This project aims to utilize a rasberry pi 3 to calculate the frequency of truck and buses on the street outside my house. Since we live near a FedEx distribution center I am particularly interested in counting the number fedex delivery trucks on my street.

Although I am just getting started it seems very easy to retrain a tensorflow deep learning algorithm through transfer learning, and uploading that model to the rasberry pi. See instructions here and tensorflow retraining here. Additional training images can be found in Imagenet here and from the rasberry pi camera feed,

I started with the Adafruit Rasberry pi 3 camera kit, although it's much cheaper to buy through Arrow.com here. Full kit for 88$ plus free next day shipping (at least to DC)! Camera setup instructions can be found here

cd ~/Documents/tensorflow-for-poets-2/
source activate tensorflow   # activate anaconda environment with tensorflow

IMAGE_SIZE=224
ARCHITECTURE="mobilenet_0.50_${IMAGE_SIZE}"

# retrain 
python -m scripts.retrain \
  --bottleneck_dir=tf_files/bottlenecks \
  --how_many_training_steps=500 \
  --model_dir=tf_files/models/ \
  --summaries_dir=tf_files/training_summaries/"${ARCHITECTURE}" \
  --output_graph=tf_files/retrained_graph.pb \
  --output_labels=tf_files/retrained_labels.txt \
  --architecture="${ARCHITECTURE}" \
  --image_dir=RetrainingPhotos

These are a first pass at retraining the classifier using about 30 examples of fedex trucks and other trucks/cars and only 500 training steps.

python -m scripts.label_image \
    --graph=tf_files/retrained_graph.pb  \
    --image=Testing/car2.jpeg

2017-11-18 12:12:59.255913: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX

notafedextruck 0.999974
fedextruck 2.60127e-05

python -m scripts.label_image     --graph=tf_files/retrained_graph.pb      --image=Testing/fedex2.jpeg
2017-11-18 12:27:00.936555: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX

fedextruck 1.0
notafedextruck 9.88756e-11

python -m scripts.label_image     --graph=tf_files/retrained_graph.pb      --image=Testing/whitetruck.jpeg
2017-11-18 12:28:09.792163: I tensorflow/core/platform/cpu_feature_guard.cc:137] Your CPU supports instructions that this TensorFlow binary was not compiled to use: SSE4.1 SSE4.2 AVX
fedextruck 0.999915
notafedextruck 8.54759e-05

One issue is that each periods images might be a little different based on camera placement. For this reason I wanted to create a simple algorithm for finding the road line, and croping an area of interest (AOI) big enough to capture only cars on the road.

You can find the tutorial here ./Find_Street_Yellow_Line_Tutorial.ipynb.

The following techniques are used:

• Color Selection
• Canny Edge Detection
• Region of Interest Selection
• Hough Transform Line Detection
• Line estimates and AOI masking

Hough Line Detection Results

Linear Estimate of Line Feature

AOI Masking

The next step is to see how our classifier does is the wild. Here we retrain TensorFlows light MobileNet classifier to identify FedEx trucks from my window>

You can find the tutorial here Retrain_Tensor_Classifier_Tutorial.ipynb.

Classification with Probability of Class

Although it gets confused by the purple car, its class probability is very low. Therefore we might still have a good classifier with a little bit of filtering.