Pedestrian Detection Combining RGB and Dense LIDAR Data [la] [Fusion-DPM]

Submitted on 30 May. 2014 17:33 by
Cristiano Premebida (ISR-UC, University of Coimbra)

Running time:~ 30 s
Environment:1 core @ 3.5 Ghz (Matlab + C/C++)

Method Description:
DPM on Velodyne-dense map* with NMS (thr=0.4) + DPM
on RGB-images with NMS + SVM-based
rescoring strategy + final NMS.

(*) Depth map are obtained using an upsampling
method on LIDAR data only.
NMS: Non-Maximum Suppression.
DPM: Deformable-Part Based Models (release5).
DPM(release5): 5 components, extra_octave=true, NMS
with threshold=0.4. LibSVM with RBF-kernel (for
detection windows re-scoring); features are
Latex Bibtex:
author = {C. Premebida and J. Carreira and J.
Batista and U. Nunes},
title = {Pedestrian Detection Combining RGB and
Dense LIDAR Data},
booktitle = IROS,
year = {2014}

Detailed Results

Object detection and orientation estimation results. Results for object detection are given in terms of average precision (AP) and results for joint object detection and orientation estimation are provided in terms of average orientation similarity (AOS).

Benchmark Easy Moderate Hard
Pedestrian (Detection) 58.93 % 44.99 % 40.19 %
This table as LaTeX

2D object detection results.
This figure as: png eps pdf txt gnuplot

eXTReMe Tracker