Change Log

[v1.2.1] - 2021-11-24

Added

• Support for TFLite trained SSDLite MobileNet v2 models

  • TFLite compiled binary wheel files for raspberry pi devices are included in python_wheels/tflite_runtime

    • at present installing through pip is experimental on RPi 4 and not possible on RPi Zero W, must be built from source.

  • animal.py modified to run inference using tflite networks or yolov4-tiny

  • animal.py run_raw and run modified to produce a human_confidence as well as animal_confidence

  • configuration for human detection, fast animal detection (tflite model) and thresholds for each added to settings.

Fixed

- fixed output queue length in `event_rememberer` to 1 not 10, this stops the system loading way too many motion events and exhausting memory (kswaps)
- fixed indexing error in reading from motion buffer within `camera_to_disk.py`

Changed

- methods which use `animal.run` have been modified to accept human and animal values
- changes in `setup.sh` insure tflite install is attempted 

[v1.2.0] - 2021-10-31

Added

• Camera recording to disk module, modified PiCamera; DynCamera

• EventRememberer which to load events from disk into processing pipeline

• Filtermodes BY_FRAME and BY_EVENT to add functionality to process events with Filter module

• Event output added to comms.py

• modified main.py to support old and new pipelines depending on settings

• Modified settings to include parameters for: pipeline, bitrate, framerate divisor, buffer length, raw image format and detector fraction

Changed

• Motion Queue Settings semantics changed to processing settings. To better fit other pipeline which does not use a motion queue.

[v1.1.0] - 2021-08-15

Added

• Check for settings.json version vs. DynAIkonTrap version in case settings are copied from one trap to another

• Added support for multiple output video codecs and settings to choose between them

• Pillow to requirements.txt easiest way to load raw images as far as I can tell. If this can be done with OpenCV it would be nicer.

Fixed

• Implementation of UrSense interface following updated documentation

• Catches pycamera ModuleNotFoundError when running the camera trap with emulated input on desktop

Changed

• Video sensor logs to JSON for easier machine reading – parsing this back to the previous VTT output is trivial

• Interface to initialise Output – output mode is now handled internally

• Documentation – including wiki – migrated to Sphinx

Added

• context buffer so that clips of animals include “run in” and “trail off” frames.

• LabelledFrames now include a motion_status label of enumerated type MotionStatus

• filtering.py adds all frames to a motion_sequence regardless of motion score but labels frames passing through it.

• frames without motion are assigned a priority of -1.0 and a MotionStatus of STILL this ensures they are never returned by get_highest_priority() - thus never assessed for containing an animal.

• MotionQueue does not add motion sequences to its queue which do not contain motion. ie end_motion_sequence() now searches the sequence to make sure at least one frame is labelled with motion before appending to queue.

Changed

• MotionSequence class is now called Sequence

• MotionQueue class is now called MotionLabelledQueue

---

[v1.0.0] - 2021-06-12

Added

• First release of DynAIkonTrap