Real-Time Traffic Sign Detection
An Era that is bringing ease for the drivers who are getting tired of driving their vehicles manually and are wasting most of the focus and energies on driving rather than keeping that energy for their daily work. This gap is now been filled by the enormous autonomous industry working on a massive level to make self-driving cars capable of producing highly accurate decisions
There is extensive research in Object detection Algorithms nowadays; self-driving vehicles are one of the most focused areas in Research and Development (R&D). Traffic Sign detection System plays a vital role in the Autonomous Vehicle industry. It is very efficient in assisting drivers or self-driving modes in the detection of Traffic Signs. Convolutional Neural Networks gave good results in training and proved efficient, as described in the literature review. The Data-Set was taken from online repositories like Kaggle. So, this technology’s future is bringing innovations and how we will travel throughout in the coming future.
Companies like Tesla have an impact in revolutionizing the world using Autonomous Electric Vehicles. They are using intense research in Pattern Recognition, Making bounding boxes, and learning from their live big datasets towards making models that are highly accurate in solving self-driving cars’ problems.
Following are the Past Algorithms / Methods being applied to solve this problem,
- TensorFlow Library and parallel architecture of multithreaded programming CUDA.
- Viola-Jones detector, a linear classifier based on Histogram of Oriented Gradients (HOG) features and a model-based approach.
- Convolutional Neural Network (CNN)
- Fast Recurrent Convolutional Neural network (Fast-RCNN)
- Instead of using one pixel at a time, this paper focuses on utilizing the block of pixels and giving it an input vector to the SVM for color classification. The block approach is applied to diversify the training and testing.
- Support Vector Machine (SVM). Instead of using one pixel at a time, this paper focuses on utilizing the block of pixels and giving it an input vector to the SVM for color classification. The block approach is applied to diversify the training and testing.
- A modified YOLOv2
- LeNet-5 convolutional neural network
- YOLOv3
The use of artificial intelligence and machine learning in technology facilitates the detection to be more precise and accurate. After going through the literature review and the accuracy matrices of these methods, we decided to go for YOLOv3.
As YOLOv3 proved to be better than other algorithms and more efficient in detecting and classifying traffic sign detection.
We have also done experimentation on Customized CNN, First-of-all we shuffled the dataset and preprocessed them. Then we used three layers, consisting of the following details,
- First Layer: 32 Neurons using RELU Activation and kernel Size of 3*3
- Max pooling using 2*2 window
- Second Layer: 64 Neurons using RELU Activation and kernel Size of 3*3
- Max pooling using 2*2 window
- Dropout by the factor of 0.25
- Third Layer: 128 Neurons using RELU Activation and kernel Size of 3*3
- Flattening
- Two Dense Layers
o First Dense layer of 128 neurons using RELU Activation
o Second Dense layer gives probability distribution in the output using softmax and having neurons equal to the data set class size, 43.
The Loss Function Graph is shown as follows,
The Accuracy Graph is shown as follows,
Then we used a loss function named “categorical_crossentropy” and “Adam” Optimizer. Accuracy Metrics are generated to compare our results. The dataset used in this experimentation is from Kaggle, and we got 99.566 Accuracy.
