TrafficLight_Detection_Classification

Transfer learning has been used to detect and classify traffic lights. The following models were considered -

• ssd_mobilenet_v1_coco_2018_01_28
• faster_rcnn_resnet101_coco_2018_01_28

Model Training Steps

An AWS Deep Learning AMI has been used for training the model. The following steps were used -

Installing Dependencies

The training dependencies can be installed by following the steps outlined on this page

Download Tensorflow models

git clone https://github.com/tensorflow/models.git
cd models/research

export PYTHONPATH=$PYTHONPATH:`pwd`:`pwd`/slim

cd object_detection

Training data

There are 2 options for training data. Download Training data or build training data by using annotation tools like LabelImg and include the TFrecord file in the ./data folder

Download models from Tensor flow training zoo to start transfer learning

# Ensure that the models are downloaded into the /research/object_detection folder
wget http://download.tensorflow.org/models/object_detection/ssd_mobilenet_v1_coco_2018_01_28.tar.gz

unzip ssd_mobilenet_v1_coco_2018_01_28.tar.gz

Download required files

# Verify that the current directory is (/models/research/object_detection)
wget https://raw.githubusercontent.com/vishal-kvn/TrafficLight_Detection_Classification/master/config/ssd_mobilenet_v1.config

cd /data

wget https://raw.githubusercontent.com/vishal-kvn/TrafficLight_Detection_Classification/master/data/traffic_light_label_map.pbtxt

Run training

python model_main.py \
--pipeline_config_path='./ssd_mobilenet_v1.config' \
--model_dir='training_ssd_mobilenet_v1_sim/' \
--sample_1_of_n_eval_examples=1 \
--alsologtostderr

Export inference graph

python export_inference_graph.py \
--input_type image_tensor \
--pipeline_config_path ./ssd_mobilenet_v1.config \
--trained_checkpoint_prefix training_ssd_mobilenet_v1_sim/model.ckpt-5000\
--output_directory ssd_mobilenet_v1_inference_graph

Validate model using the notebook

jupyter notebook --ip=0.0.0.0 --no-browser #Ensure that a rule is added to the security group to accept inbound traffic on port 8888.

Copy the IP address of the ec2 instance and view the notebook and run all cells

Detection and Classification Results