Search

We present a new method for action recognition that employs the co-occurrence of concept features as semantic geometric context. Firstly, the semantic concept codebook is learnt by an improved subspace clustering, then the spatio-temporal interest points are labelled as meaningful features, namely concept features. After that, Multi-scale co-occurrence statistics that embeds the relative distance and direction of pairwise concept features is constructed. Unlike the traditional k-means, the features labelled by the concept codebook well represent the ingredients of objects and ensure temporal consistency. Moreover, the relative layout is the semantic geometric context that describes the changes of geometric relationships. Using the popular KTH and UCF-sports action datasets, experimental results show that the relative layouts combined with the STIPs have discriminative power for action recognition. Our method obtains promising recognition performance compared with other state-of-the-art algorithms.

We propose a tracking method using both overall information of object and local representation.The object is divided into m overlapped patches to improve tracking performance in background clutter and partial occlusions.We take the prior information into account for avoiding ambiguity in discriminative model.The dictionary is updated by Metropolis-Hastings to capture the appearance changes.We achieve comparable performances with other methods on challenging sequences. In this paper, we propose a novel visual tracking algorithm using the collaboration of generative and discriminative trackers under the particle filter framework. Each particle denotes a single task, and we encode all the tasks simultaneously in a structured multi-task learning manner. Then, we implement generative and discriminative trackers, respectively. The discriminative tracker considers the overall information of object to represent the object appearance; while the generative tracker takes the local information of object into account for handling partial occlusions. Therefore, two models are complementary during the tracking. Furthermore, we design an effective dictionary updating mechanism. The dictionary is composed of fixed and variational parts. The variational parts are progressively updated using Metropolis-Hastings strategy. Experiments on different challenging video sequences demonstrate that the proposed tracker performs favorably against several state-of-the-art trackers.

Topic models have been wildly applied in the field of computer vision, through which superior performance was yielded in various recognizing tasks. Among them, probabilistic latent semantic analysis model has earned much attention due to its simplicity and effect. But the affection of encoding and normalization methods on topic models has been ignored during the period. This paper explores the impact of encoding methods combined with different normalization on probabilistic latent semantic analysis model in the context of action classification in videos. Detailed experiments are conducted on KTH and UT-interaction datasets. The results show that an appropriate combination of encoding and normalization methods could significantly improve the performance of probabilistic latent semantic analysis model. The recognition accuracy reachs 96.44% and 93.33% on UT-interaction set1 and set2 respectively, which outperforms the state-of-the-art. Especially, we obtain 94.24% on UT-interaction set1 using sparse STIPs.

Trajectories extracted by previous methods for human action recognition contain irrelevant changes, and the Orientation-Magnitude descriptors of their shapes lack the robustness to camera motion. To solve these problems, action recognition by tracking salient relative motion points is proposed in this paper. Firstly, motion boundary detector which suppresses the camera constant motion is utilized to extract motion features. After processing the detected boundaries by the adaptive threshold, the super-pixels that contain salient points are defined as relative motion regions. Then tracking the points within super-pixels is to generate trajectories. For the trajectory shape, the pre-defined orientation assignments with coarse-to-fine quantization levels are used to produce orientation statistics. Finally, the descriptors of oriented gradient, motion boundary, oriented statistic and their combination are adopted to represent action videos, respectively. On the benchmark KTH and UCF-sports action datasets, experimental results show that the extracted trajectories can describe the movement process of object. Compared with the conventional algorithms, our method with multiple kernel learning obtains good performance.

Human action recognition is at the core of computer vision, and has great application value in intelligent human-computer interactions. On the basis of Bag-of-Words (BoW), this work presents a Huffman coding and Implicit Action Model (IAM) combined framework for action recognition. Specifically, Huffman coding, which outperforms naive Bayesian method, is a robust estimation of visual words' conditional probabilities; whereas IAM captures the spatio-temporal relationships of local features and outperforms most other common machine learning methods. Spatio-Temporal Interest Points (STIPs) and Harris corners are employed as local features, and multichannel feature description is adopted to exploit the complementarity among different features. Experiments on UCF-YouTube and HOHA2 datasets systematically compare the performance of various feature channels and machine learning methods, demonstrating the effectiveness of the approaches proposed by this paper. Finally, multiple augment mechanisms such as feature fusion, hierarchical codebooks and sparse coding are integrated into the recognition system, achieving the best ever performance comparing with the state-of-the-art.

Handling appearance variations caused by the occlusion or abrupt motion is a challenging task for visual tracking. In this paper, we propose a novel tracking method that deals with the appearance changes based on sparse representation in a particle swarm optimization (PSO) framework. First, we divide each candidate state into multiple structural patches to cope with the partial occlusions of the object. Once the object is lost, we present an object’s recovery scheme by the scale invariant feature transforms (SIFT) correspondence between two frames to reacquire the rough object position. Then the tracking state is searched in the vicinage of the rough object position using the PSO iteration. In addition, an online dictionary updating mechanism is presented to capture the object appearance variations. The object information from the initial frame is never updated in the tracking, while other templates in the dictionary are progressively updated based on the coefficients of templates. Compared with several conventional trackers, the experimental results demonstrate that our approach is more robust in dealing with the occlusions and abrupt motion variations.

Sparse representation has been shown competitive performance on single object tracking. In this paper, we extend this technique to tracking multiple interactive objects and present a novel sparse tracker under the tracking-by-detection framework, with saliency detector for objects detection and sparse representation for objects association. Furthermore, we propose an online dictionary learning scheme to capture appearance variations of objects. To avoid using trivial templates, the dictionary contains not only objects templates, but also background information, resulting in more robust estimation. The experiments demonstrate that our approach achieves favorable performance over state-of-the-art algorithms.