The VE352 video engine detects left or removed items in moderately to less busy indoor environments. Typical deployments include retail facilities where detecting removal of merchandise is desired. Camera position should be at least 15 feet above the ground, and the camera angle can be flexible, depending on the field of view you wish to capture.
Task 1: Configure VE352
To configure the VE352:
1. In Aimetis Symphony Client, from the Server menu, select Configuration. The Configuration dialog box opens. This allows you to configure devices for the currently selected server.
2. In the left pane, click Devices. The Devices dialog box opens.
3. Click the camera you want to configure for use with video analytics and click Edit.
4. Click the Analytics Engines tab and select VE352. (Optional) By default VE250 is selected, in which case you must first deselect VE250.
5. Click the Analytics Configuration tab. From the Analytic Engines drop-down box, select VE352.
6. By default, the Masks tab is shown. Leave the default settings for the Analysis FPS and Analysis Resolution.
7. Click the Persistence Mask subtab and adjust the yellow mask as required.
• Select the Erase option to erase the yellow mask.
• Select the Draw option to draw the yellow mask.
• The Size slider adjusts the pen thickness.
8. Click the Perspective subtab and set the perspective for the scene: The red lines must be drawn such that they represent a length of 5m at their location in the scene. The Person icon will be sized automatically.
9. Click the Basic tab to configure additional settings.
Option |
Usage |
Left item duration (seconds) |
Specifies how long an object must be left before an alarm will occur. Increasing the value: • Reduces false alarm due to person standing still, reflection of people walking near by, etc. • However the person would have left the scene before an alarm is obtained. Decreasing the value: • Increases false alarms because there is not enough time to determine if it is a valid alarm • The person who left or removed the item might still be in the scene when the alarm occurred. |
Initial background model learn time (seconds) |
Specifies how many seconds will be used to build background model. Set this low unless you know for sure there will not be any moving objects during start of camera because the longer the initial background model is learned the more certain the engine is of the initial background. During this time the engine is not running and as such it will not detect any left/removed items. |
Object merge distance (pixels) |
Determines the distance (in pixels) between objects where they will be automatically merged as one object, that is, the minimum distance between two adjacent alarmable objects. Increasing the value: • If parts of a single object are detected as separate objects then they are merged together into one object. • There is a risk of two actual left/removed objects being merged together. Decreasing the value: • There is a risk of getting multiple alarms on the same object. • Ensures that two adjacent left/removed items are detected as two different objects. |
Check for object drop-off |
This enables the software to check for activity around the left or removed item seconds before the alarm occurred (such as a hand reaching in). The pixel value defines the search area around the alarm.
The reach width determines how big of an area around the left/removed item you want to search for motion.
When selected, this option will reduce false alarms due to reflections, sudden lighting changes, etc. However there is a risk that a true alarm will be missed. For example when a person drops off or removes an item quickly the item might not be alarmed on.
Increasing the reach width makes it easier to find the removing/adding motion; however, with a larger value, more motion is needed before the alarm is triggered. |
Reduce false alarms from still people |
May be used to reduce alarms caused by people standing in scene. The seconds value will determine how long to filter out still objects. A person can come into the scene and stand still for longer than the Left item duration causing a false alarm. This option can be used to remove these false alarms.
When this option is enabled, an object of a person size standing still is alarmed on after Time to wait for still people (seconds) NOT after Left item duration (seconds). It is assumed a person will move within Time to wait for still people (seconds), which is set larger than Left item duration. This option ONLY affects objects of a person size as determined by the perspective configuration information. |
Reduce reflection |
Enables another filter that attempts to reduce false alarms caused by reflections. The slider controls how aggressively to eliminate reflections.
Turning on the feature: • Reduces false alarms due to reflections • Increases chance of missed alarms Move slider to right • Less aggressive reflection control • More alarms due to reflections but less missed alarms Move slider to left • More aggressive reflection control • Less alarms due to reflections but more chance of missed alarm |
Camera looking at shelf |
This should be enabled or disabled depending on camera location relative to shelf. |
10. Click OK to save configuration changes.
11. The default settings in the Advanced tab are sufficient, but you may have cases where adjusting values would be beneficial. Consider the following:
• The advanced settings are a trade-off between detecting false alarms and missing true alarms.
• The default values were set with the assumption that for 99% of the time you do not want to miss any alarms even if some are false.
• Ideally place the camera such that there is little or no occlusion of the alarmable region by a person. This way you can minimize the false alarms due to people standing still. This is particularly important if you have a very small Left Item Duration below 60 sec.
• If you cannot place the camera so there is no occlusion of alarmable region by people, place camera such that the entire person is in view when they occlude the alarmable region AND set Time to wait for still people high; at least 60 sec, perhaps more.
• Also ensure that the perspective is set properly.
• Zooming too much can at times cause false alarms do to the following reasons:
• Parts of people might be detected as left/removed items.
• Slightest lighting change will cause a large camera gain change, which will cause false or missed alarms.
Option |
Description |
Minimum fill rate % |
The default value was found to work for almost all cases. This value should be changed only as a last resort.
Increasing this value: • Reduces false alarms • Increases chance of missed alarms (starting with objects on a diagonal) Decreasing this value: • Increases the chance of false alarms • Ensures most true alarms are detected |
Time to recover or drop an item (seconds) |
Determines how long [max time] it takes to remove or drop an item. |
Minimum valid object dimension (pixels) |
Increasing the value: • Reduces false alarms on noise • Does not detect any small object. • Does not detect any object that looks similar to the background. (This happens in retail situations where the same products are stacked together.) Decreasing the value: • Increases false alarms for very small objects • Increases the chance of detecting all true small alarms |
Merge small objects closer than (pixels) |
Similar to the Object merge distance option from the Basic tab but is used to merge objects that are smaller than the Minimum valid object dimension. If there are a lot of very small objects very close by, then when combined they should make a single valid alarmable object. Increasing the value: • Increases false alarms but ensures you detect some objects that look similar to the background. Decreasing the value: • Reduces false alarms but misses some objects that looks similar to the background. We recommend that you keep this value small. |
Grayscale background update rate |
Controls the long term background update rate.
A fast update means only most recent frames are considered in obtaining the background. A slow update means only the older frames are considered in obtaining the background. Mainly effects the reflection reduction settings. Increase the update rate if you want to eliminate reflection objects but there is also a greater chance of missed detection. |
Time to clear long-term background (seconds) |
Determines how long before a learned long-term background is forgotten. The larger the value, the longer the long term background is remembered even after being occluded by some other object. If an item is dropped then removed before Time to clear long-term background seconds, then no alarm will occur for the remove event. |
Time to clear short-term background (seconds) |
Determines how long before a learned short-term background is forgotten. This value determines how flexible the engine is in adapting to changing environments without causing false or missed alarms. Large value: • Ideal for less busy environments Small value: • Ideal for very busy environments |
Background reset time after an object is deleted (seconds) |
Determines how long to reset after a valid change object is discovered. During this time another object at the current location is not allowed. Large Value: • Decreases multiple alarms on same object. • Might miss a left item at the same spot, within this time period. Small Value: • Might have multiple alarms on same object. • Will not miss multiple objects at the same spot, within this time period. |
Reset camera if this percentage of image becomes foreground in 1 frame |
In two situations, you want to stop analysis and restart everything: • Sudden change in the video between frames (camera moves) • Slow change in the video within 5 frames (florescent lighting turned on) |
Reset camera if this percentage of image becomes foreground in 5 frames. “ |
In two situations, you want to stop analysis and restart everything: • Sudden change in the video between frames (camera moves) • Slow change in the video within 5 frames (florescent lighting turned on) |
Task 2: Creating a Rule using VE352
To configure a Rule using the VE352:
1. You must configure the VE352 before creating a Rule that uses the VE352.
2. From the Server menu, select Configuration and then Rules
3. Click New. The Rule Wizard opens. Read the overview and click Next.
4. Select the check box next to the camera name. If it is a PTZ camera, select the camera tour position check box.
5. From the Select video engine for those running on camera drop-down list, select VE352.
6. Using the displayed image from the video, define the area in red (a mask) where alarms should be set.
• Select the Erase option to erase the red mask.
• Select the Draw option to draw the red mask.
• The Size slider adjusts the pen thickness.
• The VE352 does not distinguish between a left or removed item.
• Objects that are left or removed outside the red alarm zone will not cause alarms.
7. Specify what percentage threshold for an object, which is not fully inside an alarm zone, by adjusting the Alarm based on percentage of object inside alarm mask option.
8. (Optional) Alarm conditions may be filtered by enabling the Alarm only if, or OR, at least options. Both conditions can be enabled together or separately.
• The Alarm only if filter is designed for retailers who want to alarm only if a certain number of items are removed from a shelf in a pre-determined time. For example, 2 LCD monitors must be removed before the event is triggered.
• The OR, at least option is a work-around if multiple objects are removed but they are so close together it creates one bounding box (and therefore one alarm).
• Therefore by combining both filters, many small background changes or one large background change (caused by a removed item) can trigger the event.
• Alarm only, if: Defines how many alarms must occur in predetermined time
• OR, at least: Defines the size of alarm before
• Alarms Occur: Based on alarms detected
• Pixels Change: Based on pixels changed
9. Click Next to continue in the Rule Wizard.
Currently, there is no distinction between left and removed items.
If objects are too small, detection is difficult. Move camera closer to objects that are to be detected.