Traffic Lights Detection and Classification

This is a fork of original Bosch code, modified by Kung Fu Panda team to use in Udacity Self-Driving Car Engineer Nanodegree Capstone project.

Setup

The Bosch Small Traffic Lights Dataset can be downloaded here.

• Please only download rgb named archives.
• Put the files in data folder.
• Concatenate multi-part zip archives like cat x.zip.001 x.zip.002 > z.zip
• Extract the files and re-arrange so the folder structure is as follows:

data
├── rgb
│   ├── additional
│   │   ├── 2015-10-05-10-52-01_bag
│   │   │   ├── 24594.png
│   │   │   ├── 24664.png
│   │   │   └── 24734.png
│   │   ├── 2015-10-05-10-55-33_bag
│   │   │   ├── 56988.png
│   │   │   ├── 57058.png
...
│           ├── 238804.png
│           └── 238920.png
├── rgb
│   ├── train
...
├── rgb
│   ├── test
...
├── additional_train.yaml
├── test.yaml
└── train.yaml

You can verify/view the data using:

python dataset_stats.py data/train.yaml
python show_label_images.py data/train.yaml

Extracting Traffic Light Images from Annotated Pictures

In order to train a classifier we created a script to extract and save actual traffic lights images into separate folders. You can do the extraction as follows:

python save_tl_images.py data/train.yaml data/tl-extract-train

You may see the following warnings:

libpng warning: Image width is zero in IHDR
libpng error: Invalid IHDR data

Please ignore them. The saved pictures are valid and usable.

The resulting images (of variable sizes) are saved with the following naming convention, i.e. sequential number padded to 6 digits then lower-case name of the class.

-rw-r--r--      1 alexeysimonov  staff     177 27 Aug 15:55 000001_yellow.png
-rw-r--r--      1 alexeysimonov  staff     188 27 Aug 15:55 000002_yellow.png
-rw-r--r--      1 alexeysimonov  staff     245 27 Aug 15:55 000003_yellow.png
-rw-r--r--      1 alexeysimonov  staff     159 27 Aug 15:55 000004_redleft.png
-rw-r--r--      1 alexeysimonov  staff     200 27 Aug 15:55 000005_red.png
-rw-r--r--      1 alexeysimonov  staff     257 27 Aug 15:55 000006_red.png
-rw-r--r--      1 alexeysimonov  staff     228 27 Aug 15:55 000007_redleft.png
-rw-r--r--      1 alexeysimonov  staff     265 27 Aug 15:55 000008_red.png
-rw-r--r--      1 alexeysimonov  staff     220 27 Aug 15:55 000009_red.png
-rw-r--r--      1 alexeysimonov  staff     329 27 Aug 15:55 000010_red.png
-rw-r--r--      1 alexeysimonov  staff     199 27 Aug 15:55 000011_redleft.png
-rw-r--r--      1 alexeysimonov  staff     352 27 Aug 15:55 000012_red.png
-rw-r--r--      1 alexeysimonov  staff     195 27 Aug 15:55 000013_redleft.png
-rw-r--r--      1 alexeysimonov  staff     307 27 Aug 15:55 000014_red.png
-rw-r--r--      1 alexeysimonov  staff     244 27 Aug 15:55 000015_red.png

Train a Classifier

We have defined TLClassifierCNN in tl_classfier_cnn.py loosely based on CIFAR-10 network architecture. To train it update training parameters in train.py and run as follows:

python train.py

It saves checkpoints and summaries as it goes along. At the end it saves the model.

Predictions

TLClassifierCNN can load a pre-trained model to run predictions. The following script demonstrates the model loading and prediction:

python predict.py