[3F1-ES-2] Machine learning: Social application (3)

[3F1-ES-2-01] Performance Enhancement of Region-based Spatio-temporal Neural Network for Traffic Risk Estimation using Real and Virtual Datasets

〇Kwong Cheong Ng¹, Yuki Murata¹, Atsumi Masayasu¹ (1. Soka University)

Keywords:Traffic Risk Estimation, simulated image dataset, object detector YOLO

We propose a region-based spatio-temporal neural network for traffic risk estimation trained by the comparative loss function. In this model, a pretrained object detector YOLOv3 is fine-tuned using real image datasets: the KITTI and Dashcam Accident datasets and virtual image dataset: the VIPER dataset to detect moving object regions, and their features are clipped out from the middle layer of the detector. Then, these feature sequences are used to estimate the traffic risk by the spatio-temporal pattern encoding network followed by the risk estimation network. A comparative loss function is used to learn a risk estimation network by comparing pairs of two spatio-temporal patterns of moving objects. Experiments were conducted on a combination of real image datasets: the Dashcam Accident and AnAn Accident Detection (A3D) datasets and virtual image datasets: the VIENA2 dataset. The experimental results were analyzed and evaluated, and we have confirmed it is possible to estimate dangerous traffic situations using the proposed risk estimation network and the proposed network can be improved using both real and virtual image datasets for training.