Detection and Recognition of Dynamic Human Gestures in Cluttered Scenes by Kernel-Based Deformable Shape Models with Nonlinear/Non-Gaussian Bayesian Tracking and a Cascade of Dynamic Time Warping and Independent Classifiers

瀏覽統計 Email 通知 RSS Feed

簡歷

基本資料

Project title

Code/計畫編號

NSC98-2221-E019-038

Translated Name/計畫中文名

在凌亂的環境裡應用配合非線性與非高斯貝氏追蹤法之核可變形形狀模型及使用動態時間校準比對法與分類器分離的方式來偵測與辨識人體動態姿勢

Project Coordinator/計畫主持人

Chin-Chun Chang

Funding Organization/主管機關

National Science and Technology Council

Department/Unit

Department of Computer Science and Engineering

Website

https://www.grb.gov.tw/search/planDetail?id=1917505

Year

2009

Start date/計畫起

01-08-2009

Expected Completion/計畫迄

31-07-2010

Bugetid/研究經費

567千元

ResearchField/研究領域

資訊科學--軟體

Description

Abstract

"在凌亂的環境裡應用配合非線性與非高斯貝氏追蹤法之核可變形形狀模型，及使用動態時間校準比對法與分類器分離的方式來偵測與辨識人體動態姿勢動態姿勢辨識是發展先進人和電腦互動系統(human-computer interaction)的核心技術之一。這種系統的一個典型應用可為，根據事先定義的動作來辨識人的動作指令的命令解釋器。為了在一般環境下操作此系統，這種系統必須有能力在複雜環境中鎖定與分析目標肢體的動作。然而，這項技術目前仍舊是一個十分有挑戰性的問題。由於事先學習的動態動作樣版提供豐富的資訊來輔助鎖定目標肢體，我們擬將在這個三年計畫發展能在複雜環境裡偵測與辨識動態動作的技術。我們將整合可變形形狀偵測法(deformable shape detection)，動態時間校準比對 (dynamic time warping)，及非線性與非高斯貝氏追蹤法(nonlinear/non-Gaussian Bayesian tracking)來在複雜環境裡同時偵測與辨識動態動作。在這裡可變形形狀偵測法與及非線性與非高斯貝氏追蹤法將用來掌控整個偵測的程序，並用嘗試用蒙地卡羅平滑法(Monte Carlo smoothing)追蹤運動路徑。另外特殊的是，我們根據最近Lichtenauer 等人所提的動態姿勢辨識方式，將只用動態時間校準比對法於對位，對位後的動態姿勢樣本辨識則交由其他分類器負責。在這個架構下，事先學習的動態姿勢樣本所含有的空間與時間上的資訊將會被自然的應用出來。這三年擬將研究的主題如下。 􀂄 第一年將整合核可變形形狀模型，動態時間校準比對，與配合非線性與非高斯貝氏追蹤法在一般複雜環境，偵測與辨識單一肢體的動態動作。 􀂄 第二年將改進核可變形形狀模型，並使用動態時間校準比對，與配合粒子濾波 (Particle filtering)在一般複雜環境，偵測與辨識多個肢體的動態動作。 􀂄 第三年將研究如何藉由調適一個通用的動態動作模型來製作個人動態動作模型一般來減少學習動態動作樣本時間。同時我們也將研究是否可將前兩年的結果用部分結構(part-based)為主的方式來製作以容忍目標肢體更大的變形量。由於我們擬將研究的技術是整合幾個已經發展成熟的技術上，如動態時間校準比對，貝氏追蹤法，相關性過濾器(correlation filter)，核技術(kernel methods)等，其結果十分值得期待。""Detection and recognition of dynamic human gestures in cluttered scenes by kernel-based deformable shape models with nonlinear/non-Gaussian Bayesian tracking and a cascade of dynamic time warping and independent classifiers Dynamic gesture recognition is one of the core technologies to develop advanced input devices for natural human-computer interaction. A typical application of such devices is a command interpreter, which classifies the activity of a performer into one of the pre-defined actions according to the pre-learned prototypical dynamic gesture patterns. In order to operate such devices in general environments, these devices should be capable of locating and analyzing the target human motion in cluttered scenes, which is still a challenging problem. Since the pre-learned prototypical gesture patterns provide valuable cues for locating the target body part, we shall investigate the techniques of detection and recognition of dynamic gestures in cluttered scenes in this three-year project. In this three-year project, we shall integrate the kernel-based deformable shape model, the dynamic time warping algorithm, and nonlinear/non-Gaussian Bayesian tracking together to simultaneously detect and recognize dynamic gestures in cluttered scenes. Here, the kernel-based deformable shape model with Bayesian tracking will be used to control the whole detection process, and Monte Carlo smoothing methods will be used to trace out the trajectory of the target. According to the approach of Lichtenauer et al., the dynamic time warping algorithm will be only used to align captured motion pattern against the prototypical dynamic gesture patterns. The aligned motion pattern will be classified by other classifiers. Based on the framework which will be developed, the space and time information embedded in the prototypical dynamic gesture patterns will be utilized naturally. The topics of the three-year project are as follows. 􀂄 In the first year, we shall aim at integrating three algorithms, namely, the kernel-based deformable shape model, the dynamic time warping algorithm, and nonlinear/non-Gaussian Bayesian tracking, together to locate a single body part and recognize the type of the action of the body part in moderately cluttered scenes. 􀂄 In the second year, we shall focus on improving and integrating the kernel-based deformable shape model, the dynamic time warping algorithm, and the particle filtering, together to locate multiple target body parts and recognize the type of the action in cluttered scenes. 􀂄 In the third year, we shall focus on adaptation of generic dynamic gesture models of the target body part to those for an individual to reduce the complexity of the learning phase. Meanwhile, we shall study the possibility of a part-based implementation of the systems developed in the first two years to make the systems tolerate more shape deformation of the target body part. Since the technology which will be developed is based on several well developed algorithms such as the dynamic time warping algorithm, the kernel-based deformable shape model, nonlinear/non-Gaussian Bayesian tracking, the correlation filter, and the kernel method, the result of this project could be expected."

Keyword(s)

動態姿勢辨識
動態時間校準比對法
貝氏追蹤法
粒子濾波
可變形形狀偵測
dynamic gesture recognition
dynamic time warping
Bayesian tracking
particlefiltering
deformable shape detection