CCTV Test: Object-based Video Analytics (Part 1)
One of the first video ‘analytics rules’ to emerge in the security sector, way back in the analogue days before terms such as VCA or IVA were invented, was the detection of object removed violations. It seemed very advanced at a time when video motion detection was considered to be the pinnacle of security technology. Today, ‘object removed’ and ‘object left’ analytic rules are typically only mentioned in the lists of additional rules for devices or software packages. Benchmark took a look at some of the object-based options to find out if these basic IVA functions offer anything of use to installers and integrators.
When it comes to video analytics, the headline features tend to be those that allow the creation of the more ‘exciting’ applications. Demonstrations and screen grabs used to promote the technology tend to involve intruders scaling fences at critical infrastructure sites, vehicles being pursued through high risk environments and individuals being tracked as they travel across campus-type applications. The general impression being generated is one of smart and innovative solutions tackling crime at its most insidious.
Whilst it is very true that video analytics does deliver in all of these scenarios, the technology can also offer simpler solutions to more basic problems. Talk to those tasked with delivering IVA, and they will tell you – if they’re being honest – that for every high-end deployment of the technology, there are dozens of others that carry out more menial tasks.
While the technology can still shine in such applications, they often don’t get highlighted because they lack the ‘sexy’ nature of the more challenging roles that IVA fills.
Of course, for most installers or integrators, it is this ‘bread-and-butter’ business that forms the backbone of customer demands, and as a result the core functionalities of any IVA set-up are likely to be more important than the headline-grabbing features.
When you look down the list of rules incorporated in many IVA packages, ‘object left’ and ‘object removed’ are often present. These are very common scenarios, and are important to a wide range of customers. The reason they sometimes don’t attract the headlines is because, on the face of things, they appear to be very simple.
If the implementation of these rules is done well, they are not only simple, but also offer a solution for a wide range of applications. In this test, Benchmark sets out to see if this functionality is reliable and robust in a very basic scenario.
One thing that can be confusing is what manufacturers mean by ‘object left’ and ‘object removed’. The term ‘object’ is used by some to include people, vehicles, targets, etc..
Some use the terminology to describe an alert caused by the arrival of an ‘object’; for example, if a car drives into a region of interest. However, with ‘object left’ analysis, expected objects behaving in ordinary ways should not trigger any type of alarm. Equally, with ‘object removed’, temporary obscurations or other environmental changes should not be confused with relevant events. Often a field trial is the only way to understand what type of object analysis is really being offered.
For the core test, Benchmark used an office which included a storage area. The goal was to detect left objects or removal of stored items. The object left and object removed tests were carried out separately; where simultaneous detection was possible, this is also reported.
ACIC: ActivityDetection Edge
ActivityDetection Edge from ACIC is an embedded video analytics software application for the Axis ACAP platform cameras and encoders. The App is available in versions for ARTPEC 3, 4 and 5 versions of the devices.
The application can support up to eight concurrent rules, and these include presence detection, line crossing, stationary object, object removed, object monitoring and motion detection.
With regard to object-based rules, the software can detect object left, object removed, or both violations simultaneously.
The software is claimed as being able to adapt to environmental conditions, and boasts what the manufacturer claims is a configuration interface that delivers ease of installation. Integration with many of the leading VMS solutions is possible.
The App is simple to install onto any compatible Axis camera (ACIC provide a compatibility list with details of required firmware). It is uploaded via the Applications menu of the edge device, and the licence is applied via a simple file.
ActivityDetection Edge can be used with a number of browsers (IE, FireFox and Chrome are supported, according to the manufacturer); the configuration menu requires that a Java applet is run. Some IT departments are suspicious of Java following past vulnerabilities, but the current version has thus far not raised concerns in the data security world.
Once the elements are loaded onto the device, there are two menu options: Main Page and ActivityDetection Configuration.
We opted to use IE for set-up, but found the ACIC interface to be a bit flaky at times. For example, the Configuration menu appeared as it should, but text and numerical values could not be entered. A few screen refreshes later, this started to work. However, the main screen – where results are checked – would not load and the various elements displayed incorrectly. The main page did appear correctly with live video loaded in Chrome, but there were then other issues with the Configuration menus. The end result was that the easiest approach was to use two browsers, which wasn’t ideal.
IE has become established as the browser of choice in security (rightly or wrongly), so our thinking is that App developers must ensure their products are fully IE compatible. Lesser manufacturers seem to be able to deploy the browser with ease, so we expect a security company selling browser-dependent software to be able to do so too!
Initial loading of the config screen does prompt you to run a wizard. This initiates a calibration procedure which is required to establish scene depth and object size. This is done by drawing ‘segments’ and identifying their height (or length) to allow for perspective. Once this is done, you can start creating rules.
Rules can be generated either by using the wizard or a manual set-up process. The wizard gives some indication of the way the rules are structured, so it’s worth running through it a few times to get your head around the ACIC processes.
The first task is to establish a region of interest. These can be named, but the software will link regions, detection types and object classes by number. Next you need to create an object class. Choices include automatic settings for people or vehicles (these can be tweaked if necessary), plus a custom option which allows you to define minimum and maximum sizes, plus speed.
The final step is to configure a detection rule. This links the region of interest with the object class, and defines the method of detection. Object-based rules are for object left, object removed or any violations. A time period can be specified.
Once the rules are established, these can then be checked, and tweaked, using the main page. When the configuration is complete, the App then uses the Events functionality of the Axis device, via the Action Rules menu, to set up appropriate alarm actions.
Initially the App was run with single rules specified. Object left was the first tested, and the performance was good, with all left items being detected in a timely fashion.
Switching to object removed did see some item removals missed. We noticed that these seemed to occur if there was a lot of activity in the scene prior to the removal. As most risks are associated with objects being removed by individuals – who might loiter until an opportunity arises – we did feel that this was something of a weakness in the App.
Interestingly, with both rules selected (the ‘Any’ setting), the object removed detections did seem to be marginally more reliable. However, in general, the object left detections were around 98 per cent accurate, whilst object removed seemed to be around 90 per cent accurate.
Bosch: Dinion IP 7000
The Dinion IP 7000 is an HD1080p static box-type camera from Bosch. Its headline features are common across the well-established Dinion range and include CBIT (content based imaging technology) which uses video content to adjust the processing levels, intelligent dynamic noise reduction, area-based encoding, regions of interest and edge recording. Pertinent to this test, it also includes intelligent video analytics as standard.
With regard to the IVA, the rules include line crossing, loitering, motion, object in detection field, condition change, route following, entering or exiting a detection field, crowd detection, counting, object left, object removed and camera tamper. There are also general settings for global change and image deterioration.
The Bosch IVA functionality also allows metadata to be gathered relating to scene motion, without triggering the IVA. Given that this feature is license-free, it is more powerful than many similar options.
Our test camera was received without any manuals or utility CD. This is a frustration, as the default settings as given in the documentation we found on the Bosch website were not correct for this model. We eventually discovered an on-line download for the Bosch utility (called IPHelper) and this immediately identified the camera and allowed the network settings to be changed.
Once logged in, you’ll notice that the Bosch GUI has had something of an overhaul, and it does look refreshed and more modern. The ActiveX viewing element is still used, but is no longer loaded directly from the camera. Instead it needs to be downloaded from the Bosch website, which does necessitate on-line internet access from the server.
The process isn’t the fastest. Without the viewer, you do have low frame rate M-JPEG images, but set-up of IVA and other functions won’t work properly without the ActiveX add-on. Ever since Microsoft changed IE to prevent easy ActiveX implementation, it has been obvious that many cameras would need changes to the config pages. However, this approach seems clumsy.
Once up and running the VCA configuration can begin. As we had no manual, we relied on the Help function (F1 brings this up). In truth, there’s not a lot there that’ll confuse any installer or integrator worth their salt.
The configuration is split into three menus: Tasks, Metadata Generation and Metadata Inspection. The Tasks menu allows the setting of up to eight rules, and the various parameters depend upon the type of rule.
With regard to object-based rules, the supported analytics detect Idle Object (item left) and Object Removed. Bosch uses the term ‘object’ for other targets too, but these are two options we concentrated on.
The first setting defines the detection area (this can be the whole scene or a generated zone) and the ‘intersection’ of the object required to generate an alarm. The type of object is then identified. A wide range of criteria can be set including size, shape, colour, etc., which can be used to filter out unwanted activations.
The feature also allows a differentiation between started/stopped items and placed/taken items. These can also have different debounce times.
Once completed, the rule and its definitions are completed. The next menu, Metadata generation, allows performance configurations to be carried out, defining 3D elements such as perspective, to allow for enhanced accuracy.
Because of the depth of configurations, coupled with the somewhat basic Help function providing the only documentation, correctly setting up the Dinion IP 7000 takes a little longer than some of the other options. However, you are rewarded with a good degree of accuracy.
Once set up, the accuracy was consistent, and all attempts at removing or leaving objects in the protected area were detected. Both rules can be used simultaneously without any performance issues or drop in accuracy. The rules can also be supplemented with additional IVA elements if required.
Significant scene changes such as lighting switches did not generate false alarms, and even with racking the introduction of objects was not confused with removals, and vice versa!
The VB-M740E from Canon is an external bullet-type HD720p camera designed for ease of installation. The ruggedised unit makes use of H.264 and M-JPEG compression, its headline features include low light performance, enhanced bandwidth and bit-rate management, edge recording and external protection to IP66, PoE and privacy masking.
The camera is fitted with a 2.55-6.12mm varifocal lens (F1.2–F1.8), which delivers 2.4x optical zoom and is motorised.
The camera also includes integral IVA functionality and zone protection. With regard to the IVA, the supported rules include moving object detection, abandoned object detection, removed object detection, camera tampering detection, passing detection and intrusion detection.
The VB-M740E is supplied with a set-up CD, a paper quick start guide (nine different language versions are included), a mounting template, safety wire and an allen key.
In the past we have asked why the people at Canon don’t get an IP utility like the one that Axis – Canon’s sister company – offers. Sadly, it’s not yet the case, although you do get Canon’s installation software which is a little more time-consuming.
There is a single installation package, and this installs the Camera Management Tool, Camera Angle Setting Tool and an Admin Tools Certificate.
The Camera Management Tool discovers connected devices on the LAN, and allows the IP address to be configured. Unless you intend to use Canon’s management software, the next stage for many will be to complete the camera configurations and switch to control via a VMS or NVR. However, you may need to return to the tool to access the Admin Tools; more about this in a moment.
When you log into the camera, the settings menu appears immediately. However, the viewer page will hang for a while before showing an error message if you have not added the camera to Trusted Sites in IE. Once it is added, the video stream is shown.
One frustration when carrying out general configurations is that not all menus include a preview of the video. It’s something we’ve commented on before, and it does mean that during initial set-up you have to keep changing screens to see what effect your tweaks are having.
A quick scan of the menus will send you back to the PDF manual, as you’ll not find any configuration options for the IVA. These need to be accessed via the Admin Tools menu, which can only be pulled up from the camera’s top page. As an aside, this menu also includes the settings for privacy masking and recorded video management.
There is no simple or obvious way to access the Admin Tools from the viewer or settings pages, which is frustrating. To gain access to the menu you’ll need to leave the browser and return to the Camera Management Tool and double click on the IP address of the relevant camera.
It’s hard to understand the logic behind making the set-up different to the vast majority of professional cameras out there. Once you’ve been through the process once you then know what to do, but we didn’t feel the approach was more beneficial or secure than the more straightforward approach taken by the majority of other manufacturers.
Once you’ve gained access to the Admin Tools menu, you can start to configure the IVA rules. Along with object left and object removed, the options include moving object, intrusion, passing and tampering.
To configure an IVA event you select the required rule. A detection zone can then be created; this can be either rectangular or polygonal. A target size can then set as a percentage. An indicator allows this to be set with ease. The final setting is to qualify the event that occurs on activation of a rule.
The process is relatively simple, which also underlines the fact that it is relatively basic when compared to some IVA options. Indeed, Canon makes a point of stating that the IVA functionality should not be relied upon where a very high degree of accuracy is required.
The Object Abandoned rule worked well, and where activations were occasionally missed, this was invariably with small objects in busy scenes. Target size is set by percentage, and an icon displays the equivalent size during set-up.
We did check the manual to see if it offered any advice on the IVA configuration, but the information was minimal. However, if you want to improve the accuracy of detection, the answer seems to be to ensure camera positioning and the viewed scene are sympathetic to – and emphasise – the area being protected. Expecting the camera to watch over a wide area and detect a small bag being concealed is never going to deliver outstanding results!
The Object Removed rule was a little more prone to missing events. Initially we used a number of packages on a rack. Accuracy was around 70 per cent with a general view (the packages were around 5 per cent of the screen height). By zooming the view to mainly cover the rack itself, accuracy increased to around 85 per cent. The best results were achieved with larger objects in a relatively sterile scene.
Object left and object removed analytic rules have a wide number of potential uses, and while they don’t attract the headlines like some other IVA rules, they still offer a degree of opportunity for many installers and integrators.
ActivityDetection Edge from ACIC is a decent add-on for Axis cameras, and does feature a number of IVA rules. In general these work well, and although in our test a few incidents were missed, these could probably be tweaked out over longer term use. Where the App falls short is in regard to its browser-based interface. We ended up using both IE and Chrome (the former for the live view and configurations, and the latter for the App main page) which isn’t really acceptable. If the interface was improved, it would certainly rank higher.
The Dinion IP 7000 is a well established device from Bosch, and the software changes made earlier this year now allow full access to the IVA functionality via the camera rather than the Video Client. This makes configuration faster, but you still need to spend time on it to get the performance right. If you do, the two object rules both deliver a high degree of accuracy, and these can be applied simultaneously. In our test scenario, the Dinion didn’t miss any genuine alerts.
Canon’s VB-M740E is firstly a video surveillance camera, and secondly a device designed for use with IVA. The rules are simpler than the Bosch camera, and as a result the installation will need to be designed with achieving object-based detection in mind. If this is done, then the camera does a deliver a more-than-decent degree of accuracy. In a busier general scene, then some events will be missed if they include smaller objects, or if there is a lot of motion which obscures the target items.