Prohibitory Sign Detection Using Machine Learning

Abstract

With advancement, technology is becoming increasingly creative in every field. It is atomized using machine learning and artificial intelligence approaches. The detection of traffic signs is a core technology used in automated vehicles. Real-world applications of traffic sign recognition include self-driving, traffic monitoring systems, and safe driving. Very limited work has been done on the Pakistani traffic signs after finding a huge gap and knowing the importance of growing technology and demand. In the results, we first created a dataset since Pakistan lacked a dataset of this kind, and then we used algorithms for machine learning to produce the desired outcomes. To process the gathered images and recognize traffic signs, we employed a variety of deep learning algorithms, including CNNs (convolutional neural networks) and SVMs (support vector machines). We have used YoloV4 and Darknet53 with pre-trained weights of conv.137 for detection purposes, and we have used Google's GPU for training due to limited resources. We got an accuracy of 83 on the new dataset (mAP=83). Although the accuracy of 83 was promising, it was challenging to keep this rate with our limited resources, as we had to terminate the training process after three weeks on Google's GPU after 16,600 iterations and at 83 Mean Average Precision. Then, I took a few real-world images of different classes to check the performance of the training model. The results show the images with boundary boxes and predicted class categories.