Yaobin Mao

Recognition of human activities using SVM multi-class classifier

Even great efforts have been made for decades, the recognition of human activities is still an unmature technology that attracted plenty of people in computer vision. In this paper, a system framework is presented to recognize multiple kinds of activities from videos by an SVM multi-class classifier with a binary tree architecture. The framework is composed of three functionally cascaded modules: (a) detecting and locating people by non-parameter background subtraction approach, (b) extracting various of features such as local ones from the minimum bounding boxes of human blobs in each frames and a newly defined global one, contour coding of the motion energy image (CCMEI), and (c) recognizing activities of people by SVM multi-class classifier whose structure is determined by a clustering process. The thought of hierarchical classification is introduced and multiple SVMs are aggregated to accomplish the recognition of actions. Each SVM in the multi-class classifier is trained separately to achieve its best classification performance by choosing proper features before they are aggregated. Experimental results both on a home-brewed activity data set and the public Schuldts data set show the perfect identification performance and high robustness of the system.

Home environment fall detection system based on a cascaded multi-SVM classifier

Fall detection system for intelligent home care for elderly people is presented in this paper. The system includes human blob detection by non-parameter background substruction method, feature extraction from two minimum bounding boxes, and fall detection by a cascaded multi-SVM classifier. Besides falling down, other daily activities such as walk, jogging, sitting down, squatting down and immobility are also taken into consideration. A three-stage cascade of SVM classifiers is made to distinguish fall action from other activities. Each SVM classifier is first trained and tested separately to achieve its best classification performance by choosing proper features and corresponding kernel function. Then the combined classifier is trained to detect falls. A perfect correct identification rate of 98.13% on a real activity video set by experiments demonstrates the robustness and the utility of the system.

Robust Real-Time Detection of Abandoned and Removed Objects

This paper presents a robust real-time method for general detection of abandoned and removed objects from surveillance videos. The system introduces a unique combination of a new pixel-wise static region detector and a novel abandoned/removed object classifier based on color richness. In the static region detection phase, two backgrounds are constructed respectively to build foreground and stationary masks which are then used to update a static region confidence map. Static regions are thus extracted from the confidence map and further classified into abandoned or removed items by comparing color richness between the background and current frame. Our algorithm is easy to implement, robust to small repetitive motions, illumination change and can handle object occlusion. Experimental results on two public video databases which are shot in different scenarios demonstrate the robustness and practicability of the proposed method in real-time video surveillance.

SVM-based abnormal activity detection for home care

Abnormal activity detection for intelligent home care is presented in this paper. The activities have been catalogued into six possible classes, such as standing, sitting, squatting, walking, jogging, and falling down, among which falling down including on-marching falling down and in-place falling down is regarded as the abnormal activity. The output of background subtraction is employed directly to obtain the binary human-body images and only centroid track and figure width of human blob are selected as features for recognition. Activities are sub-divided into moving activities and quasi-static activities in terms of the horizontal movement of the centroid of body blob. And then SVM classifiers are used to recognize respectively the on-marching falling down and in-place falling down from above two classes of behaviors. A home-brewed activity database is obtained and the experimental results are: the correct identification rate is 100 percent for on-marching falling down and the minimum correct identification is above 90 percent for in-place falling down activity.

Cost-Sensitive Boosting: Fitting an Additive Asymmetric Logistic Regression Model

Conventional machine learning algorithms like boosting tend to equally treat misclassification errors that are not adequate to process certain cost-sensitive classification problems such as object detection. Although many cost-sensitive extensions of boosting by directly modifying the weighting strategy of correspond original algorithms have been proposed and reported, they are heuristic in nature and only proved effective by empirical results but lack sound theoretical analysis. This paper develops a framework from a statistical insight that can embody almost all existing cost-sensitive boosting algorithms: fitting an additive asymmetric logistic regression model by stage-wise optimization of certain criterions. Four cost-sensitive versions of boosting algorithms are derived, namely CSDA, CSRA, CSGA and CSLB which respectively correspond to Discrete AdaBoost, Real AdaBoost, Gentle AdaBoost and LogitBoost. Experimental results on the application of face detection have shown the effectiveness of the proposed learning framework in the reduction of the cumulative misclassification cost.

Estimation of crowd density using multi-local features and regression

Crowd analysis is an important issue in intelligent visual surveillance systems. In this paper, a tracking-free solution to crowd density estimation is presented. The method consists of four steps: each motion parts are first extracted from video frames through motion segmentation; then eight kinds of low-level image features including blob area, Harris corner, KLT feature points, contour number, contour perimeter, ratio of perimeter to area, edge and fractal dimension are calculated; to eliminate the errors introduced by perspective effect and occlusion, both geometric correction and overlapping compensation through proper weight assignments are performed; finally, multiple regression model is used to estimate pedestrian numbers. Various experiments are performed on three video data sets and the encouraging results show that the proposed algorithm not only can perform crowd density estimation correctly but can operate in real-time.

Visual tracking by separability-maximum online boosting

Recently, visual tracking has been formulated as a classification problem whose task is detecting the object form the scene with a binary classifier. And online boosting, which adapts the binary classifier to appearance changes by online feature selection, has been investigated by researchers. However, online boosting generally suffers from drifting if the tracking error accumulates. To reduce tracking error, separability-maximum boosting (SMBoost), together with a two stage online boosting paradigm (online SMBoost), is proposed and applied to visual tracking. SMBoost uses a separability based cost function that defined on the statistics. And online boosting is therefore split into two individual stages: online statistics estimating and separability-maximum classifier training. Experiment on UCI machine learning datasets shows that SMBoost is more accurate than batch AdaBoost and its online variation. And benchmark on public sequences indicates that feature selection with online SMBoost is more effective and robust comparing with previous online boosting algorithm. To track a visual object stably, online SMBoost saves more than 50% classifier complexity, and achieves 108 fps.

Multiple Faces Tracking via Statistical Appearance Model

Recently, appearance based methods have become a dominating trend in tracking. For example, tracking-by-detection models a target with an appearance classifier that separates it from the surrounding background. Recent advances in multi-target tracking suggest learning an adaptive appearance affinity measurement for target association. In this paper, statistical appearance model (SAM), which characterizes facial appearance by its statistics, is developed as a novel multiple faces tracking method. A major advantage of SAM is that the statistics is a target-specific and scene-independent representation, which helps for further video annotation and behavior analysis. By sharing the statistical appearance models between different videos, we are able to improve tracking stability on quality-degraded videos.

Visual tracking by separability-maximum boosting

Abstract. Recently, visual tracking has been formulated as a classification problem whose task is to detect the object from the scene with a binary classifier. Boosting based online feature selection methods, which adopt the classifier to appearance changes by choosing the most discriminative features, have been demonstrated to be effective for visual tracking. A major problem of such online feature selection methods is that an inaccurate classifier may give imprecise tracking windows. Tracking error accumulates when the tracker trains the classifier with misaligned samples and finally leads to drifting. Separability-maximum boosting (SMBoost), an alternative form of AdaBoost which characterizes the separability between the object and the scene by their means and covariance matrices, is proposed. SMBoost only needs the means and covariance matrices during training and can be easily adopted to online learning problems by estimating the statistics incrementally. Experiment on UCI machine learning datasets shows that SMBoost is as accurate as offline AdaBoost, and significantly outperforms Oza’s online boosting. Accurate classifier stabilizes the tracker on challenging video sequences. Empirical results also demonstrate improvements in term of tracking precision and speed, comparing ours to those state-of-the-art ones.

An efficient data-scalable algorithm for image orientation detection

Image orientation detection is a useful, yet challenging research topic in intelligent image processing. Existing methods generally train a detector on ensemble data-set which is little scalability when new image samples with novel scenes come. This paper proposes a data-scalable algorithm for image orientation detection using bagging, a method aggregates several classifiers trained independently on non-intersecting sub data sets. By the proposed algorithm, when new classifiers trained on novel data sets are added, the prediction accuracy increases. In the paper, more representative feature set and more efficient learning algorithm are adopted to remedy the possible decrease of detection accuracy caused by the curtailment of the training data for single classifiers. Compared with previous work, the proposed algorithm has great competitiveness in terms of data-scalable ability, prediction accuracy, training and detection complexity.