A method of vehicle flow training and detection based on ResNet50 with CenterNet method

Nowadays, the increasing costs of law enforcement have become a considerable problem for the traffic department. With the insufficient police force, the department cannot separate the daily statistics of traffic flow at the intersection. In this case, understanding the daily traffic flow at the intersection helps to make the traffic light timing scheme, plan the path direction, understand the urban congestion, and so on. Therefore, we propose a vehicle detection model which can be deployed in NVIDIA Jetson AGX Xavier and is based on TensorFlow. To improve the accuracy and the speed of detection, we use CenterNet, a keypoint-base network model, and the backbone is the ResNet50. Hence, we can extract features of the image and use regression to predict the box of any object. The experimental results show that: the accuracy of vehicle flow identification is up to 98.23%, small vehicles 98.30%, medium vehicles 98.64%, and small vehicles 97.75% in a straight-ahead environment, meet the actual use requirements.