearance. In Figure 42, the sensor 1 gives Sophie's desk in c6. The sensor 2 15 is on the desk without a camera. The sensor 3 overlooks the office under surveillance of cl. - Detection on d1-sensor 1 c6-sensor weapon 2 (Sophie's office). - Detection on d1-sensor 2 (desktop without camera) weapon cl-sensor 5, c3-sensor 7 (corridor) and c6-sensor 3 (Sophie's office). 20 - Detection on d1-sensor 3 c1-sensor 1 which confirms. In Figure 43, the intruder is in the office monitored by and. - Detection on c1-sensor 1 weapon dlG. - Detection on c1-sensor 2 d2D weapon. - Detection on c1-sensor 5 weapon c3-sensor 7 (corridor) and c6-sensor 3 25 (Sophie's office). - Detection on cl-sensor 4 weapon c3-sensor 1. - Detection on c1-sensor 3 weapon c3-sensor 8 and c2-sensor 5 which confirms. [0125] Now that we have eliminated Paul and Sophie, let's take the course of the intruder. The intruder is autotracked by dl. Three sensors are positioned on the building's exits. Like a dynamic mask, the sensors follow the movements of the dome. Note: it is possible to postpone the sensors on the panoramic image of the dome and detect the intrusion of the intruder into these sensors, then its disappearance. The sensor 1 gives Sophie's desk in c6. The sensor 2 is on the desk without a camera. The sensor 3 overlooks the office under surveillance of cl. - Detection on d1-sensor 1 c6-sensor weapon 2 (Sophie's office). - Detection on d1-sensor 2 (desktop without camera) weapon c1-sensor 5, c3- sensor 7 (corridor) and c6-sensor 3 (Sophie's office). - Detection on d1-sensor 3 c1-sensor 1 weapon that confirms. [0126] The intruder is in the office monitored by cl. - Detection on c1-sensor 1 weapon dl G. - Detection on c1-sensor 2 weapon d2D. - Detection on cl-sensor 5 weapon c3-sensor 7 (corridor) and c6-sensor 3 (Sophie's office). - Detection on cl-sensor 4 weapon c3-sensor 1. - Detection on c1-sensor 3 weapon c3-sensor 8 and c2-sensor 5 which confirms.

Then, as illustrated in Figures 44 and 45: - Detection on c2-sensor 4 c3-sensor 2 which confirms. - Detection on c3-sensor 7 c6-sensor 3 weapon that confirms. - Detection on c6-sensor 4 c3-sensor 6 weapon that confirms. - Detection on c3-sensor 3 c4-sensor 3 weapon that confirms. - Detection on c4-sensor 2 weapon c5-sensor 4 which confirms. - Detection on c5-sensor 3 weapon c3-sensor 4 which confirms. Note that the motion detection algorithm must be sufficiently powerful to successively detect an appearance and then disappearance of the intruder via the same sensor. - Detection on c3-sensor 4 c5-sensor 3 weapon which confirms - Detection on c5-sensorl weapon d2G sensor 3 which confirms. The peculiarity of the thermal dome is that its small opening angle coupled to the autotracking of the target center makes it possible to follow it on a plane by simply postponing the intersection of the optical axis on the maintaining the target curve level of the site. This is a little less true for visible domes whose angular aperture can reach 60 °. However, when autotracking works and the image, we can consider that this measurement is sufficient to locate the target on a plane, as shown in Figure 46. Each dome is associated with a panoramic image of his vision 360 ° in equirectangular projection. This image allows the calculation of the panoramic vision increasing the real-time vision of the camera.

Real-time vision is a floating window on panoramic vision. The two windows can be synchronous or not: the operator may need to circulate in the pan while keeping an eye on the target, then ask for the resynchronization of the pan on the real-time window, as shown in Figure 47. In the figure 48 illustrates the equirectangular projection projects the latitude (pan) on the x-axis and the longitude (tilt) on the y-axis. The red segment E on the image below corresponds to a circular arc centered on the thermal dome. The detection of crossing this segment by the intruder can be used to trigger an alarm. In FIG. 49, in the case of the thermal dome towards the thermal dome, the thermal domes cover considerable distances (up to 600m) and, at night, they are more effective than the visible domes. They are intended to monitor common areas in the periphery of buildings. When a target has been locked by a thermal dome, it can be lost by the latter for 2 reasons: too large distance of the dome and target hidden / hidden on the site. The mediator of the line connecting the two domes is the line of equal distance from a target to the two domes. This line can be adjusted to the buildings and different masks present on the site. When the trajectory of the target crosses this line, the second dome is theoretically better positioned than the first to follow the target; the size of the target (in number of pixels) must be larger. If one switches from one dome to the other crossing the line of equal distance, there can be a phenomenon of perpetual tilting from one dome to another when the intruder journeys along. line. In fact, as long as the first dome continues to track the target, there is no point in switching to the second dome. On the other hand, the second dome must be ready in case the first loses the target. We therefore propose the following operation: - Detection of the crossing of the line of equal distance. - Focusing the second thermal dome towards the target. - Waiting 15s of confirmation by the videosensor -If confirmation, locking the target and autotracking; -3 If no confirmation, focus towards the last raised position and wait confirmation 15 seconds, and so on. - Failover of the second dome in the alarm channel if the target is lost by the dome 1 or size less than a threshold to be determined [0134] In the case of the thermal dome towards the visible dome, with partial recovery, it is possible to project, for each visible dome, a hanging circle whose radius is a function of the zoom power of each dome. This radius takes into account the most favorable case (the day). At night, depending on the quality of the lighting, the limit of attachment is no longer a function of the dome but of the range of the lighting. In our example, the limit follows the closing of the site. By day, we propose the following operation: - Detection of the crossing of the hanging line of the dome; - Focusing of the visible dome towards the target; - Waiting 15s of confirmation by the video sensor: - If confirmation, locking the target and autotracking, -3 If no confirmation, focusing towards the last position raised by the thermal dome and wait confirmation 15s, so on, - Tilting of the dome visible in the alarm channel if the target is lost by the thermal dome. At night, the operation is identical, but based on the detection of crossing the lighting limit: - Detection of crossing the lighting limit. - Focusing the visible dome towards the target. - Waiting 15s of confirmation by the video sensor: confirmation, locking the target and autotracking -) If no confirmation, focus towards the last position raised by the thermal dome and wait confirmation 15s, so on. - Failover of the visible dome in the alarm channel as soon as the target is lost by the thermal dome. In the case of the thermal dome towards the visible dome, without partial recovery, FIG. 50, as long as the thermal dome retains its target, it remains at the alarm channel. When the thermal dome reaches the range limit, its probability of returning to its target is low and the probability of returning to a target different from the original target becomes significant. It is then possible to initiate a recovery procedure by the visible domes or domes closest to the detection of the crossing of the thermal dome catch limit. We propose the following procedure: - Detection of the crossing of the line of attachment of the thermal dome. - Trigger on the visible dome of a slow scan of a limited area centered on the last known position of the target. - Failover of the visible dome in the alarm channel if the target is lost by the thermal dome and taken back by the visible dome, if no thermal dome has hung the target. In the case of the visible dome towards the thermal dome without partial overlap, the return to the thermal dome (s) must be made as soon as possible, as soon as the target leaves the zone defended by the visible domes and a thermal dome can locate the target. When the target reaches the limit of range of a visible dome, we can trigger a recovery procedure by the nearest thermal dome (s). By day, as long as the target can be tracked by the visible domes, it is more interesting to stay in the visible. At night, it is the opposite: it is more interesting to return to the thermal. During the day, we therefore propose the following operation: - Detection of the crossing of the hanging line of the dome. - Triggering on the thermal dome of a slow scan of a limited area centered on the last known position of the target. - Failover of the thermal dome in the alarm channel if the target is lost by the visible dome and no visible dome is candidate. At night, we propose the following operation: - Detection of the crossing of the lighting limit of the visible dome. - Triggering on the thermal dome of a slow scan of a limited area centered on the last known position of the target. - Switching of the thermal dome in the alarm channel as soon as the target is lost by the visible dome. In the case of the visible dome to the thermal dome, the return to the (s) thermal domes (s) must be made as soon as the target leaves the area defended by the visible domes. By day, as long as the target can be tracked by the visible domes, it is more interesting to stay in the visible. At night, it is the opposite: it is more interesting to return to the thermal. By day, we therefore propose the following operation: - Detection of the crossing of the hanging line of the dome. - Focus of the thermal dome towards the target - Waiting 15s of confirmation by the video sensor: confirmation, locking the target and autotracking, - ~ If no confirmation, focusing towards the last position raised by the visible dome and wait confirmation 15s, so on, - Failover of the thermal dome in the alarm channel if the target is lost by the visible dome and no visible dome is candidate [0141] By night, we propose the following operation: - Detection of the crossing of the lighting limit. - Focusing of the thermal dome towards the target - Waiting 15s of confirmation by the video sensor:) If confirmation, locking the target and autotracking, -) 'If no confirmation, focus towards the last position raised by the visible dome and waiting confirmation 15s, so on - Switching the thermal dome in the alarm channel as soon as the target is lost through the visible dome. [0142] Case of the visible dome to visible dome with partial recovery of the fields of view. As for the thermal domes, the mediator of the line connecting the two domes is the line of equal distance from a target to the two domes. However, unlike thermal domes, with 60 degrees of opening, the report of the optical axis of a visible dome on the site plan is an inaccurate location: when crossing the line is detected, the intruder can be in a more or less wide area upstream or downstream of this line. By day, we therefore propose the following operation: - Target entry detection in the snap sector common to both domes. - Focusing of the second visible dome towards the target. - Wait 15s confirmation by the video sensor: - * If confirmation, lock the target and autotracking, no confirmation, focus towards the last position raised by the first dome and wait confirmation 15s, so on, -) If loss of target, focus towards the last position raised by the first dome and wait confirmation 15s, so on. - If the first dome loses the target and the second hangs: tilting of the second dome to the alarm channel. - If the two domes lose the target: the first one that resumes it switches to the alarm channel; if a thermal dome has hung the target, and no visible dome is a candidate, it switches to the alarm channel. At night, we therefore propose the following operation: - Target entry detection in the snap sector common to both domes. - Focusing the second visible dome towards the target. - Wait 15s of confirmation by the video sensor: -If confirmation, lock the target and autotracking, --If not confirmation, focus towards the last position raised by the first dome and wait confirmation 15 seconds, so on, 25 -) 'If the target is lost, focus towards the last position raised by the first dome and wait confirmation 15 seconds, and so on. - If the first dome loses the target and is hooked by another dome, tilting of this dome in the alarm channel with priority to thermal domes. - If the 2 domes lose the target: the first one that resumes it switches to the alarm channel, with priority to the thermal domes. [0145] In FIG. 51, represents the visible dome towards visible dome without overlapping fields of view. When the target crosses the hanging limit of a dome, the best-positioned domes must focus towards the last raised position and wait for the video sensor to go off. By day, we therefore propose the following operation: 15 - Target exit detection of its attachment sector towards the second dome. - Focusing the second visible dome towards the target. - Wait 15s of confirmation by the video sensor of the dome 2: confirmation, locking the target and autotracking, 20 non-confirmation, focusing towards the last position raised by the first dome and wait confirmation 15s, and so on, loss of the target, focusing towards the last position raised by the first dome and waiting confirmation 15s, so on. - If the first dome loses the target and the second hangs: tilting of the second dome to the channel. - If the two domes lose the target: the first one that resumes it switches to the alarm channel. At night, we therefore propose the following operation: - Detection of the target's exit from its attachment sector towards the second dome - Focusing of the second visible dome towards the target - Waiting 15s of confirmation by the video scanner of the second dome: -? If confirmation, lock the target and unconfirmed autotracking, focus towards the last position raised by the first dome and wait confirmation 15s, so on,) If loss of the target, focus towards the last position raised by the first dome and wait confirmation 15s, so on. - If the first dome loses the target and is hooked by another dome, tilting of this dome in the alarm channel with priority to the thermal domes. 20 - If the 2 domes lose the target: the first one that resumes it switches to the alarm channel, with priority to the thermal domes. Figures 52 and 53 show the dome to fixed camera. We are interested in the camera 1. To enter his field of vision, an intruder can come (with direction of crossing): 15 25 - from a building exit (sensor 1); - the left fence (sensor 2); - from the other side (sensor 3), and - from the back (sensor 4). If the target is autotracking on the left or right thermal dome: - Detection of the crossing of the lighting limit. - Depending on the crossing point: arming sensors 2 and 4 or arming sensors 2 and 3. - Wait for 10 minutes confirmation by one of the sensors - If confirmation, switching the camera in the alarm channel, -> If no confirmation, disarming of the sensors [0150] If the target is in autotracking on the visible dome opposite: - Detection of the crossing of the dome snap zone limit towards the camera. - Arming of the sensor 3. - Waiting 10 mn of confirmation: - If confirmation, switching of the camera in the alarm channel, - If no confirmation, disarming of the sensor. If the target is autotracking on the rear visible dome: - Detection of crossing the dome catchment limit of the dome to the camera. 20 - Arming the sensor 4. - Waiting 10 mn of confirmation: -) If confirmation, switching of the camera in the alarm channel, -If not confirmation, disarming of the sensor. Figures 52 and 54 show the following case of the dome to fixed camera. We are interested in the camera 2 whose field of vision overlaps that of the visible dome. [0153] The visible dome has hung the target: - Arming the sensor 3 as the target is locked by the dome. - Wait for confirmation of the sensor: confirmation, followed by the target in the field of vision of the camera (autotracking) - If the dome loses the target: -) Switching the camera in the alarm channel. [0154] The thermal dome locks the target: - Detection of the crossing of the lighting limit. - Focus of the visible dome towards the target and arming of the sensors 2 and 3 of the camera: -) If the visible dome locks the target, it autotracke -) If the camera locks the target, it autotracke - Si the thermal dome loses the target, tilting of the dome in the alarm channel, if it still has the target, the camera in the opposite case, even if it no longer has the target. - If the camera loses the target, arming sensors 2, 3 and 4 of the external cameras 1,2 and 3 for 10 minutes (example of recovery procedure). - Disarming the sensors after 10 minutes if the target was not picked up by any camera. In the case of the fixed-to-dome camera, illustrated by FIGS. 52 and 55: • Camera 1 - Output of the camera field via the sensor 4: Positioning of the thermal and visible domes towards the camera, - Waiting for confirmation 10 minutes , ~ Autotracking of the target by any dome locking the target, Insertion to the alarm channel in the following order of priority: thermal dome, visible dome - Output of the camera field via the sensor 2: -Positioning the thermal domes and the visible dome facing the camera, towards the camera, -) 'Waiting confirmation 10mn, -3 Autotracking the target by any dome locking the target, Insertion to the alarm channel in the following order of priority: thermal dome, visible dome. - Output of the camera field via the sensor 3: -) 'Positioning of the thermal dome and the visible dome facing the camera, towards the camera, -> Waiting confirmation 10mn, -> Autotracking of the target by any dome locking the target , -) Insertion in the alarm channel in the following order of priority: thermal dome, visible dome. In the case of the fixed camera towards fixed camera, illustrated by FIGS. 52 and 56: [0157] We are interested in the cameras 2 and 3. - Output of the field of the camera 2 via the sensor 4: Arming of the sensor 3 of the camera 3, -Waiting confirmation 10mn, -Insertion of the camera 3 to the alarm channel if no dome is a candidate. - Passing of the sensor of the camera 3 towards the camera 2: Arming the sensor 4 of the camera 2, - Waiting confirmation 10 minutes, - Inserting the camera 2 to the alarm channel if no dome is a candidate. [0158] Let's follow an individual from the property limit inside a building: - The sensors are permanently active or temporarily deactivated (entry of the site on or any point of detection, until authorized access, for example) . Activating a sensor pre-alerts a thermal or visible dome that confirms and self-tests the target. - The individual's journey is reported both on the panoramic image of the dome and on the site plan. Virtual dams are included, symbolizing the clashing areas of neighboring domes and fixed cameras. - The virtual crossing of a hooking area by the intruder pre-alerts one or more domes, which confirm, hang and self-tap. Each dome must confirm on its snap area only: the sensor zone should not be extended to the entire image in order to avoid confirming on site staff. Therefore, if the snap does not occur in the 15s, a new set point is given to the dome that tries again to detect the target. - When the thermal dome loses the intruder, the hypertracking motor designates the dome that will take over from the thermal dome in the alarm channel. - If the initial dome loses the intruder and no other dome follows the intruder, he must pre-alert himself (he can find his target later) and other domes whose attachment sectors are the same. more likely. One can imagine that a motion-scanning zone scan mechanism is automatically set up on the nearby thermal dome and neighboring domes to maximize the chances of finding the intruder. - When a visible dome hangs the intruder, it autotraque. It is clear at this level that if a person from the site is confused with the intruder, the chances of finding the intruder diminish. This is why, if autotracking is done in a wide shot, we keep a global view of the site and the possibility of recovering from an error. Moreover, we lose the opportunity to identify the target. But, if the operator takes the hand, he can decide to zoom on the target. - Domes and visible cameras can hypertrack the same individual, on fairly wide areas of attachment. As far as possible, individuals already present on the site before the arrival of the intruder must be identified and neutralized. They should therefore be transferred to the site plan. - When the intruder approaches a building, the dome pre-alarms the indoor cameras: we have seen that the positioning of sensors on the opening or passage points required by indoor cameras was often necessary to restrict the choice of potential cameras. Only these sensors confirm for the cameras concerned. The software must therefore accept the dynamic arming and disarming of: - sensors, - sectors and zones of attachment, and - sensors on opening or forced points of passage. The software must be able to automatically neutralize the sectors around the people present on an area before an intrusion is detected and follow the movement of these people from fixed camera to fixed camera (no autotracking) to limit the number candidate channels. The software must be able to constitute video clips of initial intrusion confirmations and photos of each confirmation hooking, and make a choice when two confirmations create a possible tree of the alarm channel.

The chosen camera feeds the alarm channel and the candidate camera creates a new tree called the candidate channel. A priori, if the choices made are the right ones, a candidate channel is more likely, either to feed the alarm channel in the long run, or to die, ie the target tracked stabilizes on the site or , not being a real target, is lost by the camera. As shown in Figure 19, the operator must be able to view on the same window, the course of the alarm channel and the appearance of candidate channels. He must be able to view the snap confirmations photos. The operator has the alarm channel and the photos that mark each camera change of the alarm channel and each start of a new tree of a candidate channel. When it connects live on the site, it connects to the alarm channel and reread, in a window, the history of the alarm channel since the intrusion. Each change of video resource in the alarm channel triggers transmission of the associated thumbnail. [0166] By clicking on a thumbnail, the operator can view the image in its native format. By clicking on a candidate channel, the operator can replay the corresponding video record or switch the live video source into the augmented reality engine. There are two possibilities where to situate the hypertracking: decentralized hypertracking, and centralized hypertracking. [0168] Decentralized hypertracking (cooperative hypertracking): like the players on a football field (the cameras), the best positioned player automatically takes the ball (the target) in agreement with his teammates. This assumes a collective intelligence of domes and cameras and a rule allowing the best-positioned camera to fit into the alarm channel.

Hypertracking seems logically reserved for a centralized intelligence: the geolocation of targets seems to be a good reason. But before ruling out the hypothesis of a collective hypertracking, let's push the reasoning to the end. Let's take up the problem of the football match. A team of players prepositioned on the field (but with some latitude of movement) pursues a ball whose trajectory is not mastered by the group since the opposing team can take control. The ball is the official target. The marking of the opposing players is as many secondary targets as players who can be dangerous when approaching the goal. The behavior of the players is perfectly regulated by the intelligence of collective game: it is very rare that 2 players of the same team compete for the ball. [0170] Now back to our domes and cameras pursuing an intruder and review the features expected by the remote surveillance. [0171] Alarm channel: Each camera or dome that considers itself best positioned relative to the official target automatically inserts into the alarm channel, based on a collective intelligence that settles the confits. For simplicity, one can imagine, for example, a multicast channel dedicated to the alarm (such as a TV channel) on which each camera pursuing the target 20 emits successively, in a concerted manner with others. The central monitoring station must be able to re-read the alarm channel, from the detection of the intruder to the live view of an operator. The first solution is to use a local video recorder filming the alarm channel on site. The second solution is to allow each camera to keep its own recordings. The third solution is to register the alarm channel at the central station. This solution seems the simplest because it does not imply additional material on the site. But technically, it involves an automatic connection of the central monitoring station to the alarm channel, before the doubt is raised by the operator, so before deciding whether or not the site must be followed by an operator. This solution therefore has the disadvantage of consuming incoming bandwidth before deciding whether or not the site will be processed. But it has an undeniable advantage: it is politically correct because it goes in the direction of the customer advantage. Indeed, the images of the site can be diffused on a screen in analysis room, likely to be visualized by an operator analyzer, even if it is not yet officially seized of the mission of surveillance of this site. Candidate Channels: The monitoring of the candidate channels by the remote surveillance center involves: the setting up and the broadcasting of parallel missions allowing several cameras to track an unofficial target; • their distribution on a candidate channel; • registration of cameras contributing to a candidate channel or registration of candidate channels; • access by the remote monitoring center to all potential candidate channels; • access by the central monitoring station to the successive recordings of the cameras contributing to the different candidate channels or to the recording of the candidate channels; When a camera hangs a target, it must report it to the central station. Indeed, the operator must be able to evaluate the track followed by this camera and the following cameras, in case of failure found on the alarm channel. The information from the central monitoring station by the camera may be the snap image itself or a link to this image transmitted as a thumbnail. The camera must then: • either keep itself the following images: the monitoring station will have to search them in the camera, and this, for each camera; • or transmit them on a candidate channel: the candidate channels must be recorded by equipment external to the cameras, either on site or at the central station (if the number is not too important) [0175] Assuming that each camera provides its own recordings, these must be attached and indexed on the candidate channel number. In the event that an external device manages the recordings, the following problem arises: it is not known, a priori, how many channels it will have to record. At worst, there should not be more candidate channels than cameras on the site (to demonstrate) and, for the monitoring center, the multivision of the candidate channels will remain less bandwidth consuming than the multivision of all the cameras of the site. site. The problem posed by the solution of collective hypertracking is that existing cameras can not today be taken into account in hypertracking. There will therefore be no short-term market for this technology. However, it goes in the direction of integration closer to the camera and certainly deserves to be studied closely. Centralized Hypertracking: Option 1: the hypertracking function is integrated into the video recording software of the site. Option 2: Hypertracking is a feature provided by site-independent equipment that accesses video resources: • Directly: Sends stream requests and decodes streams. • via the video recording system of the site (subscription to the resources of the recorder via SDK). Centralized hypertracking on the video recorder of the existing site. In this solution, it is considered that a software solution of the Milestone or Genetec type integrates a hypertracking module. The advantage for the end customer is: • no multiplication of equipment; no multiplication of flows on the network (a single stream for recording and image processing); • a priori, cost benefiting from wider dissemination of software. In this solution, we see three forces of algorithms (as Sony does in its software): standard level: applicable to most indoor cameras; advanced level: reserved for indoor and outdoor cameras and domes. The advanced level consuming more computing resources than the standard level, one can imagine that some of the hypertracking treatments are performed in the thermal and visible domes and external cameras. By way of example, the Sony domes and cameras embody advanced motion detection which can be exploited as an alarm by the Milestone and Sony software (not by Genetec) and in the form of Metadata by the Sony software ( not by Milestone). This explains the difficulty (or the lack of will) of the video surveillance software vendors when it comes to integrating a proprietary feature. The same problem arises in a more basic way for controlling domes with absolute pointing. Milestone and Sony use the absolute pointing of the Sony domes, not Genetec, which until now has refused to integrate this type of control too proprietary. In order to minimize the development effort of the integrators, Sony, Axis and Bosch have come together at the global level to harmonize their protocols and their metadata. The next version of the Sony RealShotManager software will recognize the Bosch and Axis cameras. Afnor has just launched a working group to develop a 3-year standard for interoperability in digital video surveillance. A priori, telemonitoring operators are not represented in this working group. To move in this direction, Scutum must convince at least one software integrator Milestone or Genetec type and probably a specialist in intelligent video detection. Hypertracking centralized on equipment other than the site recorder. There are two possibilities: • this equipment alone acquires and decompresses the video streams before processing: it is already necessary to eliminate this solution which requires a permanent work of integration of new IP surveillance cameras. Or limit it to a few generic JPEG and MPEG4 streams. • this equipment recovers uncompressed video streams via the site's registration system: this requires developing on the basis of the SDK of the software publisher. The possibilities offered by software publishers are different; Genetec, for example, works on the native video stream and manages the decompression, Milestone entrusts the SDK manufacturers decompression. ISS, does both, depending on the camera manufacturer. Genetec seems to be the safest way. This is the choice of Evitech, which recovers the flows via Genetec and returns the alarms from its own calculations to Genetec.

So you have to: convince an intelligent video detection specialist to integrate hypertracking into their own video analytics engine, on the one hand, and have an aggressive price policy for indoor camera licenses, on the other hand go ; to develop a proprietary engine strictly limited to hypertracking; The first path seems the most durable because likely to evolve with the progress of video intelligence. Another embodiment also relates to a multithreaded hypertracking, where each thread would be specific to a camera, trying to keep as much as possible the independence of each thread vis-à-vis a main thread. This would be to simulate on a centralized server the behavior of N cameras acting collectively. In order to ensure optimum operation of the invention the system comprises means for calibrating the rotation values of the domes. The first step of the calibration of the domes consists in aligning their rotation values with a common reference frame. [0190] There are two methods: the alignment of the domes on the infinite and the alignment two by two. In the context of the alignment of the domes on the infinite and as illustrated in FIG. 57, when a characteristic point situated at infinity (several Km) is visible from most of the domes of the site, the method the simplest is to take a first reference dome (dome 0), then: • measure the pan value of this point; • for each dome, align the value in pan of this point with the reference value. For the domes that can not visualize the same reference point, they must be aligned in pan on already calibrated domes with which they share a very distant characteristic point.

After calibration, the very far points should display identical pan values when they are pointed in the panoramic images of different domes. In the context of the alignment of the domes two to two t as illustrated in Figure 58, another technique of calibration of the rotation values is to align the domes 2 to 2: • Take a first reference dome; • A second dome is aimed at from the 1st dome; • We measure the value pan = p1; • With the second dome we aim for the first dome; • We measure the value pan = p2; • The rotation correction R on the 2nd dome is applied so that p2 = pl + 180 °. The other domes are thus aligned in the same way, choosing the more distant domes rather than the nearest domes. The second calibration step consists, for each dome, of entering their GPS coordinates (x, y and altitude h) in the configurator, as illustrated in FIGS. 59 to 61. The altitude values must be validated by the method consisting, for each dome, to point in the panoramic image the foot of other visible domes and check the value returned by the calculation: • h1-n1 = D tan (t1); • h2-n2 = D tan (t2). [0194] As far as possible, intermediate points on the right joining the two domes should be checked to be aligned with the other dome in each pan. The altitudes are computed using the previous matching method and found in the associated matrices of both panes and on the plane. [0195] Assuming that we know for each dome the positions and 5 altitudes (x, y, h). We can calculate the distance of the two domes: • D = root ((x2-x1) 2 + (y2-yl) 2) And the absolute angle R of the line joining the two domes, in the common frame of reference: • R = Atan [(yl-y2) / (x1-x2)] If we calculate the angles (pan) of the target relative to this line: • p'1 = p1-R • p'2 = p2-R We have: • dl cos (p'l) + d2 cos (p'2) = D • dl sin (p'1) = d2 sin (p'2) Hence for the non-zero values of p'2 (target not aligned with the domes): d2 = dl sin (p'l) / sin (p'2) • dl [cos (p'1) + cos (p'2) sin (p'1) / sin (p'2)] = D • dl = D sin (p'2) / [cos (p'1) sin (p'2) + cos (p'2) sin (p'1)] Let be: • dl = D sin (p ') 2) / sin (p'l + p'2) • d2 = D sin (p'l) / sin (p'l + p'2) [0196] When two domes aim at a target not aligned with the domes, its position is uniquely determined by the pan values of both domes. We also have for the non-zero values of t1 and t2 (target not aligned on the horizon): • d1 = (h1 - n) / tan (t1) • d2 = (h2 - n) / tan (t2 ) • dl cos (p'l) + d2 cos (p'2) = D Where: • (h1 -n) cos (p'l) / tan (t1) + (h2 -n) cos (p ') 2) / tan (t2) = D We put: • cl = cos (p'l) / tan (tl) • c2 = cos (p'2) / tan (t2) [0198] Hence the altitude of the target: [0199] • n = h1cl + h2c2ùD / (c1 + c2) [0200] By pointing the same point in two panoramic images, we can calculate a unique altitude value to be reported in the matrices associated with each panoramic . [0201] The association with each panoramic image of a matrix of altitudes is necessary to take into account: • the important elevation changes on the site • mounds, terraces, mounds, vats, roofs or terraces accessible by the intruders, when they are visible by one or more domes [0202] The whole of the site can not be raised. It is therefore necessary to focus only on the probable paths of the potential targets, paths along which the non-taking into account of the contour lines could lead to a detrimental geolocation error. [0203] What errors and what harm can be feared? Remember the objectives to achieve. [0204] The system must: • detect the target by a sensor • confirm it with a dome • hang it with the same dome [0205] The operator must then be able to: • monitor and evaluate the evolution of the target • choose the relevant video resources • control the intervention teams for the elimination of the threats [0206] The threats consist of: • the target itself • checkpoints on the trajectory of the target [0207] Let us study, for each of these steps, the data that will have to be entered into the system by the configurator. The best positioned dome must focus towards the sensor in alarm: it must learn its position and save it as a preposition (prerecorded PTZ value). The conventional method consists in controlling the dome in a configurator window by centering the sensor at the center of the image, and then associating with the sensor the preposition obtained by the get PTZ method. [0210] The transfer of the sensor position on the panoramic has, at first sight, no interest. No more than the geolocation of the sensor, vis-à-vis the operational missions of the operator. The value of the altitude of the sensors is therefore not necessary. If multiple domes can confirm the target, the same procedure applies. Note that only one dome will be broadcast on the alarm channel. The autotracking algorithm does not use the geolocation of the target, but regularly repositions the optical axis of the dome on the feet of the target (the lowest points of the segmentation). The configurator does not provide any value for this function. The evolution of the camera target camera is allowed by the hypertracking engine, as shown in Figures 63 and 62. A hypertracking engine error can be induced by a geolocation error. The hypertracking engine works as follows: • autotracking a target by dome 1; • transfer of its trajectory to the panoramic image of dome 1; • detection of crossing a limit stall / snap 20 at the panoramic image; • transmission of a setpoint to the dome 2 so that it aims at the indicated place. Or: • autotracking of a target by dome 1; 25 • report of its trajectory on the panoramic image of the dome 2; • detection of crossing of a stall / hook limit at the panoramic image of the dome 2; • focus of the dome 2 towards the target. In the first case, if the crossing point in the panoramic image of the dome 1 does not correspond to the point of focus of the dome 2, the dome 2 may not catch the target. In the second case, a significant error in the trajectory of the target in the panoramic image of the dome 2 can lead the dome: • not to trigger: the reported trajectory does not cut the limit of attachment while the trajectory real would cut it • to trigger unexpectedly: a sector is crossed, but when the dome focuses on the crossing point, it does not find the target. In order to reduce these cases of errors, it suffices to map the panoramic images of the domes taken two by two at the respective hooking / unhooking sectors. This matching is done by raising, in the hooking / unhooking sectors, a certain number of characteristic points (bitter) appearing in the two panoramic images. The configurator must therefore allow: • to designate a point of the panoramic 1; • to designate the corresponding point of the panoramic 2; • calculate the altitude n of this point; to enter this value in the matrices of altitudes of the two domes; • leave this point on a layer. As shown in Figure 63, from the dome 0 can be mounted on the dome 1 or the dome 2. [0218] At this level, it is considered to have achieved a minimum parameter to achieve the fixed objective. The curves must, however, be extrapolated by the calculation (hyperbolas, polynomials of degrees 3). [0219] Figure 64 represents it after extrapolation. [0220] The operator has 3 methods for inserting a control point (PTC): • point the precise location of the PTC directly in the operator geolocation channel; • point in the real-time image; • point in the panoramic image. Direct pointing in the operator channel is the simplest method when the location or object to be controlled is perfectly identified on the map. It has the advantage over the following methods of precise positioning of the PTC on the site plan. On the other hand, the transfer of the PTC on the panoramic image can be erroneous, because of the error on the altitude of the PTC. This method is not the most used, because the vision of the operator is focused on the intruder rather than the site plan. The pointing in the real image is more natural: the operator clicks on what he sees. The score is recorded on the underlying panoramic image. This method assumes that the target is not moving too fast, otherwise the PTC may exit the real-time image, the time that the operator responds. The pointing in the panoramic image then allows the operator to point (judged) the PTC. The pointing methods in the real or panoramic image have the advantage of accurately designating the PTC in the panoramic image, but the disadvantage of a report on the more or less precise plane, depending on the altitude accuracy of the PTC in the panoramic image. Indeed, we know that the closer we get to the horizon, the more the error on the altitude measurement leads to a significant error in the report on the plane. Similarly, when the intruder is in an uncalibrated or low altitude calibrated region. [0224] To fulfill its mission of eliminating threats, the operator must lead the speaker on the specific locations of the PTC. In the HMI, the site plans of the 2 geolocation channels are operational guides for the operator, indicating the relative positions intervening / target and intervening / PTC. [0225] There are two working hypotheses: the speaker is geolocated by GPS. its position is automatically reported in the geolocation channels; • The speaker is not equipped with GPS. [0226] To which are added the two following possibilities: • the intruder has already been eliminated when the PTCs are checked: the alarm channel is potentially available for the monitoring of the speaker; • the intruder is still on site when the PTCs start to be checked: the alarm channel is not available. [0227] In order to limit the situations, it will be considered that the alarm channel and the associated panoramic image must remain at the maximum available for the monitoring or the recovery of the intruder and must not be mobilized by the operator to follow the intervener, only when he needs the augmented vision to direct the stakeholder on the PTC. [0228] The pilot's control to a PTC is in an embodiment carried out by the operator, FIG. 65. To begin his PTC control mission, the operator clicks on the PTC in a geolocation channel. and displays the image of the dome that marked the PTC in the preview window, image centered on the PTC. [0229] To roughly control the speaker to the PTC, the operator must, in a first time, use the geolocation channels and give guidance to the speaker on the approximate location of the PTC. [0230] When the speaker is near the PTC, the operator is able to tell him precisely its location (this is the case when he pointed the PTC directly in the geolocalisation operator channel), or he can pilot the speaker once he sees it live in the preview window, or he must refer to the panoramic image to tell him precisely the PTC. In the latter case, the operator will have to hang the speaker with the dome. [0231] First case: the speaker is not geolocated. The operator must use the dome that marked the PTC to locate the speaker, hang it and autotracker. He has several techniques: • He sees the speaker in the preview channel: he then engages autotracking on the speaker as a target; • It knows the position on the plane of the speaker (according to the indications of the latter): he then points this position on the plane, which directs the dome towards the speaker and allows him to see it in the preview channel ; he then engages autotracking on the stakeholder as a target; • He starts an automatic zone scan; • It manually searches for the speaker by driving the dome in the preview channel, or exceptionally by mounting the dome in the alarm channel. We see that in the absence of a means of geolocation, the operator can spend time hooking the speaker. [0232] Second case: the speaker is geolocated. The GPS position of the speaker is plotted on the panoramic image of the dome marking the PTC, with inaccuracy relative to the altitude values. But once the intervener autotracked by the dome, its position reported in the panoramic image becomes precise, since it comes from the pan and tilt values of the dome. Its relative position with respect to the PTC thus becomes precise.

To hook the speaker, we will postpone its GPS trajectory on the panoramic image of the dome and detect the crossing of a limit of attachment. If the postponement of this trajectory is false, its intersection with the snap limit is false and the dome will focus next to the target.

Once the speaker hung by the dome in the preview window, the operator can switch the dome in the alarm channel to take advantage of the increased vision of the scene. [0233] The configurator must make it possible to cross the limits of daytime clashes of the domes with absolute positions (GPS) of the plane. It must therefore: • provide for a procedure for manual matching of the panoramic image with plan bitter • or, provide for an automatic matching procedure, for example by engaging autotracking on the intervention vehicle in the context of a learning procedure [0234] Both methods must be possible, for the good reason that, if during the installation phase, it is possible to envisage using the GPS of the intervention vehicle, afterwards technicians will no longer have this resource. The configurator must therefore: • designate a point of the panoramic 1 • designate the corresponding point on the plane • calculate the altitude n of this point • enter this value in the corresponding altitude matrix • leave this point on a layer [0235] Only the access routes used by the intervention vehicle must be calibrated, as shown in Figure 66. [0236] At this level, it is considered to have achieved a minimum parameterization to reach the target. objective set. The points must then be extrapolated to the left and right to take into account the widths of the tracks, then, thanks to image analysis algorithms, the segments must be connected together taking into account the characteristics of the site ( buildings, obstacles), as shown in Figure 67. [0237] The problem that image analysis will encounter in defining the hanging limit of a dome is the possible failure to take into account the natural masks of the site, which can lead to a spontaneous appearance of the intervener in the area of competence, without the limit of attachment being crossed. In the example shown in FIGS. 68 and 69, it is possible to go from zone 1 to zone 2 without crossing the hanging limit by taking advantage of the pole mask or by following the plant hedge. A solution would be to program the image analysis algorithm so that it takes into account these masks in the snap limit. Thus, in the example below, in both cases the intruder would be detected at the crossing of the new limit. [0239] In the absence of three-dimensional (3D) vision of the scene, the problematic seems complex, even on high resolution panoramic images. One can, however, consider the assistance of the human resource in the process of analysis. Thus, in the example below, the segmentation of the grid by the video analysis may be rejected by the technician setting the site because the target remains visible through the grid. [0240] The definition of the night snap limit must be defined from the night panoramic image, probably by image analysis assisted by the technician. Thus, it is understood that the invention is not limited to the embodiments described and illustrated. It is furthermore not limited to these exemplary embodiments and the variants described.

Claims (11)

REVENDICATIONS1. A method of monitoring one or more sites in a field of view of at least one image capture means connected to a remote station, the method comprising the steps of: - detecting the presence of a suitable event triggering capture of at least one image by the image capturing means towards said event; transmission of said at least one image to the remote station through a video stream, characterized in that said video stream is initially modified by the integration of reference elements adapted to the scenic content of each of the images composing this video stream so as to identify the cause of said event. 2. Method according to claim 1, comprising a prior step of acquiring at least one reference panoramic image in the equirectangular format of the field of view of at least one image capture means. 3. Method according to the preceding claim, comprising a step of transmitting said at least one reference panoramic image to the remote station. 25 4. Method according to claim 2, comprising a series of calculation steps, from said at least one panoramic panoramic image, of panoramic images in rectilinear format each corresponding to an enlarged vision of said scene with respect to the vision. of each image composing said video stream. 30 5. Method according to one of the preceding claims, comprising a series of steps of merging said computed panoramic images and images composing said video stream in order to constitute an augmented reality video stream restoring, on the remote station, a field of view. of said artificially expanded scene with respect to the initial video stream. 6. Method according to any one of the preceding claims, wherein said reference elements relate to one of the information relating to one of the following elements: designation and identification of the cause of the event; location of the cause of the event and / or - indication of the buildings and infrastructures of the site (s). 7. Method according to any one of the preceding claims, comprising a step of automatically tracking the movement of the cause of said event by at least two image capturing means, characterized by the following steps: tracking the displacement of the cause of said event by the image capturing means 1; - input detection of the cause of said event in an area of the field of view of said image capture means 1 or in an area of the field of view of an image capture means 2; focusing the image-capturing means 2 towards the cause of said event; confirmation of entry of the cause of said event in an area of the field of view of said image-capture means 2 and capture of at least one image of the cause of said event by said image-capture means 2, and tracking the movement of the cause of said event by the image capture means 2. 8. Method according to any one of the preceding claims, comprising a preliminary step of parameterizing at least one zone of detection in the field of view of at least one image sensor, characterized by the following steps: - association of an altitude matrix with said panoramic image making it possible to enter, for any point of the reference panoramic image, the altitude value at this point, marking on a layer superimposable on said panoramic image of said characteristic points of said zone; detection, - calculation of the altitude values at these characteristic points of said detection zone and entry of said values in the altitude matrix, and - complete manual parameterization on said layer of said detection zone, with the possible assistance of means of image analysis. 9. System for monitoring one or more sites situated in a field of view of at least one image-capturing means, for carrying out the method according to any one of Claims 1 to 8, comprising at least one at least one sensor associated with said at least one image capture means adapted to detect an event capable of triggering the capture of at least one image by the image capture means in the direction of said event so as to transmit said at least one image at the remote station through a video stream, characterized in that the system comprises processing means able to modify said video stream by integrating reference elements adapted to the scenic content of each of the images composing this video stream so as to identify the cause of said event. 10. System according to claim 9, comprising a unit for automatically tracking the cause of said event by the different means of capturing images of the same site. 11. System according to the preceding claim wherein the image capturing means is a thermal dome.