traffic_light_detection

Implementation of a traffic light detector for the udacity sdc nanodegreee using tensorflow object detection API

1. download the Bosch dataset

You should first download the Bosch Small Traffic Lights Dataset, and rearrange the folder structure as following:

data
├── train.yaml
├── test.yaml
├── additional_train.yaml
├── rgb
│   ├── train
│   ├── test
│   ├── additional
│   │   ├── 2015-10-05-10-52-01_bag
│   │   │   ├── 24594.png
│   │   │   ├── 24664.png
│   │   │   └── 24734.png
...

Note that the image paths contained in the test.yaml files should be modified to the relative path to the data folder.

2. install the tensorflow object detection API

You can find the tensorflow object detection API here, and follow the installation instruction here.

3. create the tfrecord data file

python create_tf_record.py

4. train the detection model

4.1 download the pre-trained models

download faster_rcnn_resnet101_coco_2017_11_08 and ssd_mobilenet_v1_coco_2017_11_17 model files from here, put the model files in models folder as following:

models
├── faster_rcnn_resnet101_coco_2017_11_08
│   ├── frozen_inference_graph.pb
│   ├── ...
├── ssd_mobilenet_v1_coco_2017_11_17
│   ├── frozen_inference_graph.pb
│   ├── ...
├── faster_rcnn_resnet101_traffic_lights
├── ssd_mobilenet_v1_traffic_lights

4.2 train the ssd_mobilenet model

python object_detection/train.py --logtostderr \
  --pipeline_config_path=config/ssd_mobilenet_v1_traffic_lights.config \
  --train_dir=models/ssd_mobilenet_v1_traffic_lights

4.3 train the faster_rcnn_resnet101 model

python object_detection/train.py --logtostderr \
  --pipeline_config_path=config/faster_rcnn_resnet101_traffic_lights.config \
  --train_dir=models/faster_rcnn_resnet101_traffic_lights

4.4 export the detection model

python object_detection/export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path config/ssd_mobilenet_v1_traffic_lights.config \
    --trained_checkpoint_prefix models/ssd_mobilenet_v1_traffic_lights/model.ckpt-30727 \
    --output_directory models/ssd_mobilenet_v1_traffic_lights/export
    
python object_detection/export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path config/faster_rcnn_resnet101_traffic_lights.config \
    --trained_checkpoint_prefix models/faster_rcnn_resnet101_traffic_lights/model.ckpt-25445 \
    --output_directory models/faster_rcnn_resnet101_traffic_lights/export
    
python object_detection/export_inference_graph.py \
    --input_type image_tensor \
    --pipeline_config_path config/faster_rcnn_inception_v2_traffic_lights.config \
    --trained_checkpoint_prefix models/faster_rcnn_inception_v2_traffic_lights/model.ckpt-28506 \
    --output_directory models/faster_rcnn_inception_v2_traffic_lights/export
    

5. evaluate the detection model

python object_detection/eval.py --logtostderr \
    --checkpoint_dir=models/faster_rcnn_resnet101_traffic_lights/export \
    --eval_dir=models/faster_rcnn_resnet101_traffic_lights/eval \
    --pipeline_config_path=config/faster_rcnn_resnet101_traffic_lights.config
    
python object_detection/eval.py --logtostderr \
    --checkpoint_dir=models/ssd_mobilenet_v1_traffic_lights/export \
    --eval_dir=models/ssd_mobilenet_v1_traffic_lights/eval \
    --pipeline_config_path=config/ssd_mobilenet_v1_traffic_lights.config

run detection_test.ipynb

6. prepare the classifier data

python prepare_classifier_data.py

7. train and evaluate the classifier

run train_classifier.ipynb

8. export the classifier model

 python freeze_graph.py --input_graph=models/classifier/graph.pb \
    --input_checkpoint=models/classifier/model.ckpt \
    --input_binary=true \
    --output_graph=models/classifier/frozen_graph.pb \
    --output_node_names=prediction

9. test the tl_classifier

 python test_tl_classifier.py input output