Intelligent Transport Systems – From Research and Development to the Market Uptake. First International Conference, INTSYS 2017, Hyvinkää, Finland, November 29-30, 2017, Proceedings

Research Article

Vehicles Recognition Based on Point Cloud Representation

Download
156 downloads
  • @INPROCEEDINGS{10.1007/978-3-319-93710-6_9,
        author={Patrik Kamencay and Robert Hudec and Richard Orjesek and Peter Sykora},
        title={Vehicles Recognition Based on Point Cloud Representation},
        proceedings={Intelligent Transport Systems -- From Research and Development to the Market Uptake. First International Conference, INTSYS 2017, Hyvink\aa{}\aa{},  Finland,  November 29-30, 2017, Proceedings},
        proceedings_a={INTSYS},
        year={2018},
        month={7},
        keywords={Point cloud Deep learning Convolutional Neural Network SVM Stereo system 3D model Vehicle CNN SSCD},
        doi={10.1007/978-3-319-93710-6_9}
    }
    
  • Patrik Kamencay
    Robert Hudec
    Richard Orjesek
    Peter Sykora
    Year: 2018
    Vehicles Recognition Based on Point Cloud Representation
    INTSYS
    Springer
    DOI: 10.1007/978-3-319-93710-6_9
Patrik Kamencay1,*, Robert Hudec1,*, Richard Orjesek1,*, Peter Sykora1,*
  • 1: University of Zilina
*Contact email: patrik.kamencay@fel.uniza.sk, robert.hudec@fel.uniza.sk, richard.orjesek@fel.uniza.sk, peter.sykora@fel.uniza.sk

Abstract

The following article is dedicated to techniques for recognition of vehicles on the road. By using 3D virtual models of vehicles, it is possible to create database of point cloud. The SSCD algorithm for training and testing was used. First for each 3D model the point clouds were created. Then from each point cloud one hundred pictures were rendered from different projections. Creation of filtered dataset was done by selection six angles from these projections. This dataset contains 100 models of vehicles divided into 5 classes. In summary, final non-filtered dataset contains 10 000 pictures, filtered dataset consist of 600 pictures. Dataset was used in support vector machine (SVM) and convolutional neural network (CNN) for training and testing in ratio 80:20. The result for SVM was 40%, this was done because non-filtered dataset contains many similar projections. Moreover, the size resulted in long duration of experiment (<90 h). Therefore, other experiments were done with filtered dataset. In filtered dataset, best result in SVM was 79% with RBF kernel. For the next experiment, CNN was used. With data augmentation the result was 80%, without 89%.