Usage of fully convolutional network with clustering for traffic light detection

In this paper we consider a traffic light detector constructed on the basis of a fully convolutional neural network for segmenting traffic lights on image and subsequent clustering, which allows us to obtain bounding boxes for traffic lights. The proposed approach is compared with one of the most ef...

Full description

Saved in:
Bibliographic Details
Published in2018 7th Mediterranean Conference on Embedded Computing (MECO) pp. 1 - 6
Main Authors Yudin, Dmitry, Slavioglo, Dmitry
Format Conference Proceeding
LanguageEnglish
Published IEEE 01.06.2018
Subjects
clustering
Clustering algorithms
Convolutional neural networks
deep learning
detection
Detectors
embedded system
fully convolutional network
Image recognition
single shot multibox detector
Task analysis
traffic light
Training
Online AccessGet full text

Cover

More Information
Summary:In this paper we consider a traffic light detector constructed on the basis of a fully convolutional neural network for segmenting traffic lights on image and subsequent clustering, which allows us to obtain bounding boxes for traffic lights. The proposed approach is compared with one of the most effective object detectors - Single Shot Multibox detector (SSD). We implemented algorithms for objects detection on an embedded system based on the NVidia Jetson TX2 platform. Traffic light detection recall for the proposed approach is better than SSD and higher than 0.9 on both the testing and training samples from relatively small data set (500 training images and 107 testing images). Time of traffic lights detection on one frame is about 50 ms. The results of the traffic light detection prove the possibility of applying the approach based on fully convolutional neural network with clustering for embedded autonomous vehicle control systems and driver assistance systems.
DOI:10.1109/MECO.2018.8406049