
Figure 1: Hikvision AI Cloud three-layered architecture [1].
Axis Communications
https://www.axis.com/
Axis is a pioneer in surveillance cameras that are Internet-enabled. Axis is based in Lund, Sweden. Given its lengthy history of innovation, Axis is considered one of the industry leaders in employing AI and machine learning towards video surveillance. Axis is also a pioneer in including computing capabilities on cameras themselves. As discussed earlier, having computing resources on cameras themselves allows for real-time detection of objects of interest in video surveillance.
Bosch Security and Safety Systems
https://www.boschsecurity.com/
Bosch is located near Stuttgart, Germany and has a strong history in security and video surveillance cameras. Although it was only more recent that they invested heavily in Internet-enabled cameras, their longstanding experience with video surveillance has brought them to the top-of-the-line quickly.
Dahua Technology
https://www.dahuasecurity.com/
Dahua Technology is a Chinese company that has similar offerings as Hikvision. Dahua and Hikvision are the leading video surveillance companies based in China and a consideration of video surveillance systems should include both.
Honeywell
https://www.security.honeywell.com/
Honeywell is based in Charlotte, NC, USA and has a similar strong history to Bosch in video surveillance where they were later to invest in Internet-enabled cameras. However, their current product offerings are advanced.
A classification of this image may identify the image as a potential security threat to be investigated while object detection would specifically identify the abandoned bag.
It is important that machine learning models are evaluated for their performance. As mentioned above, an important evaluation measure is the robustness of different objects a model can detect. There are many different sophisticated types of evaluation metrics commonly used in machine

Figure 2: Detection of an abandoned object in a public area [5].
learning with a simple but powerful one being a confusion matrix. A confusion matrix compares the number of correct and incorrect predictions made by a model with a set of data where the answers are known. Figure 3 shows a typical confusion matrix where 50 images were correctly classified as being an object, 10 images were incorrectly predicted to be the object and are not, 5 images were incorrectly predicted to not be the object and is, and 100 images were correctly predicted to not be an object and are not.

Figure 3: Confusion matrix to evaluate a machine learning model.
n security, it is important to minimize both the number of errors where the model incorrectly predicted an object and it is not (false alarms that use up valuable human resources to investigate) and images that were incorrectly predicted to not be an object when it actually is (failing to identify a security threat). Performance measures of models against known answers is important to report in order to improve models and to convey to others the expected accuracy of models against new video footage.