3. Introduction
Input video frame :640X480 resolution
Standard digital video camera(30-fps video)
Feature extraction :color ,texture
Classification method :k-NN and ANN
Four class problem , accuracy : ~80%
{off−road, urban, major/trunk road,multilane motorway/carriageway}
Two class problem , accuracy : ~90%
{off−road, on−road}
A near real time classification rate of 1Hz. (i.e.,one frame
classification per second)
3
4. Introduction
Colour based off-road environment and terrain type
classification[3]
Texture and neural network for road segmentation[7]
Detection and classification of highway lanes using
vehicle motion trajectories[31]
4
7. Feature extraction and representation
Color Features
Each such channel of each subregion as the normalized histogram
distribution.
Mean , standard deviation , and entropy.
A given color value (indexed k = 1, . . . , L) occurs with
probability pk
Each color channel is summarized as a color feature vector of
combining the histogram (quantized to 10 “bins”)
13-value , 7 channel , 91-D color descriptor for each image frame.
7
8. Feature extraction and representation
Texture Features
grey-level co-occurrence matrix (GLCM) statistics
localized orientation is defined with as {N,S, E,W,NW,NE,SW,SE}
Original image Co-occurrence Matrix
8
9. Feature extraction and representation
Texture Features
grey-level co-occurrence matrix (GLCM) statistics
Co-occurrence Matrix Stochastic Matrix
9
10. Feature extraction and representation
Texture Features
grey-level co-occurrence matrix (GLCM) statistics
10
11. Feature extraction and representation
Texture Features
grey-level co-occurrence matrix (GLCM) statistics
M(i, j) is the (i, j)th entry in GLCM M with dimension(colsx
rows)
11
12. Feature extraction and representation
Texture Features
grey-level co-occurrence matrix (GLCM) statistics
horizontal standard deviation and mean (σI, μI ),
vertical standard deviation and mean (σJ, μJ ).
12
13. Feature extraction and representation
Texture Features
Gabor filters allow the study of the localized spatial distribution of
the texture.
The magnitude of the Gabor filter response identifies varying local
texture frequencies and orientations in the image.
Use for the extraction of more gradual (low-frequency) textures
and more generally create discriminative texture descriptors.
13
14. Feature extraction and representation
Texture Features
The use of only the (N,E) GLCM directions is within this visual
discriminatory context.
The Gabor filter in use is itself summarized as a quantized
histogram
(10 “bins”)
Mean, standard deviation, and entropy
23-value ,3 subregions , 69-D texture feature descriptor for each
image frame.
14
15. Feature extraction and representation
Edge-Derived Features
Three additional edge-derived features specific to the road-
edge subregion.
Use Canny edge detector
The set of edges, connected contours , and straight line detected
is then summarized by the entropy.
15
16. Feature Classification
A 163-D combined feature vector per image frame.
K-NN and ANN
Training set : 800 image frames (200 per class ,for four
classes)
Four-class
classes = {off−road, urban, major/trunk road,
multilane motorway/carriageway}
Two-class
classes = {off−road, on−road}
{on−road} = {{urban} ∪ {major/trunk road} ∪
{multilane motorway/carriageway}}
16
17. Experimental result
Two test video sequences
Video sequence Duration Consist of
Video sequence 1 40-s 10-s segments of
{off-road, urban, major/trunk
road, multilane
motorway/carriageway}
Video sequence 2 50-s 10-s segments of
{urban, major/trunk road,
multilane
motorway/carriageway}
Test image frame 20-s segments of {off-road}
Road environment
600(20-s) urban
600(20-s) major/trunk road
30-fps video
600(20-s) multilane
motorway/carriageway
900(30-s) off-road
17
18. Experimental result
K-NN classification
Fig. 2. Video sequence 1. Classification results using k-NN varying
parameter k.
18
19. Experimental result
K-NN classification
Fig. 3. Video sequence 2. Classification results using k-NN varying
parameter k.
19
20. Experimental result
ANN classification
We employ a classical two-layer network topology with H hidden
nodes.
163 input node , 2 or 4 output node (two class or four class).
The ANN is trained using I iterations.
The general range of parameter H ={10, . . . , 60} and I = {150, . . . ,
700}.
20
22. Experimental result
ANN classification
22 Fig. 4. Examples of successful ANN classification for four road
environments
23. Experimental result
ANN classification
23 Fig. 5. Examples of successful ANN classification for two road
environments (ANN configuration: H = 15 Nodes; I = 200).
24. Experimental result
Extended Sequence Results
full video sequences (representing the complete set of viable data
gathered over several hours in varying environments)
24
26. Conclusion
An ANN classifier gives ∼90%–97% successful
classification for two class ; ∼80%–85% for four-class.
A k-NN classifier implies the inherent feature overlap
within the current feature space ,so it’s difficult to classify.
Future work
Subregion optimization
Alternative computationally efficient texture measures
The efficient of varying weather and lighting condition on
performance
26
