Progress

Read the research paper and understood the results.
Downloaded the Bosch data set for traffic lights.
Learning about YOLO architecture.
Downloaded the darknet YOLO framework and setted up the framework to train the custom dataset.
Ran the framework and trained till 1000 images and saved the weights learnt so far.
Started training with bosch data in google colab.

The trained results were not able to detect on the test images- possibly because of error in creating the train.txt and test.txt files.
Retraining the model with the dataset again. 

The issue was with the framework, the training was not perfect but when a different framework from https://github.com/AlexeyAB/darknet was used, the same weights could detect the traffic lights.

07/12 - 08/12
Trained with Yolov4 with mAP of 19%.
Forked the darknet repository to personal github as it was not possible to fork in tamu github.
Made changes to the "detector.c", "image.c" files to provide a json file contained the Left_x, top_y, w and h of the image along with class and Image name details. THis will be used to crop images around the bounding box and create a new data set for the classfication CNN.
Added json file reading functionality along with conversion to image objects with bounding box information. Will be used to crop the images to create new images and data set.
Added required utils
Implemented the util to extract cropped images of the taffic lights from the original image.


09/12 - 10/12
Trainined Yolo with map @0.5 25.27%
Obtained json file of results containing the bounding box information of the images.
Implemented the functionality to create a dataset of cropped images. 
Matched the detected bounding boxes with the yaml file bounding box to read the original labels
Created a dataframe consisting of names of cropped images along with their labels
Created CNN model for training
Tested the classification framework from end to end
Trained the classification network for 50 epochs and achieved an accuracy of 80%.