利用骨架與深度資訊擷取有意義之特徵以辨識日常活動

Daily Activity Recognition Using the Informative Features from Skeletal and Depth Data

陳冠伃(Kuan-Yu Chen)

指導教授 ： 傅立成

國立臺灣大學/電機資訊學院/資訊網路與多媒體研究所/碩士(2015年)

https://doi.org/10.6342/NTU.2015.10780

摘要

近年來，動作辨識已成為在電腦視覺領域中相當熱門的研究主題，其中以”辨識日常活動”更能實際運用於我們真正的生活環境裡。為了使系統能夠以最自然的方式來解讀人類精細的動作，我們採取以視覺為基礎來設計系統。雖然至今已經有許多關於活動辨識上的研究，但若要提高系統的實用性，則解決資料變異性與環境雜訊的問題便日趨重要。因此，在本篇論文中，我們的目標為建立出一個使用深度攝影機的強健式日常活動辨識系統。
　　我們的方法主要專注於分析含有重要信息量的動作狀態，因為根據我們的觀察，大多數的日常活動只和一些特定的身體部位相關，尤其集中在上半身，例如頭和手之間的互動。基於這樣的概念，我們提出兩種新穎且具有直覺物理意義的特徵描述子，分別為位置移動量直方圖以及基於局部深度運動圖的賈伯小波表徵，這些特徵可以有效率地萃取具有鑑別力的姿勢與動態線索。藉由結合骨架與深度資訊，利用各自的優勢並強調其可靠的局部特徵來加以強化辨識能力。最後，我們運用主成分分析和線性判別分析方法來有效降低特徵的維度，然後將得到的特徵向量訓練出一支持向量機，以達到日常活動辨識之目的。經過實驗結果評估，本論文中所提出的方法不但能夠有效辨識日常動作，且相較於其他方法更表現出優越性。由於我們提出的活動辨識方法能處理真實環境下容易遇到的問題，這將有助於未來實際應用於照護系統或人機互動的介面上。

關鍵字

動作辨識 ； 日常活動 ； 深度影像 ； 骨架關節點 ； 富有信息量的特徵

並列摘要

Human action recognition has become an active topic of computer vision research in recent years, and recognizing activities of daily living (ADLs) is practically helpful in our daily life. Vision-based system is chosen so that computer can understand complicated activities in a nature way. However, it still remains some challenges such as intra-class variations and environmental noise in this field. In order to solve such problems, we present a robust activity recognition system with a depth sensor.
We mainly focus on the motion analysis of informative body parts, since most activities are much associated with these particular parts, e.g., head and hands in upper body. Based on the idea, we propose two novel features with intuitive physical meaning, which are Histogram of Located Displacements (HOLD) and Local Depth Motion Maps (L-DMM) based Gabor representation. They can capture discriminative posture and motion cues from skeletal joints and depth data respectively. Combing the advantages of joint and depth features as well as emphasizing the reliable parts can enhance the robustness of classification ability. Finally, we apply Principal Component Analysis (PCA) and Linear Discriminant Analysis (LDA) for effective dimension reduction, and then use Support Vector Machine (SVM) to acquire the classification results. The experimental results have shown effectiveness of our method and demonstrated superior performance over the state-of-the-art works. This approach can benefit to several applications such as health care system and human-computer interaction (HCI).

並列關鍵字

Activity recognition ； Activity of daily living ； Depth maps ； Skeletal joints ； Informative feature

參考文獻

[1] P. Turaga, R. Chellappa, V. S. Subrahmanian, and O. Udrea, "Machine recognition of human activities: A survey," IEEE Transactions on Circuits and Systems for Video Technology, vol. 18, pp. 1473-1488, 2008.

[2] J. K. Aggarwal and M. S. Ryoo, "Human activity analysis: A review," ACM Computing Surveys (CSUR), vol. 43, p. 16, 2011.

[3] C. Zhang and Y. Tian, "Rgb-d camera-based daily living activity recognition," Journal of Computer Vision and Image Processing, vol. 2, p. 12, 2012.

[4] P. Dollár, V. Rabaud, G. Cottrell, and S. Belongie, "Behavior recognition via sparse spatio-temporal features," in 2nd Joint IEEE International Workshop on Visual Surveillance and Performance Evaluation of Tracking and Surveillance, 2005, pp. 65-72.

[5] I. Laptev, "On space-time interest points," International Journal of Computer Vision, vol. 64, pp. 107-123, 2005.