External And Internal Factors
Video Analytics can be an influential part of any video management system, transforming IP video into more than a passive monitoring tool. Modern video analytic algorithms have changed the way organisations monitor and act on their security surveillance footage.
That being said, there are many factors, both external and internal, that can influence the performance of computer vision-based video analytics algorithms. Performance can vary greatly depending on the situation of the deployment setting. These factors can be broken into the following general categories:
Each of these factors play a significant role in analytic accuracy, the frequency of false alarms, and streamlining VMS operations.
This white paper will explore each of these influential factors in detail, providing valuable insight for setting appropriate video analytics performance expectations, and delivering accurate, consistent results.
Environmental factors have a tremendous influence over the performance of virtually all video analytics algorithms. There are six general environmental factors.
1. Camera Angle
The angle of the camera can influence several factors used in video analytics, including perspective, occlusion, and segmentation of objects.
2. Distance to object
Pixel size of the object is an important element to video analytics. Most video analytics require a minimum pixel size (e.g. 15 x 15). Conversely, if the pixel sizes of the objects are too large, that too can distort the performance of the analytics (e.g. reflecting light into the camera).
3. Lighting Level
Lighting can influence video analytics in a few ways. First, in order for video analytics to detect objects, there needs to be some minimum light available (unless infrared or thermal cameras are being used). Second, abrupt changes in lighting (e.g. opening of doors) can cause false conclusions.
4. Degree of Activity
The degree of activity or “busyness” of an environment has influence over the performance of video analytics. Generally, the higher the level of activity, the more false conclusions will be drawn by the video analytics algorithm.
The volatility and variance of weather (sun, rain, snow, wind, trees, clouds, shadows, etc.) can cause false conclusion for video analytics, especially in outdoor environments. Weather also has an impact on video analytics in indoor environments where there exist large glass windows and doors, and the mentioned conditions create changes to the scene viewed by the indoor camera.
The degree of change to the background of a camera view can impact the performance of video analytics. For instance, if the view of the camera includes a constantly moving escalator, this could result in false conclusions, and would be need to be considered when developing or installing a solution.
Video analytics algorithms can vary greatly on the amount of computational power needed to perform adequately. There are five general factors that influence performance.
There are many different analytics engines. One can require 10x as much CPU as another. More CPU is normally required if you want to detect small objects moving quickly. This is because the engine would need to run at a high resolution (detect small objects) and also at a high frame rate (track fast objects).
Normally you can record video at 4CIF, and do analysis at CIF which saves CPU. If you want to detect very small objects, you may have to run at 4CIF.
Most analytics engines need between 5 and 8 frames per second. The record rate could be higher. Faster moving objects require higher FPS for tracking. Even left item detection analytics often use motion tracking to cut down on false positives.
If you want to be able to search through analysed footage (e.g. objects moving near a car), you will need to store the XML metadata produced by the analytics engine. This is normally a negligible amount of HD - a few percent of what the video storage requirements are.
An analytics engine typically increases memory usage on a PC, with higher resolutions requiring more memory.
DEFINATION OF ACCURACY
Determining the “accuracy” of video analytics can sometimes be difficult, not only because of the uncontrollable factors discussed above, but also as it relates to the actual definition of “accuracy”. This means different things to different people. Thus, setting proper customer expectations is important.
Many times, a customer will have a requirement for the degree of accuracy of the video analytics in order to assess the usefulness of the technology – e.g. 95% accuracy of Wrong Direction. While performance expectations are reasonable requests, the answer can be greatly influenced by how the data set is defined.
Let’s assume that in the month of January the number of people who exit an escalator properly (in the correct direction) is 100,000 and there was only one person per month that tried to go the wrong way. Let’s also assume that the video analytics properly detected the person moving in the wrong direction, but also incorrectly determined the direction of 300 people during the month and delivered a false alarm. Depending on the data set, the “level of accuracy” of the video analytics will vary greatly.
INTERPRETATION #1: 100% ACCURACY
In this interpretation, one could argue that only once during the month a person tried to go the ‘wrong direction’ and the video analytics correctly identified that event.
INTERPRETATION #2: 0.3% ACCURACY
In this interpretation, one could argue that video analytics delivered 301 events, of which only 1 was correct.
INTERPRETATION #3: 99.7% ACCURACY
In this interpretation, one could argue that video analytics evaluated 100,000 events (total people exiting the escalator in the month), and correctly identified 99,700 events.
The picture of accuracy is complex. The boundaries of definition are multiple – e.g. over time, number of events, etc. Thus, to properly set video analytics performance expectations, it is important to have a common understanding of what measurement criteria is being used.
For any questions regarding video analytics, contact email@example.com