Improving Traffic Flow for Emergency Vehicles Using Deep Learning Techniques

The world's population has exponentially grown, which has an effect on usage of vehicles by individuals and leads to an increase in the number of cars in urbans. With the direct relationship between population and car usage, traffic management has become an important issue to be solved. For this purpose, an intelligent traffic signaling with a rapid urbanization is required to overcome the traffic congestions, and reduce cost and time of traveling. To overcome these problems, emerging computer vision and deep learning are vital candidates to handle this issue because they take an important role for managing and controlling traffic signals with great success. Nevertheless, detecting and distinguishing between objects are helpful for counting vehicles and other objects which avoid crowds and controlling signals in the traffic areas. Besides, detecting emergency vehicles and giving the priority to them is required for intelligent traffic signaling system.

The main objective of this study is to design and implement an efficient system for traffic signal systems based on custom vehicle detection.  Furthermore, the proposed system involves four phases; the first one is capturing images from both simulated and real time cameras from the roads. In the second phase, different image preprocessing algorithms are performed to the captured images as a pre-processing step. In addition, the deep learning techniques are applied to detect objects such as (regular car, police car, ambulance, and firefighter, etc..). In the last phase, the proposed system is tested to evaluate the performance accuracy of the detected vehicles.

A modified transfer learning approach has been applied to the DenseNet201 model for multiple classifications, including non-emergency cars, ambulances, police, and firefighters. The approach involves freezing the architecture of the model's layers. A high accuracy rate is obtained with this model and reaches 98.6%. Also, various optimization methods, including (Adam, Adamax, Nadam, and RMSprob) are used to improve the detection performance based on the best optimizer selection and yielded an accuracy of 98.84%. In addition, a modified version of YOLOv5 was proposed for vehicle detection, which aims to enhance the mean average precision (mAP) detection by 3%. Finally, the proposed system was simulated to reduce the waiting time at traffic signal. The experimental results demonstrate a significant reduction in waiting time, ranging from 30 to 100 seconds depending on the status.