Table 3.

Comparison with other state-of-the-art methods on the datasets HDM05-65, HDM05-14, CMU-30, and CMU-14. The results are presented in top-down order. The results using our approach are shown in bold text. * The CMU dataset with 14 classes was used, as well as the dataset recorded with their own Vicon (http://www.vicon.com/) motion capture device. Note: Normalized Trajectories (NT), relative positions of joints (PO), Temporal Differences (TD), Tree Structure Vector Quantized (TSVQ), Hierarchically Bidirectional Recurrent Neural Network with Long-Short Term Memory (LSTM) (HBRNN-L). For more details, see Section 2 and Section 4.3.2.

Dataset	Approach	Algorithm	Features	Training	Accuracy
HDM05-65	Du et al. [19]	HBRNN-L	NT	$90\%$	$96.90\%$
		HURNN-L	NT		$96.70\%$
		DBRNN-L	NT		$96.70\%$
		DURNN-L	NT		$96.62\%$
	Cho et al. [18]	Hybrid MLP $(\lambda =0.5)$	PO+TD		$95.59\%$
		Hybrid MLP $(\lambda =0.9)$	PO+TD		$95.55\%$
		Hybrid MLP $(\lambda =0.1)$	PO+TD		$95.46\%$
		Hybrid MLP $(\lambda =0.1)$	PO+TD+NT		$95.21\%$
		MLP	PO+TD		$95.20\%$
	Our approach	DNN	NT+Keyframes		$\mathbf{95.14}\%$
	Cho et al. [18]	SVM	PO+TD+NT		$95.12\%$
		Hybrid MLP $(\lambda =0.9)$	PO+TD+NT		$95.04\%$
		SVM	PO+TD		$94.95\%$
		MLP	PO+TD+NT		$94.86\%$
		Hybrid MLP $(\lambda =0.5)$	PO+TD+NT		$94.82\%$
	Du et al. [19]	DBRNN-T	NT		$94.79\%$
		DURNN-T	NT		$94.63\%$
	Sedmidubsky et al. [1]	CNN+KNN	NT		$93.9\%$
	Cho et al. [18]	ELM	PO+TD+NT		$92.76\%$
		ELM	PO+TD		$91.57\%$
HDM05-14	Our approach	DNN	NT+Keyframes	$50\%$	$\mathbf{98.6}\%$
	Sedmidubsky et al. [1]	CNN+KNN	NT		$94.3\%$
	Elias et al. [62]	CNN+KNN	NT		$87.4\%$
CMU-30	Kadu and Kuo [37]	Two-Step SVM Fusion	TSVQ	$80\%$	$99.6\%$
	Our approach	DNN	NT+Keyframes		$\mathbf{99.3}\%$
	Kadu and Kuo [37]	Two-Step Score Fusion	TSVQ		$98.2\%$
		Pose-Histogram Classifier	B-PL04		$95.6\%$
			B-PL06		$95.6\%$
		Motion-String Similarity	A-SL12		$95.6\%$
			A-SL13		$95.6\%$
		Pose-Histogram Classifier	B-PL05		$95.3\%$
			B-PL03		$92.8\%$
		Motion-String Similarity	A-ML12		$82.3\%$
			A-ML13		$80.5\%$
CMU-14	Our approach	DNN	NT+Keyframes	$85\%$	$\mathbf{98.5}\%$
	Wu et al. [44]*	Hierarchical Tree	3D Trajectories		$98.1\%$
	Wu et al. [48]*	Smith–Waterman	3D Trajectories		$97.0\%$