From Speed Profile Data to Analysis of Behaviour
Författare
Summary, in English
Classification of speed profiles is necessary to allow interpretation of automatic speed measurements in terms of road user behaviour. Aggregation without considering variation in individual profile shapes easily leads to aggregation bias, while classification based on exogenous criteria runs the risk of loosing important information on behavioural (co-) variation. In this paper we test how three pattern recognition techniques (cluster analysis, supervised learning and dimension reduction) can be applied to automatically classify the shapes of speed profiles of individual vehicles into interpretable types, with a minimum of a priori assumptions. The data
for the tests is obtained from an automated video analysis system and the results of automated classification are compared to the classification by a human observer done from the video. Normalisation of the speed profiles to a constant number of data points with the same spatial reference allows them to be treated as multidimensional vectors. The k-means clustering algorithm groups the vectors (profiles) based on their proximity in multidimensional space. The results are satisfactory, but still the least successful among the tested techniques. Supervised learning (nearest neighbour algorithm tested) uses a training dataset produced beforehand to assign a profile to a specific group. Manual selection of the profiles for the training dataset allows better control of the output results and the
classification results are the most successful in the tests. Dimension reduction techniques decrease the amount of data representing each profile by extracting the most typical “features”, which allows for better data visualisation and simplifies the classification procedures afterwards. The singular value decomposition (SVD) used in the test performs quite satisfactorily. The general conclusion is that pattern recognition techniques perform well in automated classification of speed profiles compared to classification by a human observer. However, there are no given rules on which technique will perform best.
for the tests is obtained from an automated video analysis system and the results of automated classification are compared to the classification by a human observer done from the video. Normalisation of the speed profiles to a constant number of data points with the same spatial reference allows them to be treated as multidimensional vectors. The k-means clustering algorithm groups the vectors (profiles) based on their proximity in multidimensional space. The results are satisfactory, but still the least successful among the tested techniques. Supervised learning (nearest neighbour algorithm tested) uses a training dataset produced beforehand to assign a profile to a specific group. Manual selection of the profiles for the training dataset allows better control of the output results and the
classification results are the most successful in the tests. Dimension reduction techniques decrease the amount of data representing each profile by extracting the most typical “features”, which allows for better data visualisation and simplifies the classification procedures afterwards. The singular value decomposition (SVD) used in the test performs quite satisfactorily. The general conclusion is that pattern recognition techniques perform well in automated classification of speed profiles compared to classification by a human observer. However, there are no given rules on which technique will perform best.
Avdelning/ar
- Trafik och väg
- Matematikcentrum
- Mathematical Imaging Group
Publiceringsår
2009
Språk
Engelska
Sidor
88-98
Publikation/Tidskrift/Serie
IATSS Research
Volym
33
Issue
2
Länkar
Dokumenttyp
Artikel i tidskrift
Förlag
Elsevier
Ämne
- Infrastructure Engineering
- Mathematics
Nyckelord
- Behaviour analysis
- Pattern recognition
- Speed profile
- Clustering
- Supervised learning
- Dimension reduction
Status
Published
Forskningsgrupp
- Mathematical Imaging Group
ISBN/ISSN/Övrigt
- ISSN: 0386-1112