The attached technical note describes h ow to benchmark and differentiate video analytics solutions. ... View more

The attached technical note describes w hich camera settings influence the performance of Essential or Intelligent Video Analytics in Bosch IP cameras, and what the best practices are. ... View more

This article  describes idle and removed object detection with Intelligent or Essential Video Analytics in FW 6.30, how to configure it and what the limitations are.   Applications: - Detect parked cars in no-parking zones - Detect blocked emergency exits - Detect the removal of museum exhibits - Detect suspiciously left packages in high-security areas - Detect idle bags in walk-through areas Limitations: - Not available on thermal cameras - Does not work in crowds - No abandoned baggage detection - No differentiation between baggage and persons sitting or lying down - Does not work in front of moving background like wind in trees or water - Does not work during illumination changes or camera brightness adaptations   Step-by-step guide   Abandoned baggage myths and facts: “Abandoned baggage”: A bag that has been carried by a person and has been placed on the ground (“idle”). Afterwards, the person left the bag and the surrounding area (“abandoned”). Application target areas are often crowded train station platforms or equally crowded airport waiting areas. THIS IS NOT POSSIBLE AT ALL: - The bags are often fully occluded by the crowd of people and thus not visible anymore. Something that is not visible cannot be detected. - The people are waiting, standing or sitting, and not moving any more than the bags do. Current idle object detection technology cannot differentiate whether a person or a bag is currently idle. Thus it will alarm on any waiting person as well, with a huge amount of false alerts. - Separating and tracking people in crowds is not possible with current technologies, thus the owner of an idle bag cannot be determined, and neither can be determined whether the owner left the premises and the idle bag is therefore really abandoned. - People wait at train station platforms and airports for hours at a time, much longer than any abandoned baggage should be alarmed on. - If people arrive in groups, like a family traveling, often someone is asked to look after baggage while the other leaves to buy something. Thus even while the correct owner left, the baggage is not abandoned at all, but well cared for. - The probability that a bomb is placed in a trash bin or hidden somewhere else is very high. No idle object detector will be able to find them then. - Even if an abandoned bag is detected, the time for a potential bomb to explode is too short to do anything about it. Two different solutions   for idle / removed object detection: - Static background and stable illumination: Use reference image check from tamper detection. See the tech note on tamper detection for full details and configuration. - Changing illumination or background: Use idle / removed object detection tasks.   Different kinds of idle / removed objects: Technology Idle and removed object detection is always based on the comparison of a current image to a background representation, which can be a single static background like with the reference image check from the tamper detection, or a learned background image as with Intelligent or Essential Video Analytics. Thus it also has all of the limitations from background subtraction algorithms like not being able to separate close objects as they only appear as a single foreground blob and being confused by illumination changes and motion in the background as those are changes to the background, too. For best performance, use Intelligent Video Analytics on CPP6 or higher, as this has the most advanced object reasoning using other features like optical flow, trajectory analysis and object shape analysis, as well as the best background maintenance and idle object detection based not only on one background image but a statistical representation over time. The difference between an idle and a removed object for the video analytics is basically whether edges were taken away or added into the image. Started and stopped objects are tracked in addition, allowing further evaluations. For learned background images, idle and removed object detection is essential for background image maintenance. Note   that overall illumination changes or camera brightness adaptations fully disrupt the video analytics, as the difference between background image and current image becomes too large. While Intelligent Video Analytics on CPP6 or higher with FW ≥ 6.10 still continues to track the moving objects, all idle / removed object observations are still disrupted and stopped. All other versions loose the moving objects as well. Therefore while in theory quite long idle / removed time to alarm can be supported, in practice the idle and removed time to alarm should be as short as possible to avoid illumination change disruptions.   Configuration For the   configuration of idle / removed object detection   using reference image check please see the tech note on tamper detection. All else is described here. The   basic metadata configuration   options of idle / removed object detection are the time to alarm. The same time is also used to determine when to put idle objects into the background image and remove removed objects from the background image. Note that started objects have no configurable time to alarm, as the alarm is send as soon as the algorithm verified it is a started object. Other configuration options   allow to fully enable or disable the metadata generation of idle / removed objects. For Intelligent Video Analytics on CPP6 and higher from FW 6.10 onwards, a separate idle object size for placed and taken objects can be configured. For all other cases, the minimal and maximal object sizes from Metadata Generation -> Tracking applies. Note   that Metadata Generation -> VCA masks also suppress any idle or removed objects. Note   that in front of moving background, the started, placed and taken object detection will not work. Intelligent Video Analytics on CPP6 and higher with FW ≥ 6.10 fully suppresses all started, placed and taken objects in front of detected moving background areas when Metadata Generation -> Tracking -> Noise suppression is enabled, including in Metadata Generation -> Tracking -> Ship tracking mode. In all other cases, expect a multitude of false idle and removed alarms in front of moving background. After configuring the metadata generation, an idle or removed object task has to be configured in the alarm rules. There, the alarm generation can also be restricted to a specified alarm area. As the time to alarm needs to be configured already for the metadata configuration, it is only added for information in the alarm rule wizards. ... View more

This article describes the camera tamper detection options available with FW 6.30: Scene too dark / bright, global change, reference image check. Applications: - Detect camera blinding - Detect camera being moved away from of its view - Detect the loss of sufficient overall illumination - Detect idle / removed objects, e.g. objects in front of an emergency exit Limitations: - Not available on thermal cameras   Step-by-step guide   1. Scene too dark / bright Application: Detect the loss of sufficient overall illumination. How it works: Scene too dark / bright evaluates the amount of incoming light at the sensor and generates a corresponding brightness level with values between 0 and 255. This is then compared to a user-configured threshold. The check is run every 10 seconds and needs to be true twice to alarm, thus the alarm is generated only if and after the event occurs for at least 20-29 seconds. The current brightness level is shown to aid in setting the threshold. Limitations: - The computed brightness value depends on the chosen camera / lens combination. Therefore the threshold may need to be adapted for every camera combination as well. - Not all cases of camera blinding can be detected with scene too dark / too bright. Even with covering the camera, the brightness at the sensor may still be quite high, and the brightness of artificial lights like torchlights is less than a very bright sunny day. Use reference image check to detect camera blinding instead. - Available since FW 6.30. (Prior to FW 6.30, the average brightness of the final image pixel was evaluated, however, as can be seen in the example below, current camera image quality is good enough to generate a properly bright image even from very dark or very blinded scenes, so the old alarming method was not reliable anymore) Some camera blinding situations and their corresponding brightness values. When the material covering the camera is not fully opaque or fully covering, brightness levels can be quite high. Torchlights and laser pointer were held very close to the camera to archive highest brightness possible, otherwise brightness values will be even lower.   2. Global change Application: Significant but short-term changes in the image How it works: 1. MOTION+ is running: MOTION+ signals the percentage of changed blocks in comparison to a 1 second old image. If this number is larger than the sensitivity threshold, then an alarm is immediately generated. Only the sensitive area from MOTION+ is evaluated. 2. Essential Video Analytics (CPP7) or Intelligent Video Analytics (CPP4) is running: The global change is determined by first evaluating whether too much foreground is detected, or secondly whether a comparison to the 1 sec older image is smaller than a threshold influenced by the sensitivity parameter. If this is achieved for 2 subsequent frames, then the alarm is generated. Only the sensitive area from Essential Video Analytics or Intelligent Video Analytics is evaluated. 3. Intelligent Video Analytics (CPP6/7) is running: Intelligent Video Analytics detects a significant illumination change, scene change or camera motion. Neither sensitive area nor sensitivity parameter have any influence. Limitations: - The alarm can also be set off by: o Illumination changes o Crowds of people walking through. o A person walking close to the camera and thus obstructing the camera view - Long-term stability of the uncompromised camera field of view cannot be evaluated - It cannot be distinguished which kind of camera tampering occurred - The alarm will reset as soon as the new scene is stable   3. Reference check Applications: - Detect camera blinding - Detect camera being moved away from of its view - Detect the loss of sufficient overall illumination - Detect idle / removed objects, e.g. objects in front of an emergency exit How it works: - A fixed reference image needs to be set. This reference image is then compared to the current image every 10 sec. For independence to illumination changes, only the edges in both images are evaluated. The evaluation can and should be restricted to significant and stable edges like those from buildings in the image via masking all other areas. As long as these edges are visible, then the field of view is unobstructed, the illumination level good enough, and the camera still on the desired field of view. To avoid false alerts due to short-term changes, both timer and sensitivity can be adapted. - To use the reference check for removed object detection, e.g. a painting on the wall: Restrict the evaluation to the object plus a small border around it. Adapt the sensitivity until every removal of the object raises the alarm. - To use the reference check for idle object detection, e.g. a box before an emergency exit, target the camera view on the floor or restrict the evaluation accordingly. Now select “appearing edges” for the “edge check”, and adapt the sensitivity until the camera alarms on your test objects. - The alarm will consist until the scene is again similar enough to the reference image - Performance may be improved by scheduling two different reference image check profiles for day and night. Limitations: - It cannot be distinguished which kind of camera tampering occurred - The reference check needs edges in the image which stay consistent during all hours and seasons to work properly A scene changing over day, night, and seasons The reference check for tamper detection should be focused on edges that are always visible (marked in red) to obtain best performance. Using the reference check for idle object detection: Focusing the camera view on the floor, setting the reference without any objects included, setting check to “Appearing edges” and increasing sensitivity until the counter “Seconds to alarm” starts to drop from the trigger delay time. As the trigger delay is quite large, here with 10 minutes, objects moving through will not trigger the reference check, only objects staying in the scene for longer than trigger delay. Using the reference check for removed object detection: Focusing the camera view on the object to be secured, setting the reference including this object, setting check to “Disappearing edges”. When the object is removed, the alarm will sound. Here with very short trigger delay for testing purposes only to immediately get the alarm state, increase trigger time for more robustness in practice. ... View more