Real-time vehicle detection and counting based on YOLO and DeepSORT
Thanh-Nghi Doan, Truong Minh-Tuyen
Abstract
Intelligent vehicle detection and counting are becoming increasingly important in the field of highway and transport infrastructure management. Traditional methods based on image information have shown several limitations. Especially in real-world environment conditions, real-time detection, classification and counting each type of vehicle are still a big challenge. The main purpose of this study is to develop an adaptive model that combine YOLOv4 and DeepSORT. The new model can detect object with high accuracy and fast calculation time by taking the benefits of tracking with a focus on simple, effective algorithms. Experiment results have shown that our proposed approach outperforms the original one at least 11% of AP and 12% of AP50 for most field scenarios of our dataset at a real-time speed of ~32 FPS.