第五章 實驗結果
5.1 影像系統辨識在不同角度下之人體辨識結果
如圖 5-1 所示，0o表示測試者人的正面對著攝影機，180o為人背對著攝影機。
此實驗中為三位不同測試者，每相隔 45o進行實驗，分別以 0o、45o、90o、135o、 180o、225o、270o與 315o等八種不同角度作為辨識如圖 5-2 至圖 5-10。M 表示為 待測人員與攝影機之間的距離，並且每 0.5 公尺為一間距，共有 2.5m、3m、3.5m、
4m、4.5m、5m、5.5m 等七種不同距離如圖 5-11 至圖 5-16。表 5.1 顯示受測者在 面對鏡頭的不同角度下，辨識人體存在的正確率。，實驗過程中，每一種角度皆 以七種不同距離進行辨識，且每次皆辨識 30 秒(每測試者約 50 張影像畫面，三 位測試者共約 150 張影像畫面)，藉以觀察本系統僅使用影像處理偵測環境中人 體的正確率。
圖 5-2 當人面對攝影機角度為 0 度 圖 5-3 當人面對攝影機角度為 45 度
圖 5-4 當人面對攝影機角度為 90 度 圖 5-5 當人面對攝影機角度為 135 度
圖 5-6 當人面對攝影機角度為 180 度 圖 5-7 當人面對攝影機角度為 225 度
圖 5-8 當人面對攝影機角度為 270 度 圖 5-9 當人面對攝影機角度為 315 度
圖 5-10 當人面對攝影機距離為 2.5 公尺 圖 5-11 當人面對攝影機距離為 3 公尺
圖 5-12 當人面對攝影機距離為 3.5 公尺 圖 5-13 當人面對攝影機距離為 4 公尺
圖 5-14 當人面對攝影機距離為 4.5 公尺 圖 5-15 當人面對攝影機距離為 5 公尺
圖 5-16 當人面對攝影機距離為 5.5 公尺
表 5-1 三位測試者在各種角度與距離之辨識結果
Accurate 82.75% 84.66% 82.88% 82.28% 81.65% 78.35% 78.35% 81.56%
在表 5-1 的實驗結果中，在相同距離不同角度之下，可以知道人以 0o和 180o
表 5-2 測試者在 2.5 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 125 1 0 0 4 6
Recognition rate 83.89% 86.52% 84.74% 84% 80.55% 72.84%
Average
recognition rate 82.09%
表 5-3 測試者在 3 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 122 0 0 0 4 7
Recognition rate 82.99% 85.33% 85% 82.85% 80.26% 67.56%
Average
recognition rate 80.67%
表 5-4 測試者在 3.5 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 121 0 0 0 3 6
Recognition rate 82.31% 82% 81.13% 81.08% 80% 65.77%
Average
recognition rate 78.71%
表 5- 5 測試者在 4 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 121 1 1 2 2 10
Recognition rate 80.66% 81.33% 80.95% 79.1% 77.63% 63.94%
Average
77.26%
表 5- 6 測試者在 4.5 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 121 0 0 2 2 10
Recognition rate 81.20% 79.33% 78.68% 76.05% 73.33% 61.07%
Average
recognition rate 74.94%
表 5- 7 測試者在 5 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 111 2 0 1 0 11
Recognition rate 74.49% 73.5% 71.42% 70.42% 68.91% 57.14%
Average
69.31%
表 5-8 測試者在 5.5 公尺距離之辨識結果 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 110 2 0 0 1 14
Recognition rate 72.36% 72.36% 71.21% 68% 67.5% 58.78%
Average
recognition rate 68.36%
表 5-9 2.5 公尺至 5.5 公尺之平均辨識率 Accurate rate
Distance
Standing Sitting Walking Upstairs Downstairs Lying
2.5m 83.89% 86.52% 84.74% 84% 80.55% 72.84%
3m 82.99% 85.33% 85% 82.85% 80.26% 67.56%
3.5m 82.31% 82% 81.13% 81.08% 80% 65.77%
4m 80.66% 81.33% 80.95% 79.10% 77.63% 63.94%
4.5m 81.20% 79.33% 78.68% 76.05% 73.33% 61.07%
5m 74.49% 73.50% 71.42% 70.42% 68.91% 57.14%
5.5m 72.36% 72.36% 71.21% 68% 67.50% 58.78%
Average
recognition rate 79.7% 80.05% 79.01% 77.35% 75.45% 63.87%
結果，其辨識率定義為找到畫面中人的存在並且能正確完成姿態的辨識。
從表 5-2 至 5-8 的實驗結果中，站、坐、上樓、下樓、走路、躺等六種姿態 中在七種不同距離下之總辨識率，依序為 79.7%、80.05%、79.01%、77.35%、75.45%
以及 63.87%。由於必須先辨識出人體，再進一步辨識六種人體姿態，因此辨識 率最高為辨識出人體的 81.56%，若已經確認人體後才做姿態辨識，則應能提高 姿態辨識率。
其中靜態姿態中的站立辨識率最高，躺下的辨識率則是六種姿態中最差。站 立的辨識率與上一節實驗結論相同，肩膀的橢圓樣板比對時較容易辨識，坐下姿 態則因為椅子的椅背，會讓肩膀的橢圓樣板比對造成誤判而辨識失敗，躺下則因 為人體的肩膀緊貼地面，僅剩上半部肩膀可以進行比對，使得辨識率為所有姿態 中最低。在動態姿態中，走路姿態辨識率為最高，上樓次之，下樓則為最低，主 因是人體在走路姿態時，在前後影像中所辨識出的人頭位置變動幅度最大，不容 易誤判為其他姿態。
人在下樓過程中眼睛常會低頭看著往下的樓梯，使得影像系統在低頭與抬頭 過程中誤判為上樓姿態，而移動較為緩慢與手臂的擺動幅度，則使得系統誤判為 坐下姿態。上樓的姿態辨識則與下樓姿態相同，當人抬頭看著往上的樓梯時，與 平視前方時候的人頭位置，前後位置的差異造成了姿態辨識上的誤判。
5.3.1 以加速規進行姿態辨識
表 5-10 為三位不同測試者使用加速度量規對每位測試者對站、坐、躺、上樓、
下樓、走路等六種姿態逐項做 20 次之辨識結果，每兩秒傳送一次辨識結果至機 器人端。其中辨識率最高者為躺下與站立姿態，可達 100%，辨識率最低則為下 樓的 80%。
表 5-10 加速度量規辨識六種姿態結果 Actual pose
Estimated pose Standing Sitting Walking Upstairs Downstairs Lying
Standing 60 53 0 0 0 0
recognition rate 90.67%
5.3.2 融合影像系統之姿態辨識結果
辨識站立姿態由 80.24%改善至 98.044%，坐下姿態由 79.7%改善至 98.09%，走 路由 79.53%改善至 96.42%，上樓由 77.35%改善至 87.13%，下樓由 75.45%改善 至 84.28%，而表 5-9 中加速規的辨識率，站立與躺下姿態雖然由 100%降至 98.04%
與 97.37%，但是辨識坐下姿態由 89%改善至 98.09%，走路由 90%改善至 96.42%，
上樓由 85%改善至 87.13%，下樓由 80%改善至 84.28%。
表 5-11 融合後辨識結果(2.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 58 1 0 0 1 0
recognition rate 92.44%
表 5-12 融合後辨識結果(3 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 0 0 0 1 1
Recognition rate 98.33% 96.66% 96.66% 86.66% 85% 98.33%
Average
recognition rate 93.6%
表 5-13 融合後辨識結果(3.5 公尺)
Actual pose Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 0 0 0 0 1
Recognition rate 98.33% 96.66% 96.66% 86.66% 85% 98.33%
Average
recognition rate 93.6%
表 5-14 融合後辨識結果(4 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 60 0 0 0 1 1
Recognition rate 100% 98.33% 98.33% 88.33% 83.33% 98.33%
Average
recognition rate 94.41%
表 5-15 融合後辨識結果(4.5 公尺)
Actual pose Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 1 0 0 0 1
Recognition rate 98.33% 98.33% 96.66% 90% 88.33% 96.66%
Average
recognition rate 94.71%
表 5-16 融合後辨識結果(5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 0 0 1 1 0
Recognition rate 98.33% 100% 96.66% 86.66% 81.66% 100%
Average
recognition rate 93.89%
表 5-17 融合後辨識結果(5.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 58 2 0 0 0 2
recognition rate 92.21%
表 5-18 融合後 2.5 公尺至 5.5 公尺之平均辨識率 Accurate
Distance
Standing Sitting Walking Upstairs Downstairs Lying
2.5m 96.66% 98% 95% 86.66% 83.33% 95%
3m 98.33% 96.66% 96.66% 86.66% 85% 98.33%
3.5m 98.33% 96.66% 96.66% 86.66% 85% 98.33%
4m 100% 98.33% 98.33% 88.33% 83.33% 98.33%
4.5m 98.33% 98.33% 96.66% 90% 88.33% 96.66%
5m 98.33% 100% 96.66% 86.66% 81.66% 100%
5.5m 96.66% 96.66% 95% 85% 83.33% 96.66%
Average
recognition rate 98.09% 97.80% 96.42% 87.13% 84.28% 97.61%
5.4 不同測試者之連續姿態辨識結果
表 5-19 至表 5-25 為三位測試者依序作出是坐站走路上樓下樓躺等 六種連續姿態，並且是在和攝影機不同距離下結合加速度量規，藉以觀察本系統 辨識一位使用者之連續姿態的辨識結果。
由實驗結果中發現，當人體作出連續動作時的辨識率與前一小節實驗人員維持 固定姿態下之辨識率差異不大，此實驗平均辨識率皆維持在 93%，可證明本系統 並不會因為人體姿態的轉變造成辨識率的下降。
表 5-19 融合後連續姿態辨識結果(2.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 0 0 0 1 1
Sitting 1 60 0 1 2 0
Walking 0 0 57 2 3 0
Upstairs 0 0 1 51 1 0
Downstairs 0 0 1 3 51 0
Lying 0 0 0 0 0 57
Fail 0 0 1 2 2 2
Recognition rate 98.33% 100% 95% 85% 85% 95%
Average
recognition rate 93%
表 5-20 融合後連續姿態辨識結果(3 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 58 0 0 0 1 0
recognition rate 93.6%
表 5-21 融合後連續姿態辨識結果(3.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 58 0 0 0 2 0
Recognition rate 96.66% 100% 93.33% 85% 85% 98.33%
Average
recognition rate 93%
表 5-22 融合後連續姿態辨識結果(4 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 60 0 0 0 1 0
recognition rate 93.6%
表 5-23 融合後連續姿態辨識結果(4.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 60 1 0 0 1 1
Recognition rate 100% 98.33% 93.33% 88.33% 85% 96.66%
Average
recognition rate 93.6%
表 5-24 融合後連續姿態辨識結果(5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 58 0 0 0 1 0
Recognition rate 96.66% 100% 91.66% 85% 81.66% 96.66%
Average
recognition rate 91.94%
表 5-25 融合後連續姿態辨識結果(5.5 公尺) Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 59 0 0 0 0 0
Recognition rate 98.33% 100% 90% 88.33% 86.66% 98.33%
Average
recognition rate 93.6%
表 5-26 融合後 2.5 公尺至 5.5 公尺之平均辨識率 Actual pose
Estimated pose
Standing Sitting Walking Upstairs Downstairs Lying
Standing 413 2 0 1 4 4
Sitting 2 414 0 2 2 1
Walking 0 0 402 7 24 0
Upstairs 0 0 2 368 15 0
Downstairs 0 0 4 20 356 0
Lying 2 3 0 0 1 410
Fail 3 4 12 22 19 5
Recognition rate 98.33% 100% 90% 88.33% 86.66% 98.33%
Average
recognition rate 93.6%
第六章 結論與未來展望
6.1 結論
本論文提出的基於影像處理之人體姿態辨識系統，可成功辨識站、坐、躺、
走路、上樓、下樓等六種不同的人體姿態。在室內光源無劇烈變化，且環境周圍 為單純之牆壁時，人體與攝影機之間距離在 2.5~5.5 公尺下，其整體辨識率為 79%，而本實驗室研發的人體姿態感測模組整體辨識率為 88%。經由類神經網路 結合後，可將辨識率提升至 93.5%，由於人體姿態感測模組每兩秒傳送一次辨識 結果，因此在兩秒之間發生姿態轉變的情形，就很有可能發生誤判，使辨識率略 為降低。
6.2 未來展望
本論文在影像辨識系統方面，目前只能對單一人員作姿態辨識，由於在人體 的樣本比對上較為費時，當影像中的人物快速移動下可能會無法判別，因此若未 來能夠加以改良比對方法使其運算速度上有更大的提升，將可提高其整體的正確 性。目前系統在多人情況下只能辨別出畫面中的不同人員，無法對每個人員作姿 態辨識與追蹤，因此將來若能追蹤與辨別不同人員，可提高系統的實用性。
參考文獻
[1] C. Wren, A. Azarbayejani, T. Darrell and A. Pentland, “Pfinder:Real-Time Tracking of the Human Body,” IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 19, no. 7, pp.780-785, July 1997.
[2] I. Haritaoglu, D. Harwood and L. S. Davis, “W4: real-time surveillance of people and their activities,” IEEE Transactions on Pattern Analysis and Machine
Intelligence, Vol. 22, No.8, pp.809–830, August 2000.
[3] S. J. McKenna, S. Jabri, Z. Duric, A. Rosenfeld, and H. Wechsler, “Tracking groups of people,” Computer Vision and Image Understanding, vol. 80, pp.
42-56, 2000.
[4] Linda Siana, “A Study of Human Tracking and Face Detection on A
Pan-Tilt-Zoom Camera, ” Master’s thesis, Dept. of of Electrical and Control Engineering, National Chiao-Tung University, 2005.
[5] K. Keiichi, T. Tomonaka, S. Shiotani, Y. Koketsu and M. Iehara; “Recogniaing Human Behaviors with Vision Sensors in Network Robot System,”
Proceedings of the 2006 IEEE International Conference on Robotics and Automation, Orlando, Florida, USA ,2006, pp.1274-1279.
[6] I. Haritaoglu, D. Harwood and L. S. Davis, “Ghost: A Human Body Part Labeling System Using Silhouettes,” International Conferenceon Pattern Recognition, Brisbane, Australia, 1998, pp. 77 – 82.
[7] Sergio Monteiro, and Estela Bicho, “A Dynamical Systems Approach to Behavior-Based Formation Control,” Proceedings of IEEE International Conference on Robotics and Automation, 2002, pp. 2606-2611.
[8] J.W. Davis,; “Hierarchical motion history images for recognizing human
motion,” IEEE Workshop on Detection and Recognition of Events in Video, 2001,
pp.39- 46.
[9] S. Ju, M. Black and Y. Yaccob, “Cardboard people: a parameterized model of articulated image motion,” in Automatic Face and Gesture Recognition, 1996., Proceedings of the Second International Conference on, 1996, pp. 38-44.
[10] 陳福泰, “以移動歷史影像為基礎之人類行為辨識＂,中華大學資訊工程研究 所碩士論文,2003.
[11] J.Yamato, J. Ohya and K. Ishii, “Recognizing human action in time-sequential images using hidden Markov model”, IEEE Conference on Computer Vision and Pattern Recognition, Champaign, Illinois, 1992, pp. 379-385.
[12] Toru Nakata, “Recognizing Human Activities in Video by Multi-resolutional Optical Flows,” Proceeding of the 2006 IEEE/RSJ International Conference on Intelligent Robots and Systems, Beijing, China, 2006, pp. 1973-1978.
[13] M. Soriano, S. Huovinen, B. Martinkauppi, and M. Laaksonen, “Using the skin Locus to cope with Changing Illumination Conditions in Color-Based Face Tracking, ” in Proc. Of IEEE Nordic Signal Processing Symposium, Kolmarden, Sweden, 2000, pp. 383-386.
[14] F. Chang, C-J. Chen and C-J. Lu, “A Linear-Time Component-Labeling Algorithm Using Contour Tracing Technique,” Computer Vision and Image Understanding, vol. 93, no. 2, pp. 206-220, 2004.
[15] Gonzalez and Woods, Digital Image Processing, 2nd edition, Prentice Hall, 2002 [16] 楊岱璋, “即時人體偵測追蹤的姿勢分析系統＂, 國立清華大學電機工程研
究所碩士論文, 2002.
[17] Y. Matsumoto, T. Ogasawara, A. Zelinsky, “Behavior Recognition Based on Head Pose and Gaze Direction Measurement”, Nara Institute of Science and Technology,Japan, pages 2127-2132.
Sensing System for Recognition and Accumulation of Human Behavior,”
Intelligent Robots and Systems, Japan,2000,pp. 344-350.
[19] I. Rius, J. Varona, J. Gonzalez and Juan J. Villanueva,”Action spaces for Efficient Bayesian Tracking of Human Motion,” Proceedings of the 18th International Conference on Pattern Recognition, Spain, 2006, pp.472-475.
[20] J. Zhou and J. Hoang, “Real Time Robust Human Detection and Tracking System,” Proceedings of the 2005 IEEE computer Society Conference on Computer Vision and Pattern Recognition, Canada ,2005, pp. 149-149.
[21] Kai-Tai Song and Chun-Wei Chen, “Multi-Person Pose Recognition Using a Zigbee Sensor Network,” in Proc. of 17th IFAC World Congress, Seoul , Korea , 2008 , pp. 14976-14981.
[22] T. Zhao and R. Nevatia, “Tracking Multiple Humans in Complex Situations,”
IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 26, no. 9, September 2004, pp. 1208-1221.
[23] E. Oron , “Motion estimation and image difference for multi-object tracking, ” Proc. Of IEEE Conf. On Aerospace, 1999, pp. 401-409.