An End-to-End Traffic Visibility Regression Algorithm

Hongshuai Qin,Huibin Qin

doi:10.1109/access.2021.3101323

Abstract

Traffic visibility detection plays a vital role in intelligent transportation, autonomous driving, safe driving, etc. Convolutional neural networks (CNNs) based regression and classification algorithms have been shown competitive performance in many applications, but little attention has been paid to traffic visibility identification. In this paper, we propose a trainable end-to-end system called traffic visibility regression network (TVRNet). TVRNet takes a road image as input and outputs its visibility value. TVRNet adopts CNNs based deep architecture, uses appropriate filters to extract fog density-related features, and exploits the parallel convolution for multi-scale mapping. Later, a new type of non-linear activation function called Modified_sigmoid function is used. We synthesize labeled visibility datasets comprised of multi-scene and single-scene based on the actual road sense to train the visibility regression network. Extensive experiments and comparisons with other popular algorithms are performed to verify our method in road visibility estimation.

Highlights

In bad weather, the absorption or scattering of light by atmospheric particles such as rain, snow, fog, or haze can significantly reduce the visibility of scenes [1][2]
We propose traffic visibility regression network (TVRNet), a trainable end-to-end system that explicitly learns the mapping relations between road images and their associated visibility
Structure Design of TVRNet Inspired by the DehazeNet model [19] that can learn the mapping relationship between the original foggy image and its transmittance map, we propose an end-to-end system TVRNet, to learn the mapping relationship between the foggy image and its visibility

Summary

INTRODUCTION

The absorption or scattering of light by atmospheric particles such as rain, snow, fog, or haze can significantly reduce the visibility of scenes [1][2]. TVRNet uses appropriate filters to extract features related to fog density, achieves scale invariance through four parallel convolutions, further information integration through two fully connected layers. Layers and nonlinear activations of TVRNet are designed to implement four sequential operations for medium transmission estimation, namely, feature extraction, multi-scale mapping, information integration and nonlinear regression [19]. We present layer designs of TVRNet. 1) FEATURE EXTRACTION For the task of image visibility estimation, the previously proposed methods mainly relied on low-level image cues (e.g., image gradients, contrast, hue, saturation, dark channel, etc) without adequate scene consideration or understanding. 3) FEATURE INTEGRATION In order to obtain classification or regression results, CNN generally uses fully connected layers to integrate image feature maps extracted through multiple convolutional layers and pooling layers to obtain the high-level meaning of image features. The basic learning rate is 0.001, gamma: 0.1, momentum: 0.85, and the optimization algorithm uses the Nesterov algorithm

Output Layer Activation Function Comparison

Comparison of Estimation Results of Different Models

CONCLUSION