Traffic State Detection

Object Detection

Dataset

• It consists of 3,849 photos collected from various sources, from the Internet or from videos taken on highways and roads in Hanoi.
• Various data enhancement techniques from Roboflow API and available augmentation methods of YOLO were used.
• It includes 4 types of vehicles: car, bike, bus, truck. With the number of instances for each class as follows: 9,087 cars, 6,278 bikes, 1,138 buses, 3,976 trucks.
• The dataset is divided according to the ratio: 70/20/10.
• All images are processed in the standard size of 640x640 before being used for training.

Configuration

• Three models were trained: YOLOv5n, YOLOv7, and YOLOv8n, using pre-trained models available in open-source, with a training of 100 epochs on the same dataset.
• They were trained on Google Colab with a GPU configuration of Tesla K80 and 12GB of RAM.
• They were tested on a MacBook Pro 2015 with an Intel Core i5 Dual Core CPU and 16GB of RAM.

Result

Model	params (M)	FLOPs (B)	Size	Speed (ms)	mAP@50	mAP@50-95
YOLOv5n	1.9	4.5	640	104.6	96.2%	67.5%
YOLOv8n	3.2	8.7	640	178.3	96.8%	70.9%

Congestion Classification

Algorithm

• In this problem, traffic congestion classification is determined through the calculation of the ratio x between the total area of bounding boxes surrounding the detected vehicles (V) and the area of the region of interest (R), assuming that each pixel is a unit of area.
• The region of interest is the area where the detection and counting of vehicles will be performed. It is drawn immediately after running the program.
• Below is the area of the region of interest, R.

- Below is the total area of the bounding boxes, V.

The ratio V/R is then mapped to a balanced scale and passed through a sigmoid function to distinguish more clearly between two states: congestion and normal.

Result

The result indicates a normal traffic state.

The result indicates a traffic congestion state.