在本研究當中，對於人類骨架的動作辨識提出一種確實且效率高的解決方式。該架構建立於 Sparse Coding 與 Temporal Pyramid Pooling 技術，當中稀疏編碼已證實對彩色影片中物件辨識有相當好的效果。本研究架構起始於從骨架資訊中萃取出 Moving Pose Descriptor 用以代表人類姿勢的特徵。接著將得到之特徵按照若干個身體部位做切分，再以稀疏編碼對各部位的特徵做編碼。最後，合併各部位特徵並用時間金字塔池化來提取時間序列上之資訊。分類方式採用 Linear Support Vector Machine。經實驗證實，在三個廣泛使用於骨架動作辨識與一個骨架跌倒偵測之數據集的實驗結果中本研究方法所需的訓練範本數量與辨識準確度上皆優於其他最先進的方式。

This thesis presents an efficacious and efficient approach to skeleton-based human action recognition based on the sparse coding technique, which has shown impressive results in object classification in RGB video. The new approach first extracts the moving pose descriptor feature to represent human poses from skeleton information.
Thereafter, we separate the human skeleton features into several parts and encode each part by sparse coding.
Finally, the encoded features are concatenated into compact vectors using the temporal pooled time series and then classified by SVM. Experimental results reveal that the proposed approach outperforms the main state-of-the-art works on three widespread public skeleton-based action datasets and one fall detection dataset in terms of accuracy and the number of training samples required.

Pooled motion features for first person videos
Keep it simple and sparse: real-time action recognition
Robust 3d action recognition with random occupancy patterns
Mining actionlet ensemble for action recognition with depth cameras
Sequence of the most informative joints (smij): A new representation for human skeletal action recognition
Human action recognition by representing 3d skeletons as points in a lie group
Hierarchical recurrent neural network for skeleton based action recognition
The moving pose: An e cient 3d kinematics descriptor for low-latency action recognition and detection
An approach to pose-based action recognition
Online dictionary learning for sparse coding
K-svd: An algorithm for designing overcomplete dictionaries for sparse representation
Learning sparse representations for human action recognition
Action recognition based on a bag of 3d points
View invariant human action recognition using histograms of 3d joints
Recognizing actions from depth cameras as weakly aligned multi-part bag-of-poses
Depth-based human fall detection via shape features and improved extreme learning machine
Fusing spatiotemporal features and joints for 3d action recognition
Real-time human action recognition based on depth motion maps
Histogram of oriented displacements (hod): Describing trajectories of human joints for action recognition
Coupled hidden conditional random fi elds for rgb-d human action recognition
Elastic functional coding of human actions: From vector- elds to latent variables
Grassmannian representation of motion depth for 3d human gesture and action recognition
Scale invariant human action detection from depth cameras using class templates