VehicleDetectionandTrackinginCarVideoBasedonMotionModel资源-CSDN文库

5星 · 超过95%的资源需积分: 10 45 浏览量 2011-10-15 09:30:24 上传评论收藏 1.56MB PDF 举报

Vehicle Detection and Tracking in Car Video Based on Motion Model--This work aims at real-time in-car video analysis to detect and track vehicles ahead for safety, auto-driving, and target tracing. This paper describes a comprehensive approach to localize target vehicles in video under various environmental conditions. The extracted geometry features from the video are projected onto a 1D profile continuously and are tracked constantly. We rely on temporal information of features and their motion behaviors for vehicle identification, which compensates for the complexity in recognizing vehicle shapes, colors, and types. We model the motion in the field of view probabilistically according to the scene characteristic and vehicle motion model. The Hidden Markov Model is used for separating target vehicles from background, and tracking them probabilistically. We have investigated videos of day and night on different types of roads, showing that our approach is robust and effective in dealing with changes in environment and illumination, and that real time processing becomes possible for vehicle borne cameras. ### 基于运动模型的车辆检测与跟踪技术 #### 引言随着自动驾驶技术和智能交通系统的快速发展，车辆检测与跟踪技术成为了实现安全驾驶、事故避免及自动追踪目标的重要环节。本文介绍了一种针对车内视频实时分析的目标车辆检测与跟踪方法。该方法能够在不同的环境条件下有效地识别前方车辆，并对其进行持续跟踪。 #### 技术背景与问题定义在驾驶过程中，准确感知前方车辆以及交通状况对于保障行车安全、避免事故和实现自动化驾驶至关重要。虽然现有的物体识别与跟踪技术已经取得了显著进展，但在实际应用中，尤其是在车内视频领域，仍然面临着诸多挑战。这些挑战包括光照条件的变化、复杂多变的道路环境等。因此，开发一种能够适应这些变化并在实时条件下工作的车辆检测与跟踪系统变得尤为重要。 #### 方法论本研究提出了一种全面的方法来解决上述问题，其核心思想是利用车辆的运动特征来进行车辆识别与跟踪。具体而言，该方法包括以下几个关键步骤： 1. **几何特征提取**：从视频帧中提取出车辆的几何特征，如轮廓、形状等。 2. **一维轮廓投影**：将这些几何特征投影到一维轮廓上，便于后续处理。 3. **基于时间序列的信息分析**：利用车辆随时间变化的运动行为来增强识别准确性。 4. **概率模型构建**：通过建立概率模型来描述车辆在视野中的运动特性，进而提高跟踪精度。 5. **隐藏马尔可夫模型（HMM）应用**：采用HMM来区分目标车辆与背景，实现对车辆的动态跟踪。 #### 实验验证为了验证所提出方法的有效性与鲁棒性，研究人员收集了多种类型的视频数据，包括白天和夜晚不同道路条件下的视频资料。实验结果显示，该方法不仅能够在复杂的环境下准确地检测并跟踪目标车辆，而且还能有效地应对光照变化等挑战，确保了在实际应用场景中的可行性。 #### 关键技术点详解 - **几何特征提取与一维轮廓投影**：通过对视频帧进行处理，提取出目标车辆的轮廓特征，并将其转换为一维轮廓表示形式。这种方法简化了后续的数据处理流程，同时保持了足够的信息用于识别目的。 - **基于时间序列的信息分析**：通过对车辆运动的时间序列数据进行分析，可以捕捉到目标车辆随时间变化的动态行为模式。这种基于时间的信息能够补偿仅依靠形状、颜色或类型识别所带来的复杂性和不确定性。 - **概率模型与场景特性建模**：通过建立概率模型来描述车辆在特定场景下的运动特性，这种方法能够有效地区分不同类型的车辆并减少误报率。 - **隐藏马尔可夫模型的应用**：HMM是一种强大的统计模型，适用于处理具有时间序列特性的数据。在本研究中，它被用来对目标车辆进行分类，并在复杂的背景下进行精准跟踪。 #### 结论本研究提出的基于运动模型的车辆检测与跟踪方法，在多种不同的环境条件下均表现出了良好的性能。通过对车辆的几何特征进行处理，并结合概率模型与时间序列分析，该方法不仅能够有效地识别前方车辆，还能实现实时跟踪，为智能驾驶系统的发展提供了有力支持。未来的研究方向将进一步优化算法性能，并探索更多应用场景下的应用可能性。

资源推荐

资源详情

资源评论

> FOR CONFERENCE-RELATED PAPERS, REPLACE THIS LINE WITH YOUR SESSION NUMBER, E.G., AB-02 (DOUBLE-CLICK HERE) <

Vehicle Detection and Tracking in Car Video Based on Motion Model

A. Jazayeri, H. Cai, J. Y. Zheng, Senior Member, IEEE, and M. Tuceryan, Senior Member, IEEE

Indiana University Purdue University Indianapolis (IUPUI), IN 46202 USA

This work aims at real-time in-car video analysis to detect and track vehicles ahead for safety, auto-driving, and target tracing. This

paper describes a comprehensive approach to localize target vehicles in video under various environmental conditions. The extracted

geometry features from the video are projected onto a 1D profile continuously and are tracked constantly. We rely on temporal

information of features and their motion behaviors for vehicle identification, which compensates for the complexity in recognizing

vehicle shapes, colors, and types. We model the motion in the field of view probabilistically according to the scene characteristic and

vehicle motion model. The Hidden Markov Model is used for separating target vehicles from background, and tracking them

probabilistically. We have investigated videos of day and night on different types of roads, showing that our approach is robust and

effective in dealing with changes in environment and illumination, and that real time processing becomes possible for vehicle borne

cameras.

Index Terms— In-car Video, feature detection, 1D profiling, tracking, vehicle motion, dynamic target identification, probability,

HMM.

I. I

NTRODUCTION

ENSING VEHICLES ahead and traffic situations during

driving are important aspects in safe driving, accident

avoidance, and automatic driving and pursuit. We designed a

system that is capable of identifying vehicles ahead moving in

the same direction as our car by tracking them continuously

with an in-car video camera. The fundamental problem here is

to identify vehicles in changing environment and illumination.

Although there have been numerous publications on general

object recognition and tracking, or combination of them, not

many of these techniques could be successfully applied in

real-time for the in-car video, which has to process the input

on-the-fly during the vehicle movement. This paper introduces

an effort to design and implement such real-time oriented

algorithms and systems highly adaptive to the road and traffic

scenes based on domain-specific knowledge on road, vehicle,

and control.

In-car video is from a camera facing forward (Fig. 1), which

is the simplest and most widely deployed system on police

cars. It records various traffic and road situations ahead and is

particularly important to safe driving, traffic recording, and

vehicle pursuit. Our objective is to detect vehicles ahead or

those being pursued, and track them continuously in the video.

It is not easy for a single moving camera to yield the

information quickly from dynamic scenes without stereo or

other sensors’ assistance [1]. The main difficulties are, first,

the numerous variations of vehicles in color, shape, and type.

The vast amount of vehicle samples is difficult to model or

learn [2]. Second, the vehicle detection must be done

automatically along with the video tracking. Many tracking

algorithm only assume easily detectable targets or known

initial positions [3]. Third, the in-car video may confront

drastic variation of environment and illumination [4].

Transition through shadow and sunny locations in urban areas,

dim lighting at night, loss of color on a cloudy day, highlights

on vehicles, occlusions between vehicles, and so on make the

feature extraction, recognition, and tracking unstable.

Our novel method first selects and detects the most common

low-level features on vehicles that are robust to changes of

illumination, shape, and occlusion. This avoids high-level

vehicle and scene modeling and learning that are mostly

unstable, time consuming, and non-general. Second, we focus

on the horizontal scene movement for fast processing based on

the configuration of camera and vehicle driving mechanism.

One dimensional (1D) profiles are accumulated from video

frames to show the horizontal motion directly in traces. We

track feature trajectories in such temporal profiles so that the

real time vehicle and background classification can be done in

a reduced dimension. Third, we put more weight on the

understanding of motion behavior of the vehicles than object

shape analysis in identifying targets. We use the Hidden

Markov Model (HMM), which has been widely used in

Manuscript received Dec. 1, 2009. Corresponding author: Jiang Yu

Zheng (e-mail: jzheng@cs.iupui.edu).

Digital Object Identifier inserted by IEEE

Fig. 1.A typical frame of in-car video and the processing diagram. The video

resolution is 640×480pixels

Detect features in video frame

Spatial profiling of video frame

Update and Track motion traces

Ego-motion

Likelihood PDFs

of background

and vehicle

motion

HMM

Trace ID with

robabilit

(

)

Street and vehicle

motion model

P(t-1)

> FOR CONFERENCE-RELATED PAPERS, REPLACE THIS LINE WITH YOUR SESSION NUMBER, E.G., AB-02 (DOUBLE-CLICK HERE) <

describing sequential data [5, 6], to model the continuous

movement of the features so that the results will be less

influenced by thresholding during low level preprocessing.

Identification of target vehicles will be given in terms of

probabilities.

Many related works of in-car video or images to identify

cars are shape based methods [7, 8] that usually suffer from

brittleness due to variety of vehicles and backgrounds. For

example, the symmetric property is employed in [7, 9, 10, 11].

Deformable frame models are used in [12,13,14,27]. Because

of the lack of robustness in feature extraction from complex

scenes, subsequent decisions can rarely be applied to video in

real time. Several works have employed probability in their

formulation but they are mainly applied to the occurrence and

appearance (intensity distribution) of vehicles [2,11,15,16,17,

26] or some of them are from static camera [18, 19].

The significance of this work lies in its temporal processing

in target identification. We focus on the time domain that

displays the vehicle generated motion. The motion is general

for all types of vehicles. The motion behavior of scenes in

videos is modeled with the HMM and this makes the vehicle

identification with certainty measure. Also, the profiled 1D

arrays reduce dimensionality of data, thus achieving real-time

processing to identify and track vehicles simultaneously.

In the following, we will give an overview and describe

general assumptions in Sec. II. We introduce feature detection

and tracking in Sec. III. We investigate the motion properties

of scenes in terms of likelihoods in Sec. IV. We address the

identification of feature trajectory as a vehicle or background

in Sec. V. Experimental results and discussion are given in

Sec. VI.

II. M

OTION INFORMATION IN CAR VIDEO

A. Overview of the Work

By showing the continuous motion of extracted points in car

video without color and shape information to human subjects,

we have confirmed that humans are capable of separating

vehicles from background after knowing where the source is

from. As the observer vehicle moves on the road, the relative

background motion is determined. The motion projected to the

camera frame is further determined from object distances.

Such motion shows unique properties coherent to the vehicle

ego-motion. On the other hand, the target vehicles moving in

the same direction as the camera have different motion against

the background, and thus show different optical flow from that

of the background in the video. We look into the motion

characteristics of tracked objects and separate them as static

background or moving cars in the process shown in Fig. 1.

Different from many other works that put more effort into

vehicle shape analysis in individual video frames, this work

only extract low level features such as corners, intensity peaks,

and horizontal line segments, as will be described in Sec. III.

These features are profiled to the temporal domain to ensure

the compactness of data and the robustness of processing

against the camera/vehicle shaking, noises, and irrelevant

features. The motion of scenes appears as traces in the

temporal profiles. The horizontal image velocities of features

are obtained through trace tracking.

To identify tracked traces as cars and background, we model

the motion properties of scenes in the video in Sec. IV.

Because the background flow is ego-motion orientated and

position dependent in the image frame, we examine the image

flow with respect to the position during ego-motion and

describe them by their joint probability. We estimate

likelihood probability distributions for cars and background

respectively, given the PDFs of scenes, ego-motion, and the

pursuit vehicles. With such initially calculated likelihood

PDFs, we can obtain high responses to the expected motion

events.

In Sec. V, we estimate the posterior probabilities of motion

traces as target vehicles or background based on observations,

i.e., the detected image position and velocity at each moment.

To achieve a stable classification of vehicles, we evaluate their

motion process by means of the HMM when traces are tracked

continuously. The pre-calculated likelihood PDFs and the

transition probabilities are used in the HMM for the state

estimation of traces. The continuous vehicle motion provides

more robust identification than shape features that are affected

by occlusion, illumination and background changes.

B. Dynamic Environment of In-Car Video

General assumptions we make are (a) the observer car

should not turn away completely or leave too far from target

vehicles on the road; (b) a translation of the observer vehicle is

required. For the vehicle-borne camera, a pure rotation

without significant translation, e.g., turning at a street corner,

does not provide sufficient information to separate the targets

and background motion, because such a rotation generates

almost the same optical flow in entire frame. Another

reasonable assumption is the continuity of the vehicle and

camera motion, which is guaranteed in general according to

the driving mechanism of four-wheeled vehicles.

(a)

(b)

Fig. 2 Typical views of in-car video with red arrows indicating optical flow or

image velocity. (a) Straight road, (b) mildly curved road.

Assume the camera centered coordinate system is O–XYZ

(or O-xyz) with the X axis toward right, Y axis facing down

and Z axis in the vehicle’s direction of motion. Denote the 3D

Vehicle

FOE

x=0

y=0

Horizon

x=imageSize/2

Vehicle

剩余10页未读，继续阅读

评论收藏

内容反馈

zzzzzzzggg

2015-10-19

似乎很有含金量的一篇论文。
zjr19850103

2014-08-01

目前正学习车辆检测方面，文档描述得比较深刻，检测和跟踪采用的算法值得借鉴
nk3652

2015-07-14

非常好的论文, 算法的描述,实验结果的数据都很详细
liangbingchun

2014-11-04

正在研究车辆检测，下载下来阅读一下
kaoshare

2016-03-13

正在研究，比较详细的文章

前往

页

O天涯海阁O

粉丝: 1553
资源: 90

Vehicle Detection and Tracking in Car Video Based on Motion Mode...

最新资源

Vehicle Detection and Tracking in Car Video Based on Motion Mode...

Tracking Based Structure and Motion Recovery for Augmented Video Productions

Deformable Part Model Based Multiple Pedestrian Detection for Video Surveillance in Crowded Scenes

Multi-Model Estimation Based Moving Object Detection for Aerial Video

Center-based 3D Object Detection and Tracking译文

vehicle detection and tracking_vehicledetection_trackingvehicle_

model based vehicle detection and tracking

Real-time multiple vehicle detection and tracking from a moving vehicle.pdf

Vehicle detection and tracking.zip_可见光_激光雷达_融合识别 激光_车辆识别_雷达融合

A Contour-Based Moving Object Detection and Tracking

face detection and tracking with camshift

Obstacle Detection and Tracking for the Urban Challenge 障碍物检测跟踪

Driver Fatigue Detection Based on Eye Tracking

论文研究-RGB-D Object Tracking and Occlusion Deformation Processing Based on Depth Model.pdf

detection and tracking of point features

FACE-DETECTION-AND-TRACKING-USING-A-BOOSTED-ADAPT_faces

Target detection and tracking

Improving Context Modeling for Video Object Detection and Tracking

Robust Object Tracking via Sparsity-based Collaborative Model

Moving Object Detection and Tracking in Videos

OBJECT DETECTION AND RECOGNITION IN DIGITAL IMAGES

Multi-Sensor Fusion System for Moving Object Detection&Tracking

Based.Radar.Detection.Tracking.And.Classification

code-Robust Object Tracking via Sparsity-based Collaborative Model

论文研究-A confidential target tracking algorithm in video sequences based on CS theory.pdf

A survey of vision-based vehicle detection and tracking techniques in ITS

最新资源

Vehicle detection and tracking.zip_可见光_激光雷达_融合识别激光_车辆识别_雷达融合