Analysis and Synthesis of Traffic Scenes from Road Image Sequences

Sheng Yuan,Yuting Chen,Huihui Huo,Li Zhu

doi:10.3390/s20236939

Abstract

Traffic scene construction and simulation has been a hot topic in the community of intelligent transportation systems. In this paper, we propose a novel framework for the analysis and synthesis of traffic elements from road image sequences. The proposed framework is composed of three stages: traffic elements detection, road scene inpainting, and road scene reconstruction. First, a new bidirectional single shot multi-box detector (BiSSD) method is designed with a global context attention mechanism for traffic elements detection. After the detection of traffic elements, an unsupervised CycleGAN is applied to inpaint the occlusion regions with optical flow. The high-quality inpainting images are then obtained by the proposed image inpainting algorithm. Finally, a traffic scene simulation method is developed by integrating the foreground and background elements of traffic scenes. The extensive experiments and comparisons demonstrate the effectiveness of the proposed framework.

Highlights

Traffic scene simulation and modeling has been a hot topic in the community of intelligent transportation systems
A wide range of applications has been developed based on scene analysis and synthesis, including the evaluation of unmanned vehicle algorithms [1], traffic scenes construction [2,3,4], and the advanced driver assistant systems (ADAS) [5]
We propose a method based on single shot multi-box detector (SSD) feature extraction and improved bidirectional feature pyramid network (BiFPN) feature fusion method

Summary

Introduction

Traffic scene simulation and modeling has been a hot topic in the community of intelligent transportation systems. The analysis and synthesis of traffic scene is the foundation for traffic simulation and modeling. A wide range of applications has been developed based on scene analysis and synthesis, including the evaluation of unmanned vehicle algorithms [1], traffic scenes construction [2,3,4], and the advanced driver assistant systems (ADAS) [5]. There are basically two types for unmanned vehicle evaluation methods: field test and off-line test. As the traditional field test is unsafe and demands too much time cost, the off-line test of unmanned vehicles has become popular in recent years. Based on the analysis and synthesis of traffic scenes, the off-line test method is repeatable

Objectives

Methods

Findings

Conclusion