移动车辆上的多线索行人检测与跟踪系统

需积分: 9 25 浏览量更新于2024-07-27 收藏 694KB PDF 举报

"行人检测与跟踪：Multi-cue Pedestrian Detection and Tracking from a Moving Vehicle" 这篇论文主要探讨了如何从移动车辆上实时地检测和跟踪行人，作者是D.M. Gavril和S. Munder，发表在2007年的《国际计算机视觉杂志》（International Journal of Computer Vision）第73期第1卷，第41-59页。该研究考虑了多种视觉线索的结合，以提高检测和跟踪的准确性和效率。行人检测是自动驾驶和智能交通系统中的关键问题，而从移动车辆进行行人检测和跟踪更具挑战性，因为车辆运动带来的视角变化、遮挡以及环境条件的影响。论文提出了一种多线索的行人检测系统，该系统由一系列模块构成，每个模块利用互补的视觉标准逐步缩小图像搜索空间，同时兼顾鲁棒性和效率。论文的一个创新点是连续模块之间的紧密集成。这些模块包括稀疏立体匹配、边缘分析、形状和纹理特征等。稀疏立体匹配用于估计深度信息，帮助区分前景和背景，减少误报。边缘检测则有助于识别行人轮廓，而形状和纹理特征则提供了行人特有的识别标志。这种多层次的方法可以更有效地过滤非行人对象，提高检测精度。在跟踪方面，论文可能涉及到了基于卡尔曼滤波器或其他在线学习算法的行人跟踪方法。这些算法能够处理目标间的短期消失和重新出现，以及新行人的插入，保持对目标的持续追踪。此外，由于车辆在移动，可能还讨论了运动预测和校正策略，以适应车辆速度和方向的变化。论文还可能包含了实验部分，对提出的行人检测和跟踪系统进行了实际场景的验证，并与其他方法进行了比较。实验结果通常会展示系统的性能指标，如检测率、假阳性率和跟踪精度，以证明其在复杂环境下的有效性。这篇论文对行人检测和跟踪领域做出了重要贡献，通过结合多种视觉线索和优化算法，提高了在移动平台上的行人识别和跟踪能力，对于智能交通系统和自动驾驶汽车的安全性有着重要的实用价值。

44 Gavrila and Munder

Table 1. Overview of current pedestrian detection systems. Detection performance derived by visual inspection of ROC graphs in respective

publications (approximation). “f.pos.” denotes false positives.

Sensor Coverage Detection Performance Tracking Processing

Authors type area (per frame) (per trajectory) speed Test set

Papageorgiou mono, – 70% correct, 0.15 f.pos. no 20 min 123 ped. images,

and Poggio visible (full set color wavelets) (200 MHz PC) 50 non-ped.

(2000) 70% correct, 3 f.pos. 10 Hz images

(red. grey wavelet set)

Mohan et al. mono, – 85% correct, 0.03 f.pos no – 123 ped. images

(2001) visible 50 non-ped

images

Viola et al. mono, – 80% correct, 0.5 f.pos no 4 Hz 2 sequences

(2005) visible (2.8 GHz PC) of 2000 images

(stat. camera)

Elzein et al. mono, – 69% correct, no 95s 16 ped. images +

(2003) visible 61% precision (500 MHz PC) few sequences

Shashua mono, 3–25 m 96% correct, 5.6 ×10

−6

f.pos yes 10 Hz 1–5 hrs

et al. visible (inward moving) urban drive

(2004) 93% correct, 8.3 ×10

−5

f.pos

(stationary in-path)

85% correct, 2.9 ×10

−3

f.pos

(stationary out-path)

Zhao and stereo, – 85% correct, 3% f.pos no 3–12 Hz –

Thorpe visible (per stereo ROI) (450 MHz PC)

(2000)

Bertozzi et al. stereo, – 83% correct, 0.46 f.pos no – 1500 images,

(2004) visible 1897 ped. instances

Grubb et al. stereo, – 84% correct, 0.004 f.pos yes 23 Hz 2500 images,

(2004) visible (2.4 GHz PC) 14 different peds.

Gavrila and stereo, 10–25m 61–81% correct, 78–100% cor., 3–7 Hz sequence of

Munder visible in front [0.7–23] ×10

−3

f.pos 0.3–3.5 f.pos/min 17390 images,

(current up to 4m (speed unoptimized) 694 ped. instances

paper) lateral 59–75% correct, 78–100% cor., 7–15 Hz 17067 non-ped.

[1.0–27] ×10

−3

f.pos 0.4–5.1 f.pos/min (2.4 GHz PC) images

(speed optimized)

Fang et al. mono, – 84% correct, 19% f.pos no – 289 ped. images

(2003) FIR (summer)

92% correct, 3% f.pos

(winter)

at hand. Within this work, we aim to exploit the cascade

coupling of system modules.

Cascaded classiﬁers(Viola et al.,2005)haverecently

received increasing interest due to their computational

efﬁciency, and a number of publications addressed their

optimization. For example, Sun et al. (2004) and Luo

(2005) observed that the overall cascade performance

is optimal if the slope of the log-scale ROC curve is

equal for all nodes, given that the individual cascade

nodes are statistically independent. No such assump-

tion is made by Huo and Chen (2004), who recently

proposed a ROC “frontier-following” heuristic to suc-

cessively adjust the thresholds of a classiﬁer cascade.

The idea of analyzing the optimal front of ROC points

was ﬁrst utilized by Provost and Fawcett (2001) in the

context of classiﬁer comparison. They showed that, for

any misclassiﬁcation costs, the optimal classiﬁers are

located on the ROC convex hull. Here, we extend both

ideas (Huo and Chen, 2004; Provost and Fawcett, 2001)

by developing a technique to sequentially optimize a

cascade of complex system modules, each controlled

by a number of parameters, and by incorporating user-

deﬁned constraints.

This paper builds on earlier work described in

Gavrila et al. (2004). We consider the main contri-

butions of this paper the integration of modules in a

剩余18页未读，继续阅读

ooofly

粉丝: 2
资源: 3

移动车辆上的多线索行人检测与跟踪系统

行人检测的matlab实现

基于PYTHON的行人检测与轨迹追踪

行人检测跟踪

基于车载视觉的行人检测与跟踪方法

使用kalman滤波实现动态行人检测与跟踪

行人检测与跟踪技术探析

车载视觉行人检测与跟踪技术

MATLAB实现行人检测与跟踪技术

运动行人检测与跟踪关键算法提升与应用

YOLO与DeepSORT结合实现智能行人检测与跟踪

最新资源