Deep Learning Video Action Recognition Method Based on Key Frame Algorithm

Action recognition in still images is closely related to various other computer vision tasks such as pose estimation, object recognition, image retrieval, video action recognition and frame tagging in videos. This problem is focused on recognizing a person's action or behavior using a single frame. Unlike action recognition in videos - a relatively very well established area of research where spatio-temporal features are used, these are not available for still images, making the problem more challenging. In the present work only actions that involve objects are considered. A complex action is broken down into components based on semantics. The importance of each of these components in action recognition is systematically studied.

Every year, new Convolutional Neural Network (CNN) architectures appear to deal with different problems in the activity of image and video recognition. These architectures usually work along the ImageNet dataset for looking for the best performance of the CNNs without taking into account the video task where they are used. This can represent a problem if the task is Human Action Recognition (HAR) in video, since the CNN architectures are pre-trained with an image dataset that can practically contain any object, while HAR problem requires consecutive frames of people doing actions. To prove the idea that using CNNs pre-trained on an image dataset does not always achieve the best performance on a video dataset and that, therefore, it is worth comparing the performance of different CNNs under similar circumstances for the HAR problem, this work proposes an analysis between eight different CNN architectures. Each one of the CNN was exclusively trained with RGB images, which were extracted from the frames of the different classes of videos of HMDB51 dataset. To make the classification of an activity in video, we average the predictions taking into account the successes. We also made some ensembles with the best performance CNNs to measure the improvement in accuracy. Our results suggest that Xception is a strong baseline model that could be used by the community to make their comparisons of their proposals more robust.

The two-stream deep neural networks learn contour evolution information of human activities from image frames and motion information from information evolution between image frames. Optical flow image is one of the commonly-used methods to depict the motion information between frames. The existing optical flow images are usually obtained by off-line calculations. In this paper, a deep neural network named FlowNet2.0 is used to generate optical flow images, and based on which a compound deep neural network is proposed to merge spatial and temporal information to complete the recognition of human activities from videos. Specifically, the proposed compound deep neural network is composed of two sub-networks, static-data-stream learning network and dynamic-data-stream learning network. The former is applied to extract spatial evolution information from RGB images, and the latter generates optical flow prediction images from RGB image sequence first, and extracts temporal evolution information from optical flow images. Finally, the results of two sub-networks are combined under the fusion algorithm to achieve human activities recognition. Experimental results show that the compound deep neural network proposed in this paper can effectively identify human activities from video sequences, and the average classification accuracy on data set UCF101 can reach 83.35%.

Existing RGB and CNN-based methods in video action recognition mostly do not distinguish human body from the environment, thus easily overfit the scenes and objects of training sets. In this work, we present a conceptually simple, general and high-performance framework for action recognition in videos, aiming at person-centric modeling. The method, called Action Machine, is based on person bounding boxes for instance-level action analysis. It extends the Inflated 3D ConvNet (I3D) by adding a branch for human pose estimation and a 2D CNN for pose-based action recognition. Action Machine can benefit from the multi-task training of action recognition and pose estimation, the fusion of predictions from RGB images and poses. Experiments results are provided on trimmed video action datasets, NTU RGB+D, Northwestern UCLA Multiview Action3D, MSR Daily Activity3D. Action Machine achieves superior performance and generalizes well across datasets.

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.

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 in videos is currently in the focus of scientific research due to improvements made in automatic analysis of static images and greater availability of processing power. The paper provides an overview of the key models and methods for action recognition that comprise human models and methods based on estimation of joint trajectories, silhouettes and template matching and spatio-temporal local descriptors. To deal with compound actions and activities, action semantic models are proposed with help of expert knowledge. Since the action recognition task is domain dependent, the methods and models are built and tested on domain specific databases. The paper provides an overview and description of recent video datasets that were created for developing action recognition methods, with an emphasis on datasets with additional modalities such as depth images or accelerometer data.

The task of action recognition in dark videos is useful in various scenarios, e.g., night surveillance and self-driving at night. Though progress has been made in action recognition task for videos in normal illumination, few have studied action recognition in the dark, partly due to the lack of sufficient datasets for such a task. In this paper, we explored the task of action recognition in dark videos. We bridge the gap of the lack of data by collecting a new dataset: the Action Recognition in the Dark (ARID) dataset. It consists of 3,784 video clips with 11 action categories. To the best of our knowledge, it is the first dataset focused on human actions in dark videos. To gain further understanding of our ARID dataset, we analyze our dataset in detail and showed its necessity over synthetic dark videos. Additionally, we benchmark the performance of current action recognition models on our dataset and explored potential methods for increasing their performances. We show that current action recognition models and frame enhancement methods may not be effective solutions for the task of action recognition in dark videos (data available at https://xuyu0010.github.io/arid).

3D video, which consists of a sequence of mesh models, can reproduce dynamic scenes containing 3D information. To summarize 3D video, a key frame extraction method is developed using rate-distortion (R-D) trade-off. For this purpose, an effective feature vector is extracted for each frame. Shot detection is performed using the feature vectors as a preprocessing followed by key frame extraction. Simple but reasonable definitions of rate and distortion are presented. Based on an assumption of linearity, an R-D curve is generated in each shot, where the locations of the key frames are optimized. Finally, R-D trade-off can be achieved by optimizing a cost function using a Lagrange multiplier, where the number of key frames is optimized in each shot. Therefore, our system will automatically determine the best locations and the number of key frames in the sense of R-D trade-off. Our experimental results show the extracted key frames are compact and faithful to the original 3D video.

International audience

We propose a set of kinematic features that are derived from the optical flow for human action recognition in videos. The set of kinematic features includes divergence, vorticity, symmetric and antisymmetric flow fields, second and third principal invariants of flow gradient and rate of strain tensor, and third principal invariant of rate of rotation tensor. Each kinematic feature, when computed from the optical flow of a sequence of images, gives rise to a spatiotemporal pattern. It is then assumed that the representative dynamics of the optical flow are captured by these spatiotemporal patterns in the form of dominant kinematic trends or kinematic modes. These kinematic modes are computed by performing Principal Component Analysis (PCA) on the spatiotemporal volumes of the kinematic features. For classification, we propose the use of multiple instance learning (MIL) in which each action video is represented by a bag of kinematic modes. Each video is then embedded into a kinematic-mode-based feature space and the coordinates of the video in that space are used for classification using the nearest neighbor algorithm. The qualitative and quantitative results are reported on the benchmark data sets.

Activity recognition in videos in a deep-learning setting---or otherwise---uses both static and pre-computed motion components. The method of combining the two components, whilst keeping the burden on the deep network less, still remains uninvestigated. Moreover, it is not clear what the level of contribution of individual components is, and how to control the contribution. In this work, we use a combination of CNN-generated static features and motion features in the form of motion tubes. We propose three schemas for combining static and motion components: based on a variance ratio, principal components, and Cholesky decomposition. The Cholesky decomposition based method allows the control of contributions. The ratio given by variance analysis of static and motion features match well with the experimental optimal ratio used in the Cholesky decomposition based method. The resulting activity recognition system is better or on par with existing state-of-the-art when tested with three popular datasets. The findings also enable us to characterize a dataset with respect to its richness in motion information.

In this paper, a generalized statistical tool is introduced to estimate key frames in a video sequence. The tool works based on the inter-relationship between different features of image frames in a video. The image feature vectors are plotted in feature space as points and a randomness measure is determined from the distribution of these points. The randomness measure of the feature vectors is defined with respect to simulated random point patterns and expressed as a probability value of a frame being a key frame. Since, depending on the video content more than one inter-relationship of features can be used to determine a single key frame, different probability values are derived to support a frame as a key frame. To integrate these probability values a combiner model is designed to uniquely decide the status of a key frame. The combiner model is based on the Dempster-Shafer theory of evidence. To demonstrate the idea, randomness measures, and consequently the probabilities of a frame to be a key frame, are obtained separately from spatial domain and frequency domain features. The combined probability value enhances the confidence in selecting a frame as a key frame. The result is tested on a number of standard video sequences and it outperforms the related approach

This paper focuses on human action recognition in video sequences. A method based on the optical flow estimation is presented, where critical points of the flow field are extracted. Multi-scale trajectories are generated from those points and are characterized in the frequency domain. Finally, a sequence is described by fusing this frequency information with motion orientation and shape information. Experiments show that this method has recognition rates among the highest in the state of the art on the KTH dataset. Contrary to recent dense sampling strategies, the proposed method only requires critical points of motion flow field, thus permitting a lower computation time and a better sequence description. Results and perspectives are then discussed.

In the traditional two-stream convolutional neural network, RGB image and optical flow image are trained separately and then simply fused, which results in the lack of fine-grained interaction between static image and dynamic clue, making the important clue of timing information unable to play its full role in the task of video action recognition. We propose a spatial-temporal co-attention network, which allows the attention of spatial features and temporal features to guide each other and collect basic information from the two features. The network utilizes the attention mechanism to estimate the commonality between static images and temporal information by focusing on their shared common semantics. Our experiments verify the effectiveness of the spatial-temporal co- attention network in combining spatial information and temporal information, and obtain better performance on three public standard datasets.