Analysis and eva‎luation of the method of determining the performance of self-driving cars at intersections in yellow light conditions using artificial intelligence

1/1/1900 12:00:00 AM

The impressive advances in artificial intelligence and self-driving cars have brought unique possibilities for safe and intelligent movement in the performance of cars and have improved the security of these cars and reduced accidents caused by errors. The main goal of this thesis is to design and implement an artificial intelligence system based on deep learning to detect and determine the distance between traffic lights and the self-driving car and the proper performance of the car in yellow light conditions. The necessity of this research is due to accidents caused by drivers' mistakes, including drowsiness while driving. The innovation of this research is the use of the data set of Tehran city and the implementation of a model with more accuracy than the models made in previous studies, as well as the use of a camera to detect the distance between the car and the traffic light. To implement the model, the data set specific to the lights of Tehran city was used, and the model works in the framework of the Tensorflow library. Validation of the model was done with the validation data set of Tehran city data set, and to eva‎luate the performance of the model in different conditions, including the width of the intersection and the speed of the vehicle, the images of the training set were used, which enables the car to use the four scenarios presented in this research, decide to pass through the intersection or stop at the intersection. Analyzing the results of the model by checking the output of the model such as precision, accuracy, recall, F1 score and speed of the models were eva‎luated with the results of past researches and showed that the results are correct and have higher accuracy than the existing models. Also, the best model presented in this research has an accuracy of 96% and precision of 98%. Based on the traffic light data of Tehran city, this system is able to calculate the distance of the car to the traffic light with an error of less than one percent (0.8 percent), which shows the high accuracy of the model that can provide a proper performance in yellow light conditions. This model can be used as a substitute for humans in the detection of traffic lights in self-driving cars in the yellow light mode.

خودرو خودران , چراغ‌راهنمايي , تشخيص فاصله , تك دوربين , يادگيري عميق

self-driving car , traffic light , distance detection , single camera , Deep Learning

Amirhosein Darabpour

Shahriar Afandizadeh