A project for detecting and counting mountain bikers and people walking in the forrest using a small bird house looking box, containing a low resolution thermal camera and a Raspberry Pi 2. The recording time is limited by the battery capacity, which currently provides us with around 12 hours of recordings.
This overall todo list is prioritized according to what needs to be done first. The goal in general is to create a poster and then a paper based on the event detection system for monitoring nature trails.
- Implement improvements to e.g. the detection and tracking
Capturingand annotating dataAutomated data collection by use of e.g. data connection + presentation of dataBringing the HW and SW of the prototype to a stage where the system can easily be used by non-technical persons
unfortunately, we were unable to access the camera and therefore some tasks could not be completed and a lot of time was spend on attempting to get it working
Milestones:
- Implement working Hungarian tracking in python
- Clean up data by removing beginning and end + fix missing frames
- Annotate remaining clips
- Run detector and tracker on all clips e.i. generate results
- Summerize clip information
- Present detection and tracking results
- Poster
Missing:
- Measure detector and tracker performance on raspberry pi 2
- Verify annotations
- Optimize detector parameters
- Evaluate detector https://sourceforge.net/p/dclib/discussion/442518/thread/f7f5a16d/
- Optimize tracker parameters
- Evaluate tracker
- When evaluating the detector extract some correct and some wrong detections for the paper and poster. also try to perform clustering based on the features and visualize the distribution along with labels.(example: http://cs.stanford.edu/people/karpathy/cnnembed/)
- Visualize the tracks along with the frames
Try dlib with neon enabled:
https://gist.github.com/ageitgey/1ac8dbe8572f3f533df6269dab35df65
sudo python3 setup.py install --compiler-flags "-mpfu=neon"
| Name | Time(wrong) | Place | Start time | BB anno | Event anno | Comments | Usage |
|---|---|---|---|---|---|---|---|
| 2015-09-02-12-44 | Sat 29 Sep | Hammer Bakker | 11:00:00 | ✔️ | ✔️ | ok | train |
| 2015-09-28-12-39 | Sun 30 Sep | Kongshøj | 11:00:00 | ✔️ | ✔️ | ok | train |
| 2015-09-29-12-43 | Kongshøj | Poor view | |||||
| 2015-09-30-06-23 | Fri 2 Oct | 11:00:00 | ok | test | |||
| 2015-09-30-11-17 | Sat 3 Oct | Hammer Bakker | 11:00:00 | difficult | |||
| 2015-10-05-15-56 | Man 5 Oct | Kongshøj | 11:00:00 | occlusion | test | ||
| 2016-08-08-11-00 | Man 8 Aug | Hammer Bakker | 11:00:00 | ✔️ | ✔️ | ok | train |
| 2016-08-09-14-58 | Wed 10 Aug | Hammer Bakker | 11:00:00 | ✔️ | ✔️ | ok | test |
| Sun 14 Aug | Månedalen | 11:00:00 |
| Name | Hours | MTB Count | PED Count |
|---|---|---|---|
| 2015-09-02-12-44 | 05:32:21 | 47 | 6 |
| 2015-09-28-12-39 | 04:07:39 | 64 | 2 |
| 2015-09-30-06-23 | 05:17:00 | 42 | 5 |
| 2015-10-05-15-56 | 06:33:47 | 47 | 4 |
| 2016-08-08-11-00 | 05:39:40 | 29 | 5 |
| 2016-08-09-14-58 | 05:39:35 | 26 | 3 |
| Totals: | 32:50:02 | 255 | 25 |
The project is primarily written in python, originally some parts were written in c++ but later every component should now be available in python.
More details on using either of the two primary groups of code is described in detail in their respective readmes.
"thermal capture" contains the program and scripts that runs on the raspberry pi for capturing and storing 16 bit 80x60 images at 7 fps to sd card.
"data processing" contains code for preparing and processing the frames after they have been captured on the raspberry pi.
detection
evaluation
The annotations are located under /data/annotations/ in this repository. TODO: annotate events in remaining clips(4)
Event annotation using RUBA
BB annotation using BBA
Computer: Raspberry Pi 2
Thermal camera: Lepton LWIR 80 × 60
Power: 10.400mAh