The easiest way to count pedestrians, cyclists, and vehicles on edge computing devices or live video feeds.
The Raspberry Pi Urban Mobility Tracker is the simplest way to track and count pedestrians, cyclists, scooters, and vehicles. For more information, see the original blog post [here].
Primary Components
Secondary Components
Notes
geometry/
are currently represented as stl
files which are 3d printer ready. I don't currently have a 3d printer so I used the crowd sourced printing service https://printathing.com/ which yielded great results (kind of a sales pitch, but not really. I just like the service).3D printed mounts | mounts with attached hardware |
Final setup (open) | Front (closed) |
pi
:sudo usermod -aG docker pi
UMT_DIR=${HOME}/umt_output && mkdir -p ${UMT_DIR}
cd ${UMT_DIR}
wget https://raw.githubusercontent.com/nathanrooy/rpi-urban-mobility-tracker/master/Dockerfile
docker build . -t umt
docker run --rm -it --privileged --mount type=bind,src=${UMT_DIR},dst=/root umt
wget https://github.com/nathanrooy/rpi-urban-mobility-tracker/raw/master/data/videos/highway_01.mp4
umt -video highway_01.mp4
If everything worked correctly, you should see a directory labeled output
filled with 10 annotated video frames.
First, create a new virtualenv, initialize it, then install the TensorFlow Lite runtime package for Python:
pip3 install --extra-index-url https://google-coral.github.io/py-repo/ tflite_runtime
Then finish with the following:
pip install git+https://github.com/nathanrooy/rpi-urban-mobility-tracker
Lastly, test the install by running step #6 from the Raspberry Pi install instructions above.
The default deep learning model is the MobileNet v1 which has been trained on the COCO dataset and quantized for faster performance on edge deployments. Another good model choice is PedNet which is also a quantized MobileNet v1 however, it's been optimized specifically for pedestrians, cyclsts, and vehicles. To use PedNet, simply download it from its repo here: https://github.com/nathanrooy/ped-net or clone it.
git clone https://github.com/nathanrooy/ped-net
Once the model and labels have been downloaded, simply use the modelpath
and labelmap
flags to specify a non-default model setup. As an example:
umt -camera -modelpath pednet_20200326_tflite_graph.tflite -labelmap labels.txt
Since this code is configured as a cli, everything is accessible via the umt
command on your terminal. To run while using the Raspberry Pi camera (or laptop camera) data source run the following:
umt -camera
To run the tracker on an image sequence, append the -imageseq
flag followed by a path to the images. Included in this repo are the first 300 frames from the MOT (Multiple Object Tracking Benchmark) Challenge PETS09-S2L1 video. To use them, simply download/clone this repo and cd into the main directory.
umt -imageseq data/images/PETS09-S2L1/
To view the bounding boxes and tracking ability of the system, append the -display
flag to view a live feed. Note that this will greatly slow down the fps and is only recommended for testing purposes.
umt -imageseq data/images/PETS09-S2L1/ -display
By default, only the first 10 frames will be processed. To increase or decrease this value, append the -nframes
flag followed by an integer value.
umt -imageseq data/images/PETS09-S2L1/ -display -nframes 20
To persist the image frames and detections, use the -save
flag. Saved images are then available in the output/
directory.
umt -imageseq data/images/PETS09-S2L1/ -save -nframes 20
To run the tracker using a video file input, append the -video
flag followed by a path to the video file. Included in this repo are two video clips of vehicle traffic.
umt -video data/videos/highway_01.mp4
In certain instances, you may want to override the default object detection threshold (default=0.5). To accompish this, append the -threshold
flag followed by a float value in the range of [0,1]. A value closer to one will yield fewer detections with higher certainty while a value closer to zero will result in more detections with lower certainty. It's usually better to error on the side of lower certainty since these objects can always be filtered out during post processing.
umt -video data/videos/highway_01.mp4 -display -nframes 100 -threshold 0.4
To get the highest fps possible, append the -tpu
flag to use the Coral USB Accelerator for inferencing.
umt -imageseq data/images/PETS09-S2L1/ -tpu
@inproceedings{Wojke2017simple,
title={Simple Online and Realtime Tracking with a Deep Association Metric},
author={Wojke, Nicolai and Bewley, Alex and Paulus, Dietrich},
booktitle={2017 IEEE International Conference on Image Processing (ICIP)},
year={2017},
pages={3645--3649},
organization={IEEE},
doi={10.1109/ICIP.2017.8296962}
}
@inproceedings{Wojke2018deep,
title={Deep Cosine Metric Learning for Person Re-identification},
author={Wojke, Nicolai and Bewley, Alex},
booktitle={2018 IEEE Winter Conference on Applications of Computer Vision (WACV)},
year={2018},
pages={748--756},
organization={IEEE},
doi={10.1109/WACV.2018.00087}
}