Yuexian Zou

Mobile Human Airbag System for Fall Protection Using MEMS Sensors and Embedded SVM Classifier

This paper introduces a mobile human airbag system designed for fall protection for the elderly. A Micro Inertial Measurement Unit ( muIMU) of 56 mm times 23 mm times 15 mm in size is built. This unit consists of three dimensional MEMS accelerometers, gyroscopes, a Bluetooth module and a Micro Controller Unit (MCU). It records human motion information, and, through the analysis of falls using a high-speed camera, a lateral fall can be determined by gyro threshold. A human motion database that includes falls and other normal motions (walking, running, etc.) is set up. Using a support vector machine (SVM) training process, we can classify falls and other normal motions successfully with a SVM filter. Based on the SVM filter, an embedded digital signal processing (DSP) system is developed for real-time fall detection. In addition, a smart mechanical airbag deployment system is finalized. The response time for the mechanical trigger is 0.133 s, which allows enough time for compressed air to be released before a person falls to the ground. The integrated system is tested and the feasibility of the airbag system for real-time fall protection is demonstrated.

Towards HMM based human motion recognition using MEMS inertial sensors

This paper presents a new method of human motion recognition based on MEMS inertial sensors data. A Micro Inertial Measurement Unit (μIMU) that is 56mm*23mm*15mm in size was built. This unit consists of three dimensional MEMS accelerometers, gyroscopes, a Bluetooth module and a MCU (Micro Controller Unit), which can record and transfer inertial data to a computer through serial port wirelessly. Five categories of human motion were done including walking, running, going upstairs, fall and standing. Fourier analysis was used to extract the feature from the human motion data. The concentrated information was finally used to categorize the human motions through HMM (Hidden Markov Model) process. Experimental results show that for the given 5 human motions, correct recognition rate range from 90%–100%. Also, a full combination of 6 parameters (Gx, Gy, Gz, Ax, Ay, Az) was listed and the recognition rate of each combination (total 63) was tested.

Video Image Vehicle Detection System for Signaled Traffic Intersection

In modern intelligent transportation systems, the video image vehicle detection system (VIVDS) is gradually becoming one of the popular methods at signalized traffic intersection due to its convenient installation and rich information content provided. However, in the current VIVDS, the camera usually is installed at the roadside poles or traffic light poles, which not only requires more than one camera to cover the entire intersection, but also results in serious vehicle occlusions and adverse affects on the performance of the vehicle detection and tracking. Meanwhile, it is noted that the detection rate of the black, gray and dark color vehicles (such as red, blue, and green vehicles) are poor or incomplete detection by using the traditional background subtraction method in the RGB color model. To tackle these problems, this paper presents a novel VIVDS with the new camera installation, which only uses a single camera to cover the panorama view of the interested intersection. Furthermore, a robust vehicle detection algorithm with multi-information fusion has been developed to resolve problems of detecting incompletion, which plays a key role in enhancing the vehicle detection rate in the proposed VIVDS for urban traffic surveillance. The proposed system has been tested on a traffic image sequences recorded at typical urban intersections. The experimental results show that the system offers the flexibility to detect the different color vehicles, the robustness to noise and the efficiency of computation.

Traffic incident classification at intersections based on image sequences by HMM/SVM classifiers

With the development of modern intelligent transportation systems (ITS), automatic traffic incident detection with quick response and high accuracy becomes one of the most important issues, especially for metropolitan streets that are full of signaled intersections. In this paper, we present our up-to-date research outcomes of the traffic incident detection system, which makes use of the image sequences gathered from a typical urban intersection. Basic image signal processing was used to extract image difference information for traffic image database construction. Feature extraction algorithms were then discussed and compared including PCA, FFT, and hybrid analysis of DCT-FFT. Finally, multi-classification of traffic signal logics (East–West, West–East, South–North, North–South) and accidents were realized by HMM (Hidden Markov Model) and SVM (Support Vector Machine) respectively. Experimental results showed that the hybrid DCT-FFT method gives the best features, and classification performance of SVM is superior to HMM with limited training samples, where the correction rate is 100% for SVM and 91% for HMM.

A hybrid convolutional neural networks with extreme learning machine for WCE image classification

Wireless Capsule Endoscopy (WCE) is considered as a promising technology for non-invasive gastrointestinal disease examination. This paper studies the classification problem of the digestive organs for wireless capsule endoscopy (WCE) images aiming at saving the review time of doctors. Our previous study has proved the Convolutional Neural Networks (CNN)-based WCE classification system is able to achieve 95% classification accuracy in average, but it is difficult to further improve the classification accuracy owing to the variations of individuals and the complex digestive tract circumstance. Research shows that there are two possible approaches to improve classification accuracy: to extract more discriminative image features and to employ a more powerful classifier. In this paper, we propose to design a WCE classification system by a hybrid CNN with Extreme Learning Machine (ELM). In our approach, we construct the CNN as a data-driven feature extractor and the cascaded ELM as a strong classifier instead of the conventional used full-connection classifier in deep CNN classification system. Moreover, to improve the convergence and classification capability of ELM under supervision manner, a new initialization is employed. Our developed WCE image classification system is named as HCNN-NELM. With about 1 million real WCE images (25 examinations), intensive experiments are conducted to evaluate its performance. Results illustrate its superior performance compared to traditional classification methods and conventional CNN-based method, where about 97.25% classification accuracy can be achieved in average.

Extraocular image processing for retinal prosthesis based on DSP

Simulations of artificial vision suggest that thousands of electrodes may be required to restore vision for ones with diseases of the outer retina. With the development of MEMS fabrication process for the stimulation electrode array, extraocular image processing is becoming more and more important for the retinal prosthesis systems. A Digital Signal Processor (DSP) based extraocular image processing system (EIPS) for a retinal prosthesis has been developed in this paper. The system mainly consists of a CMOS image sensor and a DSP processing system, which provides the capability of implmenting the real-time image processing with low power consumption. Furthermore, this system offers the flexibility of realizing various image processing algorithms with different specification requirements on the DSP by programming, such as different frame rate, resolution and throughput data rate. The related image processing algorithms include the image resizing, color erasing, edge enhancement and edge detection. Finally, the speed of different DSPs in the market has been evaluated and compared for achieving better performance.

Recursive robust variable loading mvdr beamforming in impulsive noise environment

This paper proposes a new recursive variable loading minimum variance distortionless response (RVL-MVDR) algorithm for robust beamforming in impulsive noise environment. It employs a new method for robust estimation of the sample covariance matrix under impulsive noise and a new method for computing the data-dependent loading level using robust statistics. Computer simulations suggest that the proposed algorithm performs considerably better than the conventional sample matrix inversion (SMI)-MVDR and variable loading (VL) MVDR algorithms in impulsive noise environment.

A robust traffic parameter extraction method using texture and entropy

This paper presents a robust traffic parameters extraction (RTPE) method for intelligent traffic system. Firstly a texture-based algorithm is introduced to solve the moving shadow problem, which occurs in traffic lane commonly. Secondly, we propose a robust exponential entropy-based and data-dependent threshold vehicle detection algorithm, named RVD-EXEN algorithm to extract vehicles feature from raw visual information for vehicle detection. On this basis, we calculate some basic traffic parameters such as traffic flow, time occupancy ratio and space mean speed. The experiments show that proposed RTPE method has the flexibility to shadow situation, robustness to noise and efficiency of computation.

Fast Extraction of Somatosensory Evoked Potential Using RLS Adaptive Filter Algorithms

This paper evaluates the efficacy of the recursive least squares (RLS) in adaptive noise canceller (RLS-ANC) for fast extraction of somatosensory evoked potentials (SEPs). The RLSANC method was verified by simulation of electroencephalography (EEG) and Gaussian noise contaminated SEP signals at different signal-to-noise ratios (SNRs). RLS was found to converge faster than the least mean squares (LMS) algorithm in ANC, i.e. SEP extraction by RLS-ANC required fewer trials than LMS-ANC. Experimental results showed that RLS-ANC with less than 50 trials could provide similar performance in SEP extraction to those extracted by the conventional ensemble averaging with 500 trials even at SNR of 20dB. Keywords-SEP; Adaptive Filter; RLS-ANC; LMS-ANC

A Robust Traffic Parameter Extraction Approach for Surveillance System at Urban Intersection

For a typical urban intersection, moving vehicle shadow and vehicle-pedestrian mixed conditions exist in traffic scene commonly. These interfering factors lead to a very low correct rate of the traffic parameters extraction. This paper presents robust traffic parameters extraction (RTPE) approach for traffic surveillance system at an urban intersection, which contains three key algorithms. First, a texture-based vehicle segmentation (TVS) algorithm is introduced to solve the moving shadow problem. Second, we propose an image exponential entropy-based vehicle exist detection (IEE-VED) algorithm to reduce the noise interference by pedestrians at the intersection, and we extract vehicle features from raw visual information to determine whether there is a vehicle in the detection zone. On this basis, the traffic parameters measurement (TPM) algorithm is introduced to calculate some important traffic parameters of the intersection for traffic management and traffic jam detection, such as traffic flow, time occupancy ratio, space mean speed and the difference of IN/OUT traffic flow. Experimental results indicate that the proposed RTPE approach is effective for traffic parameters extraction, and these parameters can truly reflect the prevailing traffic condition.