Selected Methods for Increasing the Accuracy of Vehicle Lights Detection

摘要

The paper presents selected methods for improving the accuracy of classification of headlights and taillights of the vehicles. The methods include analyzing blob properties and locations of the detections. A new feature for describing binary blob shape has been proposed. Moreover, data augmentation technique has been used to improve the results of the classification. The referenced system is based on convolutional neural networks (CNNs). New solutions have been tested with comprehensive set of video sequences (of total duration exceeding ten hours) under various weather conditions and from different road types.