Preventive Threat Detection

Yes, the models are trained on a wide variety of sceneries and backgrounds, with varied illumination and views. Essentially the solution is agnostic towards the background – as long as the object of interest is visible, the system will detect it.

The answer to this question depends on a number of factors. To begin with, it relies on the camera characteristics and the resolution, in particular. Resolution initially plays a big role, however, the incorporated zooming-tracking algorithm allows us to check the object in the original resolution the camera uses. Thus, unlike similar AI security solutions,  we are not heavily dependent on the quality of the visuals that is usually downgraded when processed through neural network platforms.

Then there is a group of characteristics that can be related to camera “picture quality”, such as stream bandwidth, encoding, etc. There are conditions of visibility to consider as well, such as the illumination, position (see Question 6 on object angle) and the pixel size of the object. The latter linearly depends on the distance from the camera and can be used for distance limit estimates. For example, the reliable minimum size limit for a gun object is around ~15-17 pixels which results in the maximum distance of up to 10-12 meters for most HD cameras.

The detection typically happens in the first 400 ms (in some cases, up to 2 seconds). When assessing the response time, take into account that most IP cameras used nowadays show some sub-second lag of the video stream. Also in the cases when the system is deployed on the cloud, the time lag that takes place when the stream reaches the cloud and the response reaches back the dashboard has to be considered.

Our Object Detection System is designed to help security units by supporting their daily operations, augmenting their capabilities and eliminating possible human-factor related flaws. Also, in case of a possible threat the alert that is sent out by the system is enriched with information crucial for quick and inclusive analysis of the threat on site and effective planning of dedicated counteractions.

Typically less than a second. In cloud deployments the response time can slightly increase depending on the client’s upload speed.

The system is based on computer vision algorithms and detection of the threat is based on visual answer analysis. This means that to detect a weapon in a bag our system has to be attached to X-Ray or millimeter-wave scanning devices. Running on CCTV cameras that operate in visual range only, the Object Detection can detect only non-concealed weapons.

No, the system is trained on all possible angles of objects of interest. Of course, in some specific cases the angle of the object would matter as the features that the system uses to classify the object are more distinct in some angles than others. For instance, if a gun/rifle is held at an angle towards the camera, more distinct features are visible compared to the cases when they are pointed directly at the camera.

An alert containing all the crucial information is compiled and delivered to end users responsible for security. There are number of customizable alerting pathways: dashboard, mobile alerting application, access point relay boards and VMS alerting API.

The limit here is only the hardware that runs the system analytics. More specifically, the main deciding characteristic is the GPUs the servers are equipped with. Other than that, Object Detection system can simultaneously accept video streams from different cameras with different characteristics.

Yes, all solutions provided by ACT can be deployed both on the cloud and on premise. Moreover, our solutions are cloud provider agnostic, as long as the cloud instance runs Ubuntu 18.04 and is equipped with a Nvidia GPU card.

Absolutely. ACT does not store any information (unless requested specifically by the user).

An alert is classified as a true alarm when the prediction of AI corresponds to reality (i.e. the object of interest is correctly identified, the action sought after is detected, etc.). A false positive is a case when the alert is triggered by mistake. Unfortunately, due to the essentially probabilistic nature of AI, the latter are inevitable in most cases.

However, due to the elaborate AI and machine learning behind Object Detection System, it can meet any level of production-grade industrial standards. Moreover, we are continuously improving ACT modules where they are retrained on mistakes to make sure the number of false alarms goes even further down with time.

The Object Detection System is camera agnostic. Most questions on the limitations and camera requirements end up receiving a simplified “rule-of-thumb” answer – if a human can see and identify the object of interest, then ODS AI will also be able to do that (and in some cases, will even outperform a human due to the integrated zooming and re-checking algorithms). As for the minimal camera parameters, these will depend on each use case and the object of interest you are trying to detect. Of course, the camera should have a digital output or at least be connected to a DVR that has one.

ACT Intrusion Detection and Perimeter Protection System can accept pretty much all the variety of stream types, such as RTSP/RTMP, HTTP, etc. Usually the minimum required resolution starts from HD (1280×720) and 5 FPS. Parameters defining the frame/image quality vary from one camera to another, but we recommend looking into such characteristics as bandwidth, encoding, and sharpness, and improving them if necessary.

The duration of alerts depends on the client’s data retention policy. By default, we offer a storage duration of one month, but this period can be configured to correspond to local policies.

Yes, it can. Our computer vision engine is indifferent towards moving and changing backgrounds and works similar to human vision – if the object is there and it is distinguishable, the system will spots it and reports. Of course, blurring of objects during PTZ shifts should be taken into account.

The Object Detection System is designed to work in challenging environments where cameras with embedded algorithms misperform. The AI engine compensates for the drawbacks imposed by demanding conditions including but not limited to poor illumination, somewhat corrupted frames, environmental factors and effects of weather.

The Object Detection System is not designed to detect unhandled or holsted weapons. It triggers the alert as soon as the weapon is in possession of a human.

Yes, it can. The maximum detection distance of Object Detection System will depend on the camera characteristics (contrast ratio, pixel crosstalk, etc). But in general, the solution complies with industry standard DRI requirements, i.e. the recognition limit (the distance at which you can determine the class of an object) is ~15 pixels for small weapons.