human action recognition | … from its eyes …

Paper review

year : 2010
employed database : six activities (aerobic data set) that are taken from a frontal-view camera; indoor; same illumination level; frame resolution 320 x 240 pixel.
application : [motion recognition] to overcome a motion self-occlusion problem due to motion overlapping.
result : recognition result > 94%; robustness analysis 93,8 %.
feature : DMHI 28-D, DMEI; seven Hu invariants; 64-D feature vector;
recognition method : appearance-based template matching; KNN; Euclidean measure.
comparation data/method : MMHI 14-D, HMHH; Zernike moments.
comment : dalam satu frame hanya ada satu action yang dikenali; In door di lab.

year : 2010
employed database : CMU human action (5 action in crowded; 1 action in a frame) 160 x 120 pixel; data in shopping mall.
application : [action detection] in complex scenes.
result : pushing button : (R/P : 0.2/1.0 & R/P : 0.8/0.9); two-handed wave : (R/P : 0.2/1.0 & R/P : 0.8/0.1).
feature : combination of MHI (motion feature), FI & HOG (appearance-based feature).
detection method : use the output of face detector or probabilistic trackers to estimates of the human bodies; use head detection; CNN.
learning method : SMILE-SVM (a novel – multiple instance learning); manually labelled to build positive & negative bags.
recognition method : locate the human actions in both spatial & temporal domains.
comparation data/method : MI-SVM; [14] method.
comment : hanya mengenali/ deteksi satu aksi.

year : 2011
employed database : 10 action; backgroundless; 8 uncalibrated cameras; 800 motion data separated into eight orientations.
application : estimate the possible orientations of an unregistered motion.
result : (reduced the searching time) for bin length 1: RR: 96%, ST: 15,1 ms; bin length 1/10: RR: 86%, ST: 11,6 ms.
feature : MHI; XOR image; eigenvector; KLT; PCA.
learning method : global eigenspaces.
recognition method : direction-oriented motion recognition approach; hierarchical eigenspaces; top-down & down-top manner.
comparation data/method : [21] [22].
comment : use SMoDB; satu aksi dalam satu frame; data di lab.

year : 2011
employed database : image obtain from a surveillance camera (outdoor scenes).
application : to find a person having a dufferent behavior or motion from others.
result : can detect the abnormal motion.
feature : Harris corner detector; coordinate space then converted to velocity, velocity magnitude & velocity orientation.
detection method : pyramidal search.
tracking method : The Lucas-Kanade tracker.
recognition method : (abnormal motion recog) is determined by the number of feature points clustered in each class; clustering by a look-up table.
comment : abnormal motion ditentukan oleh seorang yang berbeda motion dengan banyak orang; hanya mengenali 1 abnormal motion saja.

year : 2010.
employed database : Weizmann, KTH (single person actions in front of static backgrounds); UCF sports dataset; UCR videoweb.
application : to classify and localize human actions in video.
result : classification: (Weiz) 95.6% & (KTH) 92%; localization: (KTH) 0.89; classification: (UCF) 86.6% (UCR) 92.6%.
feature : spatio-temporal action Hough space; color & histograms of gradients; HSV color space & a local binary pattern.
detection method : Hough transform voting; Hough forest; spatio-temporal domain.
learning method : random trees; Houghforest structure.
tracking method : cuboid representation; cast votes for action label & the spatio-temporal center; Parzen window; the local maxima in the remaining; 3D Hough space.
recognition method : (classification) random forest.
comparation data/method : [21][29][17][23][22][31][14][18] & [20][36][34].
comment : untuk pengklasifikasi & pengalokasi aksi.

year : 2009.
employed database : Weizmann, KTH (single person actions in front of static backgrounds); locally collected gesture dataset-14 action-static background.
application : action recognition.
result : 91.07% (with dynamic background); 95.24% (with static background); 100% (Weizmann); 95.77% (KTH).
feature : kernel density estimation; QBMF; 256D shape descriptor; 256D motion descriptor.
detection method : maximizing a joint likelihood of the actor location; as [4].
learning method : joint shape & motion space via hierarchical k-means clustering.
tracking method : maximizing a joint likelihood of the action prototype; DFS on the learned binary tree; as [3].
recognition method : dynamic prototype sequence matching; frame-to-prototype matching; prototype-based sequence matching; k-NN classifier.
comparation data/method : [2][8][9][19][23][25].
comment : data lab; testing pakai moving camera and in background clutter and other moving objects; 1 aksi.

year : 2009
employed database : KTH; IXMAS multiview datashet.
application : incremental action recognition.
result : KTH : 87.7%; SR tree 30 x faster than exhaustive NN search; IXMAS : with 4 camera views : 69.6%, with voting improve 72.6%.
feature : local spatio-temporal feature; SR tree.
detection method : spatio temporal interest point detector: 2D Gaussian filter & 1D Gabor filter.
learning method : SR tree.
tracking method : –
recognition method : simple voting method by nearest neighbor (NN) feature.
comparation data/method : [25][14][22][27][3][15][23].
comment : pengenalan action untuk max 2 aksi dalam 1 frame; background statis (lab).

year : 2011
employed database : INRIA datashet.
application : action recognition.
result : KTH : detection rate: 87%; tracking rate: 49%; recognition rate: 93%.
feature : gradient feature.
detection method : HOG + integral histogram.
tracking method : pyramid Lucas-Kanade method.
recognition method : motion template, SVM.
comment : pengenalan action untuk max 3 aksi dalam 1 frame; background outdoor.