Designing a Site with Avigiln Self-Learning Vide Analytics Avigiln HD cameras and appliances with self-learning vide analytics are easy t install and can achieve psitive analytics results withut nging sftware adjustments. Avigiln's patented self-learning vide analytics autmatically adjust t the camera's field f view (FV) requiring n cnfiguratin r adjustment. Fr vide analytics t perfrm effectively, the analytics cameras (r cameras cnnected t an ACC ES Analytics Appliance) must be installed crrectly. Vide analytics enabled cameras must be: Within the height and angle guidelines. Level t the hrizn and grund plane (fr utdr r large indr areas). Installed where there is sufficient light in the area f interest with n bstructins. Within range f the area f interest fr the vide analytics t identify bjects. Mnitring a scene with enugh cntrast t classify bjects in the scene. Fr example, a persn walking in white clthes in a snw-cvered FV may prvide pr results. The fllwing infrmatin prvides a basic set f installatin parameters. Read thrugh the entire dcument befre installing cameras. Fr site requirements that deviate frm the listed recmmendatins, r when in dubt, cnsult with an Avigiln representative befre installing the cameras. General Guidelines 2 Analytics Lcatin Mde 3 Reflected Light 4 Adaptive IR 4 Lux n Target 5 Obstructins 5 Cverage Area 6 Object Velcity 7 Outdr Camera Placement 7 Self-Learning 8 Avigiln Appearance Search Feature 9 Designing a Site with Avigiln Self-Learning Vide Analytics 1
Fr Mre Infrmatin 11 General Guidelines In general, cameras shuld be installed accrding t the fllwing guidelines t achieve ptimal analytics perfrmance: Cameras must be able t see mving bjects in the field f view (FV) fr a minimum f 2 secnds. 5 secnds is recmmended fr ptimal bject classificatin. Cameras shuld be munted at a minimum f 2.8 meters (9 feet). Cameras can be tilted within 30 frm the hrizntal fr ptimal bject classificatin. Increasing the tilt angle can help in detecting targets that are directly appraching the camera. The camera shuld be tilted n mre than 45 frm the hrizntal. Cameras shuld be munted t a stable surface t minimize vibratin and mvement. Select a lens, munting height and tilt t capture the required level f detail fr classified bject detectin within the scene. General Guidelines 2
Camera FV must be level with the hrizn. Peple in the FV shuld be walking upright. Peple and cars mving parallel t the FV prvide better results than bjects mving t r frm the camera. Fr mre details related t yur particular type f camera r scene, see the related sectin. Fr mre details related t Avigiln Appearance Search feature, see Avigiln Appearance Search Feature n page 9. NOTE: Analytics n wide-angle r fisheye/panramic lenses are nt supprted at this time. Analytics Lcatin Mde In the Avigiln Cntrl Center (ACC) Client sftware, set the camera's Vide Analytics Cnfiguratin t use the Lcatin mde that best describes the scene: Outdr this ptin is suitable fr mst utdr envirnments. This setting ptimizes the camera t identify vehicles and peple. Outdr High Sensitivity nly use this ptin if yu require the system t be mre sensitive than the Outdr setting. This ptin is ptimized t run with higher sensitivity fr detecting peple and vehicles in challenging utdr scenes. Be aware that this ptin will generate mre false psitives. Large Indr Area this ptin nly detects peple and is ptimized t detect peple arund bstructins, like chairs and desks, if the head and trs are visible. Indr Overhead this ptin is ptimized fr cameras munted directly verhead and shuld nly be used when a trs cannt be seen in the camera FV. Any mvement is assumed t be human. It can be used in areas with limited space but with high ceilings, r t mnitr drs. It shuld nt be used with the Avigiln Appearance Search feature, r t detect peple traveling against the crwd. Analytics Lcatin Mde 3
Reflected Light Avid direct light surces. The camera may be temprarily blinded if bright light surces shine directly at the camera. Psitin the camera s that the sun, headlights r ther light surces d nt shine directly int the lens. Avid installing the camera in areas with drastic changes in lighting thrughut the day. Fr example an indr space with direct sunlight thrugh a skylight r large windws. Significant changes in lighting cause large shadws and different clring in the space. Such changes may generate incnsistent detectin results. Be cnscius f indirect light surces, including reflectins frm built-in r external IR illuminatrs, t avid lens flares and lss f cntrast in the image. Cameras with wide dynamic range (WDR) may be able t vercme this issue in sme instances. Avid mirrrs and ther reflective surfaces (like shiny flrs and ceilings). Reflectins may cause additinal false detectins. Adaptive IR Adaptive IR functins by adjusting the IR utput dynamically t prevent versaturatin in the scene as the light changes thrughut the night. Cameras using nly built-in IR fr illuminatin at night detect targets at a much shrter distance. Additinal illuminatin is required t cnsistently detect targets. Be aware that IR may als blur the utline f bjects and negatively impact the accuracy f the vide analytics. Yu can disable adaptive IR t help imprve classified bject detectin in the scene. Reflected Light 4
Lux n Target The recmmended minimum illuminatin is 8 lux n target fr analytic cameras. Fr nn-analytic, third-party cameras that are cnnected t the ACC ES Analytics Appliance, the minimum illuminatin requirement varies frm camera t camera. Mre light is required if the third-party camera des nt have an IR cut filter r a mnchrmatic night mde. Fr illuminating distances, it is imprtant t accunt fr lighting, weather, cntrast and camera stability cnditins. In bad weather with lw visibility, analytics shuld be cmbined with ther detectin methds t ensure a secure system. Cntact yur Avigiln representative fr advice n installing in challenging lighting situatins. Obstructins T identify bjects accurately, the scene must be clear. Fr utdr applicatins, avid placing a camera where the FV includes fliage, terrain r large bjects that cclude the subjects f interest. Als pay attentin t bstructins that can reflect IR illuminatin back t the camera and cause reduced cntrast r verexpse camera vide at night. This can be crrected by adjusting any f the fllwing: Separate the IR illuminatrs. Adjust the camera placement. Crrect the aim f the IR illuminatrs r the camera. Fr indr applicatins, a persn may be detected as lng as their upper bdy, including head and shulders, is visible. It is recmmended that a persn be fully visible fr the Avigiln Appearance Search feature t prvide better matching search results. Lux n Target 5
Fr mre infrmatin, see Analytics Lcatin Mde n page 3. Avid using analytics in crwded areas where peple are likely t verlap and blck each ther frm the FV. Overlapping bjects in the scene may cause the system t miss ptential results. Cverage Area Perfrmance is best in pen, uncrwded envirnments where peple are nt verlapping r bstructed in the FV. Install the camera in a lcatin where each bject appears in the FV fr at least 2 secnds. If an analytic rule r alarm uses a regin f interest (ROI) r beam crssing t trigger an event, make sure bjects are detected in the camera FV fr at least 2 secnds befre entering the ROI r crssing a beam. Fr advanced users, use the fllwing pixel n target recmmendatins: 24 t 36 pixels per meter (8 t 11 pixels per ft) based n 2.0MP reslutin. Maximum target size = 2/3 height f the FV. Fr the Avigiln Appearance Search feature: 72 pixels per meter (22 pixels per ft) based n 2.0 MP reslutin. Use the Avigiln System Design Tl t help yu estimate the required cverage area. The System Design Tl is designed t incrprate Avigiln analytic needs and determines the camera's maximum vide analytics detectin area in a given scene. T access the System Design Tl, g t https://sdt.avigiln.cm. Outdr Areas Be careful nt t select a cverage area that is t large, bjects may becme bscured by rain r fg even when there is enugh lighting and cntrast. Indr Areas Make sure the indr cverage area is nt t small. Lw ceilings r cnfined spaces (such as a man-trap area between secured drs) may pse prblems with establishing a scene that fits the recmmended criteria. FV shuld be at least 9 m (30 ft) wide, even if the regin f interest is much smaller. Cverage Area 6
Object Velcity Beynd the velcity recmmendatins referenced in the General Guidelines, be aware f the fllwing details: Objects that enter the FV frm behind the camera may take up t 4 secnds t be classified. If fast, lateral-mving vehicles are expected, use a wider field f view t increase the available bservatin time. Outdr Camera Placement Imagery 2016 Ggle. Map data 2016 Ggle. Make sure the camera FV verlaps t ensure adequate cverage in the blind spt immediately belw a camera. Make sure cameras can view bjects fr at least 2 secnds fr accurate bject classificatin r trigger cnfigured analytics rules. Fr mre infrmatin, see Cverage Area n the previus page. Munt cameras n a central building r structure lking ut twards the perimeter. Exceptins: Munt cameras n the perimeter if cvering exceptinally large areas. D nt munt n the central building if there is n suitable munting lcatin, r if there are bstructins in imprtant areas f the FV. Object Velcity 7
Imagery 2016 Ggle. Map data 2016 Ggle. Self-Learning Avigiln cameras include state-f-the-art bject detectin and classificatin. These cameras als emply additinal algrithms including Self-Learning and Teach by Example t reduce false detectin and alarm rates. Self-Learning is enabled by default in the ACC system. Self-learning imprves analytics decisin making pwer ver time by learning the physical relatinship f bjects mving acrss the FV. The camera r vide analytics appliance autmatically applies the mathematics f perspective t adjust itself and reduce the number f false alarms. Self-learning is activity-based t allw cameras r appliances t actively learn nly when there is mvement in the scene. Learning des nt prgress when the scene is empty r when it has lw illuminatin. Generally requires apprximately 200 high-cnfidence detectins thrughut the entire FV fr gd adjustment. The time needed t cmplete Self-Learning (100%) varies frm scene t scene, depending n activity. Self-Learning recmmendatins: Fr mst situatins, nly Self-Learning is required fr the camera t learn the scene. Ideal fr scenes where bjects are n the same plane. Self-Learning 8
Yu may want t disable Self-Learning fr the fllwing scenes: Scenes where bjects mve at different height levels. Scenes where bjects can be bserved at different levels that appear relatively clse in the FV. Fr example, verhead pedestrian bridges, train platfrms, hills and underpasses. T disable Self-Learning, see the Avigiln Cntrl Center Client User Guide. Teach by Example: Teach by Example is a feature that allws users t prvide feedback by validating the accuracy f classificatins dne by the system. Teach by Example is recmmended if the system reprts an undesirable number f false alarms after Self-Learning is cmplete, r has been disabled based n the preceding recmmendatins. Teach by Example is nt required, but can be used t help refine classificatin f peple and vehicles t further reduce the number f false alarms. Yu can perfrm Teach by Example after Self-Learning is cmplete, r when Self-Learning is disabled. If yu decide t disable Self-Learning after having executed a Teach by Example exercise, yu may need t teach the system again t accunt fr classified results that were previusly filtered by Self-Learning. It is strngly recmmended that yu reset the Self-Learning setting nce the camera is stable after initial cnfiguratin. During installatin, a camera is frequently adjusted, s any Self-Learning the system was able t d wuld becme invalid. NOTE: Always reset Self-Learning and Teach by Example after a camera is physically mved r adjusted, and if the fcus r zm level is changed. The change in the camera's FV affects the vide analytic results. In the case f nly lighting changes r IR installatin, it is nt necessary t reset Teach By Example. Hwever, adding mre examples f true and false classificatins with the new lighting will be beneficial. On the ther hand, Self-Learning shuld be reset when making lighting changes. Yu can reset Self-Learning and Teach by Example frm the ACC Client sftware. Fr mre infrmatin, see the Avigiln Cntrl Center Client User Guide. Avigiln Appearance Search Feature Avigiln Appearance Search vide analytics technlgy enables security persnnel t quickly and easily find all f the instances where a persn r vehicle f interest was captured in the recrded vide frm all cameras at a lcatin, t cmpile the vide evidence, and t answer the critical questins wh, what, where, and when with decisive actin. NOTE: ACC ES Analytics Appliance and ACC ES Cameras d nt currently supprt the Avigiln Appearance Search feature. Avigiln Appearance Search Feature 9
The system must be running ACC 6 Enterprise sftware. The netwrk vide recrder (NVR) must have the Analytics Kit installed. Ensure the camera supprts Avigiln Appearance Search. In the ACC Client sftware, pen the camera's Setup > Vide Analytics Cnfiguratin dialg bx and select the Enable Appearance Search check bx. Avigiln Appearance Search must be enabled fr each individual camera in the system. The camera must have the fllwing number f pixels n target fr ptimal Avigiln Appearance Search functinality. 72 pixels per meter (22 pixels per ft) based n 2.0 MP reslutin. Avid mirrrs and ther reflective surfaces (like shiny flrs and ceilings). Reflectins may cause additinal false detectins. It is recmmended that Avigiln Appearance Search nly be enabled n cameras using Outdr r Large Indr Area lcatin mde. The ther lcatin mde ptins may generate extraneus results. Disable the Avigiln Appearance Search ptin fr cameras using the ther lcatin mdes. Fr mre infrmatin abut the different lcatin mde ptins, see Analytics Lcatin Mde n page 3. Fr mre infrmatin abut the ACC 6 and Avigiln Appearance Search system requirements, see the ACC 6 sftware datasheet r the Avigiln Cntrl Center 6 Client User Guide. Camera Placement Fr Avigiln Appearance Search Fcus the FV f the camera t imprtant junctin pints: Entrances and exits Hallway junctin pints This is especially imprtant when using Avigiln Appearance Search feature because it helps investigatrs understand where peple travel ver time. Avid busy envirnments where images f peple ften verlap. It is difficult fr the camera t clearly distinguish different bjects in the scene if it is t busy. Fr busy envirnments, use several cameras t fcus n each junctin/exit s that yu can use the Avigiln Appearance Search feature t plt the general mvement f peple f interest. Use nnanalytics cameras fr situatinal awareness. Camera Placement Fr Avigiln Appearance Search 10
Fr Mre Infrmatin If after reading this dcument yu discver that yur site requirements deviate frm the recmmendatins, cnsult an Avigiln representative befre installing the cameras. We cannt help yu trublesht ptential issues with classified bject detectin if yu d nt fllw ur recmmendatins r seek assistance befre installing cameras. T cntact an Avigiln representative in yur area, see: http://avigiln.cm/cntact-us/ Fr mre infrmatin abut cnfiguring Self-Learning, Teach by Example and ther vide analytics features that are available in the ACC sftware, see Avigiln Cntrl Center Client User Guide. 2015-2017, Avigiln Crpratin. All rights reserved. AVIGILON, the AVIGILON lg, AVIGILON CONTROL CENTER, ACC, AVIGILON APPEARANCE SEARCH and TRUSTED SECURITY SOLUTIONS are trademarks f Avigiln Crpratin. Other names r lgs mentined herein may be the trademarks f their respective wners. The absence f the symbls and in prximity t each trademark in this dcument is nt a disclaimer f wnership f the related trademark. Avigiln Crpratin prtects its innvatins with patents issued in the United States f America and ther jurisdictins wrldwide (see avigiln.cm/patents). Unless stated explicitly and in writing, n license is granted with respect t any cpyright, industrial design, trademark, patent r ther intellectual prperty rights f Avigiln Crpratin r its licensrs. Avigiln Crpratin avigiln.cm PDF-DESIGNH3A-C Revisin: 2 - EN 20170329 Fr Mre Infrmatin 11