Recognizing Traffic Black Spots From Street View Images Using Environment-Aware Image Processing and Neural Network

Abstract

This paper proposes a novel technique to identify black spots (prone-to-accident road locations) using street view images. The proposed technique is derived based on the hypothesis that the characteristics of the surroundings of the road have an effect on the safety level of a particular spot, and is the first black spot classification technique that is fully environment-aware. Assessing four street view images around each spot, a distance-aware pixel accumulation is developed to extract information about the objects surrounding the road from a semantically segmented image. The accumulated vectors are then used to train fully-connected neural networks to identify black spots. Performance evaluations are conducted with street view images in Thailand, which represent a challenging scenario of analyzing road characteristics in developing countries, with one of the highest road traffic fatality rates and limited historical accident records. Comparisons between our proposed technique and previously proposed techniques are also provided. Experiments show that our proposed technique succeeds in classifying black and safe spots in Thailand with an accuracy of 69.91%, where 75.86% of the black spots are identified correctly. Also, the distance-aware pixel accumulation can improve the accuracy of those machine learning techniques up to 6.4%. Our findings also evidently revealed that the object surrounding the roads as well as their sizes and distances are determinants of road's accident proneness.

Keywords

• Black spot detection
• machine learning
• feature extraction
• pixel accumulation
• street view images