tooth2 / Traffic-Light-Classification

A open-cv and CNN(Tensorflow) implementation to detect Traffic light (red, yellow, green) state

Traffic-Light-Classification

A open-cv and CNN(Tensorflow) implementation to detect Traffic light (red, yellow, green) state

This project is to build a classifier for images of traffic lights usingg Open CV(1) and CNN(2) for given a dataset of traffic light images in which one of three lights is illuminated: red, yellow, or green.

The tasks will be broken down into a few sections: 0. Loading and visualizing the data : load in the images of traffic lights and visualize some of them

1. Pre-processing: The input images and output labels need to be standardized. This way, all the standardized input images are processed in the same classification pipeline, and output when you eventually classify a new image.
2. Feature extraction: extract some features from each image that will help distinguish the different types of images, and use those features to classify the traffic light images into three classes: Red, Yellow, or Green
3. Accuracy : Classification and visualizing errors that uses features to classify any traffic light image. This function will take in an image and output a label. 4. Evaluate your model: The accuracy of the classifier must be >90% accurate and never classify any red lights as green; improve the accuracy of the classifier by changing existing features or adding new features.

Here are some sample images from the dataset (from left to right: red, green, and yellow traffic lights):

The Data Set

This traffic light dataset consists of 1484 number of color images in 3 categories - red, yellow, and green. As with most human-sourced data, the data is not evenly distributed among the types.

• 904 red traffic light images
• 536 green traffic light images
• 44 yellow traffic light images Note: All images come from this MIT self-driving car course and are licensed under a Creative Commons Attribution-ShareAlike 4.0 International License.

Traning and Testing Data

All 1484 of the traffic light images are separated into training and testing datasets.

• 80% of these images are training images
• 20% are test images, which will be used to test the accuracy
• All images are pictures of 3-light traffic lights with one light illuminated.

Pre-process the data

• One hot encoded labels:
  • Red light label: [1, 0, 0]
  • Yellow light label: [0, 1, 0]
  • Green light label: [0, 0, 1]
• Resize each image to the desired input size: 32x32px

Feature extraction

• standardized image
• HSV color-masked image
  • 
• cropped image
• brightness feature (using HSV color space): convert RGB to HSV and identify 3 different classes of traffic light

Apply CNN - Define Model

Layer (type)	Output Shape	Param #
batch_normalization_1	(Batch (None, 32, 32, 3)	12
conv2d_1 (Conv2D)	(None, 30, 30, 16)	448
max_pooling2d_1	(MaxPooling2 (None, 15, 15, 16)	0
batch_normalization_2	(Batch (None, 15, 15, 16)	64
conv2d_2 (Conv2D)	(None, 13, 13, 32)	4640
max_pooling2d_2	(MaxPooling2 (None, 6, 6, 32)	0
batch_normalization_3	(Batch (None, 6, 6, 32)	128
conv2d_3 (Conv2D)	(None, 4, 4, 64)	18496
max_pooling2d_3	(MaxPooling2 (None, 2, 2, 64)	0
batch_normalization_4	(Batch (None, 2, 2, 64)	256
global_average_pooling2d_1	( (None, 64)	0
dense_1 (Dense)	(None, 3)	195

Result

• OpenCV test accuracy : 96.6330%
  • 10 misclassified traffice lights:
• Deep Learning(CNN) accuracy: 100.0000%
  •