Zhiwei Liang

Integration of semantic and visual hashing for image retrieval

Semantic hashing codes are learned by leveraging the similarity between textual structure and visual structure.Maximum entropy principle is exploited to achieve compact binary codes.Function decay principle is introduced to remove noisy visual attributes. With the rapid proliferation of large-scale web images, recent years have witnessed more and more images labeled with user-provided tags, which leads to considerable effort made on hashing based image retrieval in huge databases. Current research efforts focus mostly on learning semantic hashing functions which design compact binary codes to map semantically similar images into similar codes; however the visual similarity is not well explored for constructing semantic hashing functions. Here a novel approach is proposed to learn hashing functions that preserve semantic and visual similarity between images. Specifically, semantic hashing codes are first learned by leveraging the similarity between textual structure and visual structure; then, the maximum entropy principle is exploited to achieve compact binary codes; finally, the function decay principle is introduced to remove noisy visual attributes. Experimental results conducted on a widely-used image dataset demonstrate the superior performance of the proposed method over the examined state-of-the-art techniques.

Tracklet association for object tracking

This paper proposes a novel multi-target tracking framework, where two different association strategies are utilized to obtain local and global tracking trajectories. Specifically, a scene self-adaptive model is first utilized to generate local trajectories by constructing the association between detection responses and tracking tracklets; then, a novel incremental linear discriminative appearance model is utilized to generate global trajectories by constructing the association between local trajectories; finally, a non-linear motion model is utilized to fill the vacancies between global trajectories to obtain continuous and smooth tracking trajectories. Experimental results conducted on PETS 2009/2010 and TUD-Stadtmitte database demonstrate the proposed framework can achieve continuous and smooth tracking trajectories under the case of significant deformation, appearance change, similar appearance, motion direction change, and long-time occlusion.

Optical flow and spatio-temporal gradient based abnormal behavior detection

To improve the accuracy of the detection of local abnormal behavior, a novel method is here proposed. The main idea of the proposed method is described as follows: firstly, a video sequence is divided into spatio-temporal blobs; then, a statistical method based on the semi-parametric model is adopted to detect these blobs where abnormal behaviors most likely to appear; finally, maximum optical flow energy and local nearest descriptor are utilized to determinate whether these suspicious blobs really contain abnormal behaviors. The experimental results conducted on UCSD dataset demonstrate the effectiveness of the proposed method.

Kicking motion design of humanoid robots using gradual accumulation learning method based on Q-learning

This paper manly presented kicking design motion of humanoid robots using a reinforcement learning method which is based on the Q-learning. First, this method build a multidirectional fixed-point kicking model, which is based on the offset of kicking point, the foot space motion trajectory and ZMP stability criterion, and that makes subsequent train costs much less time. Besides, discretization of state set is also used to improve the training method. Compared to other machine learning algorithms, this method reduces the dimension of the system and solves the problem of excessive train when kicking in long distance. A series of experiments proves that the method described in this paper is feasible and effective.

Fusing Canny operator with vibe algorithm for target detection

Due to the complexity of human motion, the target detection results by using traditional ViBe algorithm are not so satisfactory. Therefore, this paper proposes a method to deal with the target detection issue by fusing an improved Canny operator with Vibe algorithm. Specifically, the ViBe algorithm is utilized to achieve the initial foreground region of a moving object; then, the improved Canny operator is applied to extract the edge information of a moving object; finally, the extracted foreground region and edge information are fused to obtain more accurate foreground region. The experimental results performed on KTH human behavior database demonstrate the effectiveness of the proposed scheme.

Kicking motion planning of Nao robots based on CMA-ES

A kicking design motion of humanoid robots is presented in this paper. This kicking design motion uses a gradual accumulation learning method which is based on the Covariance Matrix Adaptation Evolution Strategy (CMA-ES). By planning the best kicking point and the foot space motion trajectory, the first layer of learning optimization can be realized using the linear distance after kicking and the time cost about kicking point as the target. Then, the optimization of the next layer was fulfilled by employing the double balancing mechanism of the robots center of the gravity and the gyroscope sensor feedback. The learning goal was that the football contact point selection, the weighted penalty of the ankle joint and the performance of kicking were overall considered. The effectiveness of the proposed design method has been revealed in this paper through experimental results.

Local passing-ball tactics based on a Keepaway algorithm

This paper proposes a passing method based on reinforcement learning to realize multi.agent cooperation strategy under the platform of Keepaway. In this paper, a Keepaway model is built in Robocup3D simulation game using the Sarsa(λ) algorithm based on a linear function approximation. We set the time to keep the ball as the reward, and keep training to get the action, making the team hold the ball as long as possible. In this training model, the strategy of soccer robots is implemented in team collaboration mechanism, include PASS and GETOPEN strategy. The experimental results demonstrate the effectiveness of our methods.

Image classification via multi-view model

With the massive growth of digital image data uploaded to the Internet, classifying each image into appropriate semantic category based on the image content for image index and image retrieval has become an increasingly difficult and laborious task. To deal with this issue, we propose a novel multi-view semi-supervised learning framework which leverages the information contained in pseudo-labeled images to improve the prediction performance of image classification using multiple views of an image. In the training process, labeled images are first adopted to train view-specific classifiers independently using uncorrelated and sufficient views, and each view-specific classifier is then interactively re-trained using initial labeled samples and additional pseudo-labeled samples based on a measure of confidence. In the classification process, the maximum entropy principle is utilized to assign appropriate category label to each unlabeled image using optimally trained view-specific classifiers. Experimental results on a general-purpose image database demonstrate the effectiveness and efficiency of the proposed multi-view semi-supervised scheme.

Target detection via statistical model learning

This paper presents a statistical treatment of background modeling for use in target detection, where the global information and local information is added into the statistical framework to construct a robust background model to achieve accurate object detection results. Specifically, a novel self-adaptive Gaussian mixture model is proposed to construct a statistical background model based on the global information, which is utilized to deal with the target detection issue under illumination changes; for the target detection issue under dynamic background, the self-tuning spectral clustering technology is first utilized to cluster the background image, the kernel density estimation method is then utilized to construct a statistical background model based on the local information. Experimental results demonstrate that the proposed algorithm can improve the detection performance under illumination changes or dynamic background.

Information sharing method for RCRSS in communication-limited environment

Information sharing is the basis upon which agents share their local views and coordinate with each other in Multi-Agent System (MAS). This paper presents an information sharing method for RoboCup Rescue Simulation System (RCRSS) in communication-limited environment. This method has two parts: within a partition, message sending is triggered by events while among partitions by time. Simulation results in RCRSS show that this information sharing method is valid and improves rescue teams performance in communication-limited environment.