Pre-Activated 3D CNN and Feature Pyramid Network for Traffic Accident Detection

Abstract

In this paper, we present a novel traffic accident detection method using a spatio-temporal three-dimensional (3D) convolutional neural network. The proposed method consists of pre-activation ResNet and feature pyramid network (FPN) structure. To reduce computational load with preserving the detection accuracy, we propose interval input method. Experimental results show that the proposed network outperforms existing methods in the sense of both precision and recall measures.