Method

Probabilistic Joint Boosting Classifier [st] [ProbBoost]


Submitted on 28 May. 2014 09:48 by
Giovani Bernardes Vitor (Heudiasyc Laboratory)

Running time:2.5 min
Environment:>8 cores @ 3.0 Ghz (C/C++)

Method Description:
This work proposes a method for road detection in
inner-city, using a set of probabilistic distribution to model the classifier of a Joint Boosting algorithm. Differently from others works, this approach creates a different set of features merging a technique called Diston, proposed by previous works (3D information) with the Texton (2D texture and color) to compute a set of probabilistic distribution for each superpixel. The probabilistic distribution feature’s descriptor is used to model the weak classifier used in the Joint Boosting algorithm.
Parameters:
lambda = 15
h = 3
Latex Bibtex:
@INPROCEEDINGS{Vitor2014ICRAWORK,
author={Vitor, G. B. and Victorino, A. C. and Ferreira, J. V.},
booktitle={ Workshop on Modelling, Estimation, Perception and Control of All Terrain Mobile Robots on IEEE International Conference on Robotics and Automation (ICRA)},
title={A probabilistic distribution approach for the classification of urban roads in complex environments},
year={2014},
keywords={Road Detection, Computer Vision, Joint Boosting, Texton Map, Dispton Map, Watershed Transform.},

}

Evaluation in Bird's Eye View


Benchmark MaxF AP PRE REC FPR FNR
UM_ROAD 87.48 % 80.13 % 85.02 % 90.09 % 7.23 % 9.91 %
UMM_ROAD 91.36 % 84.92 % 88.18 % 94.78 % 13.97 % 5.22 %
UU_ROAD 80.76 % 68.70 % 85.25 % 76.72 % 4.33 % 23.28 %
URBAN_ROAD 87.78 % 77.30 % 86.59 % 89.01 % 7.60 % 10.99 %
This table as LaTeX

Behavior Evaluation


Benchmark PRE-20 F1-20 HR-20 PRE-30 F1-30 HR-30 PRE-40 F1-40 HR-40
This table as LaTeX

Road/Lane Detection

The following plots show precision/recall curves for the bird's eye view evaluation.


Distance-dependent Behavior Evaluation

The following plots show the F1 score/Precision/Hitrate with respect to the longitudinal distance which has been used for evaluation.


Visualization of Results

The following images illustrate the performance of the method qualitatively on a couple of test images. We first show results in the perspective image, followed by evaluation in bird's eye view. Here, red denotes false negatives, blue areas correspond to false positives and green represents true positives.



This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png

This figure as: png


eXTReMe Tracker