Documentation for Detector Pack 2.7. Documentation for other versions of Detector Pack is available too.
New Features and Enhancements
- Updated standard Object Tracker
- Implemented several neural network filters for the standard Object Tracker. Each filter classifies moving objects of a certain type only and ignores all other motion in FOV
- Introduced a neural network–based posture detection tool. Added filters which detect the following postures: standing, crouching, lying prone, shooting, hands up
- Introduced a neural tracker which uses neural networks to analyze video. The neural network classifies moving and stationary objects on which it has been trained, then the neural tracker tracks these objects. The new tracker works with all classical VMDA detection tools: movement in the area, line crossing, etc
- The neural tracker now supports the GPU of the NVIDIA GeForce GTX 1660 video board.
- Introduced counting objects detected by the neural tracker within a specified area in FOV. One neural tracker can work with several counters set up individually
- The neural tracker is now capable of generating events on appearing and disappearing objects. Added a macro that returns statuses of all detected objects: their IDs, coordinates, date and time
- Updated neural algorithms for fire and smoke detection. You can now manually select between processing hardware: GPU or CPU
- Added Alarm Expired events to fire and smoke detection tools
- Updated neural network theft detection tool at checkout (Sweethearting). Optimized memory usage
- The CAM_VMDA_DETECTOR event now contains directional vector (x, y)
- Added the VMDA.filterGivenOrTaken key for the abandoned objects detection tool which allows selecting the event type: object abandoned, object disappeared, or both
- For the queue length detection tool, the queue length threshold crossing event has been added to the Event Log
- Improved the stopped vehicle detection tool
- Detector Pack can now be installed in the background
- A second detection tool is now capable of generating alarms if the VMDA tracker operates in a separate process.
- Corrected a rare bug that prevented the VMDA tracker from launching in a separate process.
- Fixed a bug that blocked neural filter operation if the tracker's ID does not match the camera's.
- Fixed a rare bug that made neural filter crash when launched on a GPU.
- Devices available for neural filter operation are now displayed correctly.
- Detector Pack now uses the stream intended for analytics operation.
- No more video frame display delays occur in Operator Protocol on detection.
- Eliminated multiple alarms if two VMDA detection tools with different neural filters operate on the same camera.
- Fixed an issue that occurred when two trackers share a camera: one generated false alarms depending on the other's settings.