Action recognition using spatial-optical data organization and sequential learning framework

Action recognition has already been a heated research topic recently, which attempts to classify different human actions in videos. The current main-stream methods generally utilize ImageNet-pretrained model as features extractor, however it’s not the optimal choice to pretrain a model for classifying videos on a huge still image dataset. What’s more, very few works notice that 3D convolution neural network(3D CNN) is better for low-level spatial-temporal features extraction while recurrent neural network(RNN) is better for modelling high-level temporal feature sequences. Consequently, a novel model is proposed in our work to address the two problems mentioned above. First, we pretrain 3D CNN model on huge video action recognition dataset Kinetics to improve generality of the model. And then long short term memory(LSTM) is introduced to model the high-level temporal features produced by the Kinetics-pretrained 3D CNN model. Our experiments results show that the Kinetics-pretrained model can generally outperform ImageNet-pretrained model. And our proposed network finally achieve leading performance on UCF-101 dataset.

This paper considers the recognition of realistic human actions in videos based on spatio-temporal interest points (STIPs). Existing STIP-based action recognition approaches operate on intensity representations of the image data. Because of this, these approaches are sensitive to disturbing photometric phenomena, such as shadows and highlights. In addition, valuable information is neglected by discarding chromaticity from the photometric representation. These issues are addressed by color STIPs. Color STIPs are multichannel reformulations of STIP detectors and descriptors, for which we consider a number of chromatic and invariant representations derived from the opponent color space. Color STIPs are shown to outperform their intensity-based counterparts on the challenging UCF sports, UCF11 and UCF50 action recognition benchmarks by more than 5% on average, where most of the gain is due to the multichannel descriptors. In addition, the results show that color STIPs are currently the single best low-level feature choice for STIP-based approaches to human action recognition.

The accuracy of video-based action recognition depends largely on the extraction and utilization of optical flow, especially in two-stream networks. The original intention of the introduction of optical flow is to use the time information contained in video, however, the subsequent work shows that optical flow is useful for action recognition because it is invariant to appearance. In this article, we study and discuss this point of view, and propose optical flow enhancement algorithms to improve action recognition accuracy. Our enhancement algorithms improve the invariance to appearance of the representation in optical flow without losing time information, and every action recognition network with optical flow can benefit from our algorithms. We conduct a series of experiments to validate the influence of the proposed algorithms with TSN in terms of several datasets and optical flow calculation methods. As a result, we prove that first order differential algorithms are effective, TSN with our enhancement module significantly outperform original network. Based on these experiments, we also verify the importance of invariance to appearance in optical flow, and provide a reference for the follow-up study of improving action recognition accuracy.

Asymmetric 3D Convolutional Neural Networks for action recognition

We describe a “bag-of-rectangles” method for representing and recognizing human actions in videos. In this method, each human pose in an action sequence is represented by oriented rectangular patches extracted over the whole body. Then, spatial oriented histograms are formed to represent the distribution of these rectangular patches. In order to carry the information from the spatial domain described by the bag-of-rectangles descriptor to temporal domain for recognition of the actions, four different methods are proposed. These are namely, (i) frame by frame voting, which recognizes the actions by matching the descriptors of each frame, (ii) global histogramming, which extends the idea of Motion Energy Image proposed by Bobick and Davis by rectangular patches, (iii) a classifier based approach using SVMs, and (iv) adaptation of Dynamic Time Warping on the temporal representation of the descriptor. The detailed experiments are carried out on the action dataset of Blank et. al. High success rates (100%) prove that with a very simple and compact representation, we can achieve robust recognition of human actions, compared to complex representations.

IntroductionRecently, numerous studies have focused on the semantic decoding of perceived images based on functional magnetic resonance imaging (fMRI) activities. However, it remains unclear whether it is possible to establish relationships between brain activities and semantic features of human actions in video stimuli. Here we construct a framework for decoding action semantics by establishing relationships between brain activities and semantic features of human actions.MethodsTo effectively use a small amount of available brain activity data, our proposed method employs a pre-trained image action recognition network model based on an expanding three-dimensional (X3D) deep neural network framework (DNN). To apply brain activities to the image action recognition network, we train regression models that learn the relationship between brain activities and deep-layer image features. To improve decoding accuracy, we join by adding the nonlocal-attention mechanism module to the X3D model to capture long-range temporal and spatial dependence, proposing a multilayer perceptron (MLP) module of multi-task loss constraint to build a more accurate regression mapping approach and performing data enhancement through linear interpolation to expand the amount of data to reduce the impact of a small sample.Results and discussionOur findings indicate that the features in the X3D-DNN are biologically relevant, and capture information useful for perception. The proposed method enriches the semantic decoding model. We have also conducted several experiments with data from different subsets of brain regions known to process visual stimuli. The results suggest that semantic information for human actions is widespread across the entire visual cortex.

Spatio-temporal features for fast early warning of unplanned self-extubation in ICU

Recognizing Human action in video is an very active research topic. There are a growing variety of human action datasets with different video length, different practitioners. Make human action recognition becomes a very difficult topic. A majority researchers solve the problem by extracting key frames from the videos. Most paper use feature Clustering methods to extract key frames in videos. On one hand, the large variety of visual content in videos make handcraft feature isn't effective enough, since there are no fixed descriptors can describe all video cases. On the other hand, traditional clustering algorithms are easily influenced by the choice of initial clustering centers. An Unsupervised feature learning and clustering method for key frame extraction is proposed in this paper, which can be used for human action recognition. Stacked auto-encoder(SAE) is trained using videos from 10 different human actions, SAE is used as a feature extractor to learn features representing human actions. Affinity Propagation Clustering algorithm is used to select key frames from video sequences. We use a variety of videos to do the experiments. Experiments demonstrate that our method can be effectively summarizing video shots considering different human actions.

Recently, deep learning methods have been extensively applied for action recognition in videos. Most existing deep networks equally treat every video frame and directly assign a video label to all the frames sampled from it. However, discriminative action may occurs sparsely in a few key frames in a video, and other frames are less relevant or even irrelevant to the action class. Equally treating all the frames will hurt performance. To address this issue, we propose a temporal attention model which learns to recognize human actions in videos while focusing selectively on the informative frames. Our model does not need explicit annotations regarding such informative frames during training and testing. Specifically, we adopt Recurrent Neural Network (RNN) with Long Short-Term Memory (LSTM) unit and attaches higher importance to the frames which are discriminative for the task at hand. Our method consistently improves on no-attention methods, with both RGB and optical flow based deep ConvNets. We achieve state-of-the-art performance on two challenging datasets of UCF101 and HMDB51.

We propose a robust and efficient method for accurate detecting and localizing complex human action in video in space and time dimensions using spatio-temporal templates. A simple but effective motion descriptor based on the motion-compensated frame difference is designed for template representation, which is resistant to the deformation of posture and cluttered and moving background. A multi-step filtering scheme is adopted to speed up the target candidates localization and matching to the templates. For the template sequence to video registration, we present an extended continuous dynamic programming technique which can compute the matching scores for multiple trajectories simultaneously. Extensive experimental results on different videos have demonstrated the effectiveness of the proposed method.

Action recognition from videos has become one of the coveted fields in recent years because of its potential applications. In particular, action recognition methods utilizing deep learning are found to be effective. In this paper we use spatio-temporal context of videos to recognize the actions in videos using motion history images. The effectiveness of the proposed approach is demonstrated on a new dataset with typical payment related actions.

How to recognize human action from videos captured by modern cameras efficiently and effectively is a challenge in real applications. Traditional methods which need professional analysts are facing a bottleneck because of their shortcomings. To cope with the disadvantage, methods based on computer vision techniques, without or with only a few human interventions, have been proposed to analyse human actions in videos automatically. This paper provides a method combining the three dimensional Scale Invariant Feature Transform (SIFT) detector and the Latent Dirichlet Allocation (LDA) model for human motion analysis. To represent videos effectively and robustly, we extract the 3D SIFT descriptor around each interest point, which is sampled densely from 3D Space-time video volumes. After obtaining the representation of each video frame, the LDA model is adopted to discover the underlying structure-the categorization of human actions in the collection of videos. Public available standard datasets are used to test our method. The concluding part discusses the research challenges and future directions.

Recurrent neural network (RNN) has proven successful recently in action recognition. However, depth sequences are of high dimensionality and contain rich human dynamics, which makes traditional RNNs difficult to capture complex action information. This paper addresses the problem of human action recognition from sequences of depth maps using sequential deep learning. The proposed method first synthesizes multi-view depth sequences by rotating 3D point clouds from depth maps. Each depth sequence is then split into short-term temporal segments. For each segment, a multi-view depth motion template (MVDMT), which compresses the segment to a motion template, is constructed for short-term multi-view action representation. The MVDMT effectively characterizes the multi-view appearance and motion patterns within a short-term duration. Convolutional Neural Network (CNN) models are leveraged to extract features from MVDMT, and a CNN-RNN network is subsequently employed to learn an effective representation for sequential patterns of the multi-view depth sequence. The proposed multi-view sequential deep learning framework can simultaneously capture spatial-temporal appearance and motion features in the depth sequence. The proposed method has been evaluated on the MSR Action3D and MSR Action Pairs datasets, achieving promising results compared with the state-of-the-art methods based on depth data.

Understanding semantic meaning of human actions captured in unconstrained environments has broad applications in fields ranging from patient monitoring, human-computer interaction, to surveillance systems. However, while great progresses have been achieved on automatic human action detection and recognition in videos that are captured in controlled/constrained environments, most existing approaches perform unsatisfactorily on videos with uncontrolled/unconstrained conditions (e.g., significant camera motion, background clutter, scaling, and light conditions). To address this issue, the authors propose a robust human action detection and recognition framework that works effectively on videos taken in controlled or uncontrolled environments. Specifically, the authors integrate the optical flow field and Harris3D corner detector to generate a new spatial-temporal information representation for each video sequence, from which the general Gaussian mixture model (GMM) is learned. All the mean vectors of the Gaussian components in the generated GMM model are concatenated to create the GMM supervector for video action recognition. They build a boosting classifier based on a set of sparse representation classifiers and hamming distance classifiers to improve the accuracy of action recognition. The experimental results on two broadly used public data sets, KTH and UCF YouTube Action, show that the proposed framework outperforms the other state-of-the-art approaches on both action detection and recognition.

Action recognition is an important precursor for understanding human activities in videos. The current paradigm of action recognition is to classify a video sequence as a whole. However, actions usually occur only in part of a video sequence, rendering the rest of the video irrelevant for action recognition. In this paper, we propose a method for learning a subsequence classifier which can detect and classify part of a video that corresponds to the action. The subsequence classifier is trained from weakly labeled training videos whose subsequence labels are not provided, but need to be inferred during learning. We use the framework of multiple instance learning to solve two problems jointly: i) find the action subsequences in training videos, ii) train the subsequence classifier using the inferred action subsequences. To obtain a robust solution to the MIL problem, we propose a sequential algorithm that consecutively decreases the number of inferred action subsequences per video and trims their length until only one short subsequence is used as the action representative in each video. We evaluate the combination of the automatically trained subsequence classifier and the full sequence classifier on the very challenging Hollywood2 benchmark set and observe a significant gain in the performance over the baseline full sequence classifier. Moreover, a favorable performance of the subsequence classifier for temporal localization of actions in videos is evidenced on two categories of the Hollywood2 dataset.