Human action recognition is one fundamental challenge in robotics systems. In this paper, we propose one lightweight action recognition architecture based on deep neural networks (DNNs) just using RGB data. The proposed architecture consists of convolution neural network (CNN), long short-term memory (LSTM) units, and temporal-wise attention model. First, CNN is used to extract spatial features to distinguish objects from the background with both local and semantic characteristics. Second, two kinds of LSTM networks are performed on the spatial feature maps of different CNN layers (pooling layer and fully-connected layer) to extract temporal motion features. Then, one temporal-wise attention model is designed after LSTM to learn which parts in which frames are more important. Lastly, a joint optimization module is designed to explore intrinsic relations between two kinds of LSTM features. Experimental results demonstrate the efficiency of the proposed method.

This letter proposes an ensemble neural network (Ensem-NN) for skeleton-based action recognition. The Ensem-NN is introduced based on the idea of ensemble learning, “two heads are better than one”. According to the property of skeleton sequences, we design one-dimensional (1D) convolution neural network (CNN) with residual structure as Base-Net. From entirety to local, from focus to motion, we designed four different subnets based on the Base-Net to extract diverse features. The first subnet is a Two-stream Entirety Net, which performs on the entirety skeleton and explores both temporal and spatial features. The second is a Body-part Net, which can extract fine-grained spatial and temporal features. The third is an Attention Net, in which a channel-wised attention mechanism can learn important frames and feature channels. Frame-difference Net, as the fourth subnet, aims at exploring motion features. Finally, the four subnets are fused as one ensemble network. Experimental results show that the proposed Ensem-NN performs better than state-of-the-art methods on three widely used datasets.

The divide-and-conquer strategy has been widely used in cooperative co-evolutionary algorithms to deal with large-scale global optimization problems, where a target problem is decomposed into a set of lower-dimensional and tractable subproblems to reduce the problem complexity. However, such strategy usually demands a large number of function evaluations to obtain an accurate variable grouping. To address this issue, a merged differential grouping method is proposed in this paper based on subset-subset interaction and binary search. In the proposed method, each variable is firstly identified as either a separable variable or a non-separable variable. Afterward, all separable variables are put into the same subset, and the non-separable variables are divided into multiple subsets using a binary-tree-based iterative merging method. With the proposed algorithm, the computational complexity of interaction detection is reduced to O(max{n, nns ×log2 k}), where n, nns(≤n), and k(<n) indicate the numbers of decision variables, non-separable variables, and subsets of non-separable variables, respectively. The experimental results on benchmark problems show that merged differential grouping is very competitive with the other state-of-the-art methods in terms of efficiency and accuracy of problem decomposition.