
Ground Target Detection in LiDAR Point Clouds
using AdaBoost
Wenguang Zhang
ATR Laboratory
National University of
Defense Technology
Changsha, Hunan , China
zhangwenguang@nudt.edu.
cn
Yulan Guo
ATR Laboratory
National University of
Defense Technology
Changsha, Hunan , China
yulan.guo@ nudt.edu.cn
Min Lu
ATR Laboratory
National University of
Defense Technology
Changsha, Hunan , China
lumin@ nudt.edu.cn
Jun Zhang
ATR Laboratory
National University of
Defense Technology
Changsha, Hunan , China
zhj64068 @sina.com
Abstract
—
Although substantial progress has been made in
objects detecting in point clouds, the performance of most
methods is limited by the accuracy of segmentation. However,
accurate segmentation in complex scene is still an open problem.
This paper presents a novel rotation-invariant method for object
detection. It uses Adaboost to train a detector for exhaustively
scaning and testing of the point cloud scene. First, to address the
rotation-sensitive problem of 3D Harr-like features, we use
positive training samples obtained from multiple viewpoints of
the object. Then, false alarm is reduced using the prior
knowledge that the confidence of false alarm distributes sparsely
in the space. Experimental results demonstrate that the proposed
method achieves a high recall on point clouds obtained from
multiple viewpoints of the object at the low false alarm.
Keywords—object detection; AdBoost; point cloud;
Point clouds obtained by LiDAR represent the object shape
in 3D spatial space. Compared to 2D images, point clouds
provide more geometric information of the objects. Point
clouds are suitable for object detection and is not affected by
variations in illumination, viewpoints, environment temperat-
ure, and scale [1]. However, The research development in this
area is relatively slow due to the low revolution, high
manufacturing cost, and high power consumption of LiDAR
sensors. Recently, caused by the breakthrough development
and the cost reduce of LiDAR sensors, point cloud processing
becomes an active research area in computer vision. Object
detection in cluttered point clouds is an highly challenging task,
which is limited by intra-class shape variation, object
incompleteness caused by occlusion, overlapping between
neighboring objects, point-density variation, and orientation
variations.
Nowadays, a large number of approaches [2]-[10] have
been proposed to detect objects in point clouds, such as
buildings, doors, mailboxes, street lamps, and trees. However,
the performance of these methods is limited by the accuracy of
object segmentation. The segmentation errors caused by under-
segmentation and over-segmentation will decrease the object
detection performance. Although some point cloud
segmentation algorithms have been proposed in the literature
[11][12], accurate object segmentation in complex scenes is
still a highly challenging problem. Pang et al. [13] proposed a
training based 3D object recognition method. In that method, a
3D detection window is moved to search over the 3D image.
The similarity between each point cloud cluster within the
detection window and the target object trained by the detector
is calculated. Once the input point cloud has been exhausitively
checked by the scaning window, positive match instances are
detected and further processed by non- maximum suppression.
The target object is identified using the best matching and
confidence thresholding. This method dose not require any
segmentation, and achieves a high recall but low precision on
engineering LiDAR data.
Based on literature [13], we propose an improved object
detection method using AdaBoost. To address the rotation-
sensitive problem faced by 3D Harr-like features, we use
positive training samples obtained from multiple view points of
the object. Then, the false alarm is reduced using the prior
knowledge that the confidence of false alarm is sparsely
distributed in the space.
The main contributions of our method are twofold. First,
3D Harr-like features and AdaBoost training procedures are
combined for ground object detection. Second, an improved
rotation-invariant object detection method is proposed, which
has achieved a better performance compared to the method
proposed in [13].
The rest of this paper is organized as follows: Section II
describes the proposed method. Section III presents extensive
experimental results and analyses of the proposed method.
Finally, Section IV gives the concluding remarks.
II. THE P
ROPOSED
A
LGORITHM
A. AdaBoost Based Detector Training
Learning is a basic skill for human to perceive the world,
the task of machine learning is to imitate human’s learning
capability using machines. In this paper, we propose a 3D
object detection method based on AdaBoost. An overall flow
diagram of our algorithm is shown in Fig.1.
Our algorithm is divided into two modules: offline training
and online detection. We first train a detector for each object
offline using the Adaboost training procedure with training
samples generated from a labeled object library. Then, the
detector exhaustively scans and tests the point cloud, resulting