Real-time detection of vehicle queues in traffic with deep learning

Nowadays, due to the rapidly increasing number of vehicles and traffic density, traffic management and queueing detection have become an important issue. Especially in cities with high population density, the increase in traffic density has a negative impact on the daily life of people in traffic. In addition to adversely affecting daily life, traffic density causes problems such as prolonged journey times to and from destinations, increased fuel consumption of vehicles and consequently rapid depletion of energy resources. Therefore, traffic management and queuing detection play an important role in traffic management. In recent years, artificial intelligence supported systems have started to be proposed in order to solve the negative situations in traffic more easily. With these artificial intelligence-based systems, monitoring the traffic flow, determining the density in traffic and detecting and effectively managing queuing depending on this density helps to reduce the negative effects in traffic. In this thesis, it is aimed at performing traffic component analysis. The study consists of four stages. In the first stage, vehicle detection and classification are performed with YOLO, one of the deep learning algorithms, using visual traffic data sets. The model detects and classifies five different vehicle types such as car, motorbike, truck, bus and bicycle. The precision value achieved by the detection model in vehicle class prediction was calculated as 86%, the recall value as 87% and the mean average precision value as 93%. In the second stage, vehicle counting and tracking of the vehicles passing through the relevant region was performed with the SORT algorithm. In the third stage, the speeds of the counted vehicles were determined. In the last stage, queuing detection was performed using the detected speed information of the vehicles. In the queueing detection process, the average speed of the vehicles was used. The performances of the algorithms used for vehicle class detection, vehicle counting, vehicle speed detection and queueing detection were evaluated using both our own videos and videos obtained from internet sources. Keywords: Queuing, Vehicle detection in traffic, Deep learning, YOLO, SORT