yshastri66 / Zebra_cross_detection

Object detection project which detects Zebra Cross from driver's point of view.

Zebra Crossing Detection

This is the model to predict the Zebra Cross in real time from driver's view.

Key features are :

• This model is trained with 1024*1024 images which is of real time camera quality.
• I used nuScenes deep drive dataset and also some custom images for training
• I took 2300 images out of 67k images from Nuscenes dataset and 200 real images from camera
• Here I used Tensorflow 2 object detection which is the latest object detection library of google
• I used SSD Resnet 50 1024*1024 model for training purpose
• I got DetectionBoxes_Precision/mAP@.50IOU nearly 82% which is good for 2D image detection in road.

Steps to test the model:

1.Clone this repository by downloading zip or by running following command in Git or terminal :

git clone https://github.com/yshastri66/Zebra_cross_detection.git

2.Create a new python environment and install the requirments using requirment.txt file by executing following commands :

conda create -n env_name
conda activate env_name
pip install -r requirments.txt

3.Put all your images which you want to detect in 'test' folder.

4.Run the jupyter notebook image_testing.ipynb

5. comment out following lines if you are not using GPU :

physical_devices = tf.config.list_physical_devices('GPU')
tf.config.experimental.set_memory_growth(physical_devices[0],True)

6. All your images with predictions will be saved in predictions folder.

Sample predicted images are given in the folder sample_predictions. Look into it for better understanding.

If you are having any douts, feel free to contact me below:

Yashodhara Shastri G

Contact : [e-mail]