The inclusion of integral IVA in edge devices is seen by many as a positive step, and in a number of applications it is. However, the benefits of IVA can only be realised if the performance is consistent, accurate and reliable. Not all edge-based IVA is the same, and much like VMD a decade ago, there are some implementations that don’t quite match up to the hype. Are we simply expecting too much from the technology or is it poorly applied?
It is very difficult to define conditions under which video surveillance is more likely to be required. As evidence proves, the benefits of video surveillance can be realised around-the-clock, in a variety of applications, for a wide range of needs and requirements. However, statistics do show that a significant percentage of crimes take during the hours of darkness, for good reason.
Firstly, many businesses and organisations are closed for business at night. Sites are unoccupied, and in areas where there are few residents or businesses operating around the clock, the numbers of passers-by are minimised.
When it comes to crimes against property or premises, these are more prevalent in isolated locations such as commercial areas or industrial estates, during the night or periods of darkness. Some do occur in city and town centres or populated areas, but the greater threats in such areas at night come from crimes against the person.
Adding to this is the fact that criminals prefer to go unseen. The risks of being spotted and identified increase significantly during periods of daylight.
Capturing high quality images during hours of darkness is not a new challenge for the video surveillance industry. Over the years many approaches have been tried, and new technologies have been introduced. The challenge has also be made harder, as demand for higher resolutions introduces smaller and smaller pixels onto which the right amount of light needs to fall.
The biggest problem with low light applications is that when insufficient light falls onto the sensor, the camera’s processing engine will amplify the signal in order to create an image. As well as amplifying the important details, it also amplifies interference and unsuitable parts of the signal, and this results in image noise.
While processing engines can do certain things to reduce the amount of noise in an image, they don’t eliminate it, and noisy images create their own problems. Firstly, compression engines see noise as change, and that change occurs across the entire scene. As a result, the captured image exhibits a lot of change from the reference I-Frame, and so every subsequent P-Frame is completely rewritten. This erodes the efficiency of compression.
Not only does the video footage require significantly more bit-rate for transmission, but it also takes up more storage space. As a result, the processing engine in the camera works harder, which in turn can impact on other processing functions.
The IVA burden
While it could be argued that processing power in many edge devices has increased in recent times, it must also be considered that due to these increases, manufacturers are adding ever more features and functions. One which has become almost a staple feature of modern cameras is the inclusion of video analytics.
Edge-based video analytics vary from manufacturer. Some provide a number of flexible and credible options, while others include one or two rules with minimal levels of configuration.
Some integrators take the view that integral IVA functionality will be hit and miss because it’s effectively ‘free’, and as a result it won’t be as good as a licensed option. However, the reality is that the functionality is paid for when the product is specified, and if it is listed as a provided features sales benefit – which it typically is – then it should work, and should stand up to comparison with competitive alternatives.
One issue to be considered when looking at using integral IVA in low light applications is that often the basic licence-free options have minimal configurable discriminations. Most allow the creation of minimum and maximum sizes, and whilst these will ignore global scene changes if applied correctly, some will see noise as individual elements of motion. As a result, the only approach is to assess individual cameras in the location you intend to deploy them.
This is especially the case where integral illuminators are included with the camera. While the addition of infrared illuminators in cameras does provide some benefits, this approach does have limitations because of the way the LEDs are placed. Rather than being optimised to provide even and consistent lighting, the LEDs often are forced to fit into whatever space is left once the camera module is put into the housing.
Typical approaches include locating the LEDs around the lens, or having a small cluster of illuminators close to the lens. These approaches are fine for general surveillance, but when you are seeking a consistently and evenly illuminated scene for video analytics, they can create issues.
As is typical with such devices, there is some degree of light ‘pooling’ in the centre of the image, with the edges being darker and more susceptible to noise. While traditional thinking is that a camera has the region of interest in the centre of the image, the reality is that with higher resolution devices, a wider area is captured and all of the viewed scene should be consistent in terms of image quality.
Additionally, where integral LEDs are used, these typically are 850 nanometre arrays and as such emit a red glow. If this attracts insects or spiders, the IVA could be affected.
The various cameras were also tested using discrete illuminators, with both infrared and white light options deployed. While the illuminators were specified as closely to the integral LEDs as possible, the evidence was that the achievable ranges from the dedicated units was higher, and within the quoted range there was very little fading or uneven illumination. Even visually, the improved clarity in the images was obvious, and this was born out by the IVA performance.
With white light, there was no issues with shadows. However, when the illuminators were positioned to deliberately create problematic shadows, all three test samples created nuisance activations!
Axis: AXIS Motion Guard
AXIS Motion Guard is an advanced video analytics application which detects motion in pre-defined areas. AXIS Motion Guard is a downloadable application, available via the Axis Communications web site, which can be used with any Axis camera that supports ACAP and is loaded with firmware version 7 or later. Axis also offers a wide range of other IVA applications including AXIS Fence Guard, which creates virtual fences in a camera’s field of view, and triggers an alarm when it detects a moving object crossing the virtual line.
AXIS Motion Guard makes use of an intuitive user interface with real-time visual confirmation to deliver validation of correct configuration. The application integrates with the camera’s Action Rules management feature to automate actions when an alarm event occurs.
AXIS Motion Guard can support different profiles, which can have unique configurations. This allows specific rules and actions to be applied dependent upon a schedule, or to be manually switched. AXIS Motion Guard uses filters to eliminate the vast majority of nuisance activations. The application delivers triggers which can be used by the Action Rules function, and that gives a wide range of options with regard to managing events.
Unlike the other units, the hardware tested does not include integral illumination, so could only be used with additional infrared and white light illuminators.
Performance was good, with high levels of accuracy and consistency, even in mixed weather conditions. During dawn and dusk periods there was a very slight increase in nuisance activations, but with some tweaking these could be eliminated.
Hanwha Techwin: Wisenet X
Hanwha Techwin’s Wisenet X camera range offers a number of video and audio analytics rules. The range also supports Hanwha’s open-platform approach, allowing the use of third-party applications to enhance functionality.
Integral analytics rules include virtual line crossing, zone entry and exit, directional detection, object appear and disappear, loitering detection, motion detection, tamper protection and face detection.
With regard to line crossing, up to eight lines can be supported but may only make use of two nodes and cannot be segmented. Each line can have different directional discriminations applied. Up to eight detection zones can also be created. In order to minimise nuisance activations, eight exclusion zones can be set.
To ensure flexibility, each line or detection zone can have different filtering configurations applied, but the associated alarm actions for all lines in a specific scene must be the same. This makes sense if multiple analytics rules are being used to create a defined shape or to follow a perimeter, but it does rule out any options for using multiple lines to create staged alarm events.
The camera makes use of integral infrared LEDs, and as with other similar devices in the test, the issues occurred with IVA operations at the edges of the image area, despite the intrusions happening within (albeit at the extreme) of the quoted range. While coverage is generally good, some environmental conditions did see a few subtle intrusions missed. Over a prolonged period, nuisance activations were around 7-8 per cent.
When the camera was used with additional illumination, with infrared and white light options, IVA performance was much improved. Given the relative simplicity of the IVA, it works as expected with additional illumination, and in many systems will be enough for most needs.
Hikvision: Smart Feature Set
Hikvision’s Easy IP 3.0 series cameras include what the manufacturer refers to as a Smart Feature Set. This includes four intelligent video analytics algorithms: line crossing, intrusion into and/or loitering in a virtual zone, object left and object removed.
It also supports face detection as an alarm trigger source, as well as tamper detection options.
The virtual zone feature allows the creation of multiple detection areas. With line crossing, dependent upon which specification you read, either up to four virtual lines or a single virtual line are supported, and each can support individual configurations. The use of four lines allows more complex scenarios to be created.
Directional discriminations can be applied, as can size criteria to help filter out nuisance alarms. Sensitivity of detection is adjustable.
The Smart Feature Set option is quite basic. This does make set-up a quick and easy affair, but equally limits the use of the IVA to simpler applications. However, once deployed, it is stable enough for most mainstream applications and works as expected.
The camera tested included integral infrared illumination. Although all tested activity was within the specified range for the camera, so activity at the extremes of the quoted distance were missed. Equally, we experienced a false alarm rate of around 10 per cent due to anomalies due to noise and camera processing.
Interestingly, when the camera was used with a dedicated infrared illuminator or a white light illuminator, performance was enhanced and the nuisance activations were significantly reduced.
The camera also features basic VMD and this was also affected by low light issues.
Verdict
The test, which was looking at IVA performance as a whole rather than individual camera performance, gave fairly similar results for all devices. There was little point in trying to stabilise the IVA performance with the cameras close to their specified minimum sensitivity levels with no additional lighting. The reality of sensitivity specifications has been addressed by Benchmark before, and while advances have been made, it is not sensible to try and achieve the quoted figures for general surveillance, let alone for stable IVA deployments.
The two sample units which incorporate integral infrared illuminators performed to a decent degree using the integral LEDs, but there were issues. This had more to do with the quality of the illumination than the IVA algorithms. However, all three devices performed as expected, and showed good levels of stability and accuracy, when used with additional dedicated illuminators.
Of the two types, the better performance was achieved using white light illumination. Obviously, this allows the cameras to capture clear and detailed colour images, and noise isn’t as great an issue. While the performance with the infrared illuminators was very good, where white light can be used, it should be the first choice. Obviously, in some sites light pollution is a concern. The only point to note is that care must be taken to avoid the creation of innocuous shadows.
That said, monochrome images using even and consistent infrared illumination are high in contrast, and that is no bad thing when using video analytics. With professional dedicated infrared illuminators, all three cameras performed well and the IVA didn’t cause any concerns or display any unusable behaviour.
Of course, if you want to further increase the accuracy of intelligent video analytics in low light applications, there is one other option to consider: thermal imaging!