Search

Monocular 3D object detection is a low cost and challenging task for autonomous vehicles and robotics. Utilizing a monocular image for 3D object detection is served as an auxiliary module for autonomous vehicles and is a growing concern recently. Currently, the expensive lidar and stereo cameras have a predominant performance on accurate 3D object detection, whereas monocular based methods are considerably lower in performance. This performance gap is minimized by reforming the monocular based method as a single internal network. We exploit the correlation between 2D and 3D detection spaces, enabling 3D boxes to leverage feature maps generated in image space. The 2D and 3D proposals are extracted through a proposal generation network that is enhanced and utilized for estimating accurate 3D detection and localization. Experimental results on the KITTI dataset demonstrate that in comparison to other monocular object detection methods the proposed method considerably improved the accuracy of 3D object detection. The mean average precision of 3D object detection in front view is improved to 25% and the bird's eye view to 32% for the car class on the moderate difficulty level.

Background Constraint-induced movement therapy (CIMT) is a prominent neurorehabilitation approach for improving affected upper extremity motor function in children with unilateral cerebral palsy (UCP). However, the restraint of the less-affected upper extremity and intensive training protocol during CIMT may decrease children’s motivation and increase the therapist’s workload and family’s burden. A kinect-based CIMT program, aiming to mitigate the concerns of CIMT, has been developed. The preliminary results demonstrated that this program was child-friendly and feasible for improving upper extremity motor function. However, whether the kinect-based CIMT can achieve better or at least comparable effects to that of traditional CIMT (i.e., therapist-based CIMT) should be further investigated. Therefore, this study aimed to compare the effects of kinect-based CIMT with that of therapist-based CIMT on upper extremity and trunk motor control and on daily motor function in children with UCP. Methods Twenty-nine children with UCP were recruited and randomly allocated to kinect-based CIMT (n = 14) or therapist-based CIMT (n = 15). The intervention dosage was 2.25 h a day, 2 days a week for 8 weeks. Outcome measures, namely upper extremity and trunk motor control and daily motor function, were evaluated before and after 36-h interventions. Upper extremity and trunk motor control were assessed with unimanual reach-to-grasp kinematics, and daily motor function was evaluated with the Revised Pediatric Motor Activity Log. Between-group comparisons of effectiveness on all outcome measures were analyzed by analysis of covariance (α = 0.05). Results The two groups demonstrated similar improvements in upper extremity motor control and daily motor function. In addition, the kinect-based CIMT group demonstrated greater improvements in trunk motor control than the therapist-based CIMT group did (F(1,28) > 4.862, p Conclusion Kinect-based CIMT has effects comparable to that of therapist-based CIMT on UE motor control and daily motor function. Moreover, kinect-based CIMT helps decrease trunk compensation during reaching in children with UCP. Therefore, kinect-based CIMT can be used as an alternative approach to therapist-based CIMT. Trial registration: ClinicalTrials.gov Identifier: NCT02808195. Registered on 2016/06/21, https://clinicaltrials.gov/ct2/show/NCT02808195.

Recent studies demonstrate various methods to classify noises present in daily human activity. Most of these methods utilize multiple audio features that require heavy computation, which increases the latency. This article presents a real-time noise classifier based on a smartphone by utilizing only the mel-frequency cepstral coefficient (MFCC) as the feature vector. By relying on this single feature and an augmented audio dataset, this system drastically reduced the computation complexity and achieved 92.06% accuracy. This system utilizes the TarsosDSP library for feature extraction and convolutional neural network–long short-term memory for both classification and MFCCs determination. The results show that the developed system can classify the noises with higher accuracy and shorter processing time compared with other architectures. Additionally, this system only takes up 6 mAh of power consumption, which makes it suitable for future commercial use.

Object Stereo Vision has conventionally been one of the deeply examined areas in computer vision. Stereo matching is employed in numerous modern applications, including robot navigation, augmented reality, and automotive applications. Even though it has a long research history, it is still challenging for the edges of textureless, discontinues, and occluded regions under radiometric variation. This research article proposes a modified histogram equalization, a novel feature extraction, a spatial gradient model, and matching cost, which is robust and stable to images taken in different radiometric variations. The proposed method reduced the average percentage of bad pixels to 3.35 and reduced the relative mean square error (RMSE) up to 30.08 on the Middlebury dataset for different illumination and exposure values. Quantitative and qualitative evaluation of the proposed method demonstrates significant improvement in increasing PSNR and decreasing bad pixel percentage against radiometric variation and state-of-the-art local stereo matching algorithms.

Over the past decade, significant developments have taken place in the field of deep learning. State-of-the-art convolutional neural networks (CNNs), a branch of deep learning, have been increasingly applied in various fields such as image classification, speech recognition, and natural language processing. Due to the high computational complexity of CNNs, lots of works have proposed their CNN accelerators to address this issue. Besides, a processing element (PE) array has been recently further focused and discussed since it is responsible for the entire computations as the core of CNN accelerators. Therefore, the specialized design of a PE array becomes one of the main researches on CNN accelerators for energy efficiency and high throughput. In this brief, a throughput-optimized PE array for CNNs based on the channel-oriented data pattern is proposed. The proposed PE array features fully PE interconnection which achieves scalability. Besides, any sized convolution can be processed in the PE array while maximizing the utilization of PEs by exploiting the channel-oriented data pattern. Compared to previous works, this brief achieves $1.22\times $ and $1.25\times $ improvement in the throughput density on AlexNet and VGG-16 respectively.

With better experience of immersive scene, the virtual reality (VR) becomes the main stream of entertainments. Featuring high chromatic displaying performance, relatively higher efficiency in power and lower latency, the Active-Matrix Organic Lighting-Emitting Diode (AMOLED) is the best-choice display technology for VR. However, AMOLED still requires the higher power consumption amongst other components on smart phones. Thus, various studies attempted to leverage the property of AMOLED to lower the power consumption while preserving the visual quality. This paper proposes a perceptually saliency-oriented transformation by incorporating the concept of how human eyes work in VR environment to provide more suitable experience while saving more power consumption. Experimental results show the proposed method preserves up to 49.33% of the power demand with high perceptual quality as other existing techniques. Furthermore, the subjective experiment demonstrates the proposed method provides a better perceptual quality for VR environment.

Image/video processing in real-time is always in high demand for the quality of video. There are several factors which cause the loss of the video content, such as the type of transmission, missing data and especially data switch. Data switch generally occurs in the alternation of the video signal, which can cause the discontinuity of data during the video data stored in buffer or memory. The current method which adopts frame difference for detecting this issue may consume many resources and memory footprint. This paper presents a method which uses the video pixel prediction to detect the freezing event. The method is implemented with a video system which employs the System-on-chip (SoC) architecture with Field Programmable Gate Array (FPGA) and other components including DDR3 ram, flash, and exchange interfaces as the main processing platform that prevents this problem through freezing detection. The result of evaluation shows that the accuracy of the proposed method is above 99%, in terms of saving more logic usage and reducing the footprint of the memory on the video system.

The state-of-the-art convolutional neural networks (CNNs) have been widely applied to many deep neural networks models. As the model becomes more accurate, both the number of computation and the data accesses are significantly increased. The proposed design uses the row stationary with network-on-chip and the fast convolution algorithm in process elements to reduce the number of computation and data accesses simultaneously. The experimental evaluation which using the CNN layers of VGG-16 with a batch size of three shows that the proposed design is more energy-efficient than the state-of-the-art work. The proposed design improves the total GOPs of the algorithm by 1.497 times and reduces the on-chip memory and off-chip memory accesses by 1.07 and 1.46 times than prior work, respectively.

The abnormal event detection becomes an important topic recently. This paper presents a method to detect the crowd gathering, as well as the commotion event after the crowd gathering. The proposed stillness model and the motion model are based on the improved background subtraction and the optical flow feature. We construct the long-term stillness level by the break bucket model and clustering the instantaneous stillness level. Then the crowd gathering event is decided by the threshold with the long-term stillness level. Furthermore, the motion model is applied for detecting the commotion event after the crowd gathering. In the experiment, we used the dataset of PET2009. The proposed method is verified by the experiment with 97% accuracy.

Surface ElectroMyoGraphy (sEMG) is widely applied to a variety of applications. Managing the power consumption of battery-constrained sEMG Wireless Body Sensor Networks (WBSN) is an important topic. In this paper, we use fatigue assessments as a case study. We apply the concept of distributed computing to explore the impact of computation allocations on the client power consumption and the requirement of architecture specifications. Regarding the CPU clock rate, we propose a power-saving method based on the ping-pong buffer mechanism and evaluate all the crucial factors which affect the power consumption such as sEMG sample rates, algorithmic computational costs, wireless throughputs, and selection of wireless technologies. To sum up, we conduct a comprehensive analysis of all possible distributed computing architectures of WBSN to determine the lowest-power WBSN architecture. The results show that the implementation based on the lowest-power WBSN architecture has lower power consumption compared with other hardwares. The average current of the proposed architecture can be reduced by 81.7% compared with the previous work. Besides, the battery life is 4.48 times that of the previous work under the continuous wireless connection equipped with the same 300mAh lithium battery. Compared with the commercial device, the battery life is 1.6 times that of the commercial one.

With the advance of deep neural networks (DNNs), artificial intelligence (AI) has been widely applied to various applications in our daily lives. These DNN-based models can be stored in portable storage disks or low-power Neural Compute Sticks. They can then be deployed in edge devices through the USB interface for AI-based applications, such as Automatic Diagnosis Systems or Smart Surveillance Systems, which provides solutions to incorporating AI into the Internet of Things (IoT). In this work, based on our observation and careful analysis, we propose a model-based deep encoding method built upon Huffman coding to compress a DNN model transmitted through the USB interface to edge devices. Based on the proposed lopsidedness estimation approach, we can exploit a modified Huffman coding method to increase the USB transmission efficiency for quantized DNN models while reducing the computational cost entailed by the coding process. We conducted experiments on several benchmarking DNN models compressed using three emerging quantization techniques, which indicates that our method can achieve a high compression ratio of 88.72%, with 93.76% of the stuffing bits saved on average.

In recent years, the surface electromyography (EMG) signal has received a lot of attention. EMG signals are used to analyze muscle activity or to evaluate a patient’s muscle status. However, commercial surface EMG systems are expensive and have high power consumption. Therefore, the purpose of this paper is to implement a surface EMG acquisition system that supports high sampling and ultra-low power consumption measurement. This work analyzes and optimizes each part of the EMG acquisition circuit and combines an MCU with BLE. Regarding the MCU power saving method, the system uses two different frequency MCU clock sources and we proposed a ping-pong buffer as the memory architecture to achieve the best power saving effect. The measured surface EMG signal samples can be forwarded immediately to the host for further processing and additional application. The results show that the average current of the proposed architecture can be reduced by 92.72% compared with commercial devices, and the battery life is 9.057 times longer. In addition, the correlation coefficients were up to 99.5%, which represents a high relative agreement between the commercial and the proposed system.

Surveillance systems play an important role in mitigating various types of misconduct. However, observing a vast number of surveillance feeds is an arduous task that can be solved with a computerized system. A motion-based estimation method is proposed to measure the weighted temporal difference between consecutive frames by using spatial mean-sigma observations. It calculates the temporal frame differences that indicate the object vectors (OVs), of which statistical characteristics are evaluated periodically to identify any crowd abnormality. The proposed method is able to accurately detect the unusual event at the earliest scene, based on original and public data sets.

Music can generate a positive effect in runners’ performance and motivation. However, the practical implementation of music intervention during exercise is mostly absent from the literature. Therefore, this paper designs a playback sequence system for joggers by considering music emotion and physiological signals. This playback sequence is implemented by a music selection module that combines artificial intelligence techniques with physiological data and emotional music. In order to make the system operate for a long time, this paper improves the model and selection music module to achieve lower energy consumption. The proposed model obtains fewer FLOPs and parameters by using logarithm scaled Mel-spectrogram as input features. The accuracy, computational complexity, trainable parameters, and inference time are evaluated on the Bi-modal, 4Q emotion, and Soundtrack datasets. The experimental results show that the proposed model is better than that of Sarkar et al. and achieves competitive performance on Bi-modal (84.91%), 4Q emotion (92.04%), and Soundtrack (87.24%) datasets. More specifically, the proposed model reduces the computational complexity and inference time while maintaining the classification accuracy, compared to other models. Moreover, the size of the proposed model for network training is small, which can be applied to mobiles and other devices with limited computing resources. This study designed the overall playback sequence system by considering the relationship between music emotion and physiological situation during exercise. The playback sequence system can be adopted directly during exercise to improve users’ exercise efficiency.

As an emissive display, the active matrix organic light-emitting diodes (AMOLEDs) endure lower power efficiency at the higher level of pixel intensities. In the existing techniques, the power consumption is lowered with a significant loss of details and hue alterations. This paper proposes a hue-preserving pixel-dimming technique using subtractive coefficients based on the decomposed hue–saturation–value color map that reduces the power consumption on AMOLED displays. An entropy-based scene detection is adopted to maintain the computational efficiency of the proposed method on video input. Experimental results on a 5.5-in 1080p AMOLED displays show that the proposed method conserves up to 73% of the displaying power with high perceptual qualities as compared with the existing methods.

This article presents an indoor positioning system based on camera and ultrasonic sensors mounted on a pair of glasses that is specifically designed for visually impaired individuals. The proposed system incorporates a recognition algorithm that is able to recognize certain color-coded markers with a detectable range of 15 m in real time on a quad-core embedded processor. In addition, microultrasonic transducers are implemented to detect obstacles in front of the glasses. The positioning system and obstacle detection can be adjusted to work independently or in parallel. The sensor data are then transmitted to a mobile phone to be processed. The results show that the proposed system is able to recognize markers and detect obstacles with low complexity and power consumption.

Robot localization mechanism is an essential feature to determine the position of the corresponding robot within an environment, particularly in the field of Standard Platform League (SPL) at the RoboCup. Despite the available input from the onboard sensors, the ground truth information is necessary for a real-time localization system. This study proposes an efficient color-based segmentation scheme using an overhead projective camera with an autonomous calibration procedure. This enhances the system robustness against lighting changes and different labeling setups for the field environment. The experimental results show that the proposed method localizes and recognizes objects with a detection rate of 96.4%.

This study proposes a computer-aided region segmentation for the plain chest radiographs. It incorporates an avant-garde contrast enhancement that increases the opacity of the lung regions. The region of interest (ROI) is localized preliminarily by implementing a brisk block-based binarization and morphological operations. Further improvement for region boundaries is performed using a statistical-based region growing with an adaptive graph-cut technique that increases accuracy within any dubious gradient. Assessed on a representative dataset, the proposed method achieves an average segmentation accuracy of 96.3% with low complexity on 256p resolutions.

Hearing aids are increasingly essential for people with hearing loss. For this purpose, environmental noise estimation and classification are some of the required technologies. However, some noise classifiers utilize multiple audio features, which cause intense computation. In addition, such noise classifiers employ inputs of different time lengths, which may affect classification performance. Thus, this paper proposes a model architecture for noise classification, and performs experiments with three different audio segment time lengths. The proposed model attains fewer floating-point operations and parameters by utilizing the log-scaled mel-spectrogram as an input feature. The proposed models are evaluated with classification accuracy, computational complexity, trainable parameters, and inference time on the UrbanSound8k dataset and HANS dataset. The experimental results showed that the proposed model outperforms other models on two datasets. Furthermore, compared with other models, the proposed model reduces model complexity and inference time while maintaining classification accuracy. As a result, the proposed noise classification for hearing aids offers less computational complexity without compromising performance.

Articulation disorder means having difficulties during pronunciations, leading to incorrect articulations and unclear sentences. Articulation disorder has been a common child language issue. Currently, there is no any unified sayings for articulation disorders classification in the Taiwan's medical field. Thus, a speech therapist is required for analysis and treatment in hospitals. After a series of pronunciations, a speech therapist will make an analysis based on childrens pronunciations. Children will return to the hospitals for months continuously to improve their conditions. Nevertheless, the treatment can only benefit children with articulation disorder by receiving treatments in hospitals, slowing down the treatment cycle. The purpose of this work is to automate the diagnosis for articulation disorder by combining the latest AI's convolutional neural network (CNN). Results show that LeNet-5 which achieved 94.56% Top-1 accuracy with the smallest model size is more suitable to apply articulations disorder application on mobile devices.

In recent years, the application of surface electromyography (sEMG) has increasingly more prominent, while the development of deep learning algorithms cannot be ignored. In this paper, a deep neural network that combined 3D-convolutional layers and a long short-term memory layer (LSTM) is proposed. Meanwhile, we propose data augmentation methods and a transfer learning algorithm to improve our performance. This proposed method is shown to outperform the other networks in sEMG-based recognition of lower limb abnormality. The experiments with 94.12% accuracy show that our method is effective for recognition of lower limb abnormality.

Most of the head-mounted displays take the active-matrix organic light emitting diode (AMOLED) as the primary display panel because of its displaying superiorities. Yet, the AMOLED displays are still regarded as power-hungry components; in order to reduce the power consumption of AMOLED displays, the input image would be suppressed based on the proposed dynamic lightness adjustment algorithm that incorporates the depth information from the stereoscopic images which indicates the saliency, and the lightness of image pixel-wisely. The experiments reveal that the proposed method could achieve the approximately high power-saving rate with lower computational overheads compared to the existing methods.

Packet classification has been recognized as one of the most significant functions in contemporary network infrastructures. Furthermore, a number of modern applications such as IoTs contain very strict constraints on the latency of network transmissions. This paper presents the design and implementation of a novel packet classification based on FPGA architecture. The proposed design contains a Latency Compression Scheme (LCS) to achieve the low-latency packet processing. Furthermore, this structure supports 12-tuple fields for the modern Internet traffics. The experimental results show that the proposed packet classification scheme reduces the delay of packet processing by 2.18× compared to the state-of-the-art works.

This article presents a method for implementing the deep learning object detection based on a low-cost edge computing IoT device. The limit of the hardware is a challenge for working the pre-trained neural network model on a low-cost IoT device. Hence, we utilize the Neural Compute Stick (NCS) to accelerate the neural network model on a low-cost IoT device by its high efficiency floating-point operation. With the NCS, the low-cost IoT device can successfully work the pre-trained neural network model and become an edge computing device. The experimental results show the proposed method can effectively detect the objects based on deep learning on an edge computing IoT device. Furthermore, the objective experiment demonstrates the proposed method can immediately infer the neural network model for images in average 1.7 seconds with only one of the NCS and the neural network model can reach average 9.2 fps for the video sequences with four NCSs acceleration. In addition, the discrepancy of the neural network model between the edge device and the edge server is less than 2% mean average precision (mAP).

Active matrix organic light-emitting diode (AMOLED) display is an emerging technology that has been widely used in mobile devices. It incorporates a varying power-demand character, which is particularly inefficient for displaying bright colors. This paper presents a power-saving scheme for AMOLED display based on the pixel dimming transformation. Initially, the proposed method suppresses luminance and reconstructs colors of any overexposed region. Then, each pixel is transformed using coefficients derived from the SSIM metric, of which the set point is determined from the prior image. Weighted coefficients are employed specifically for high-frequency visible spectra. According to the measurements of three datasets on a smartphone with AMOLED display, up to 75% of displaying power can be reduced with mean visual saliency index of 0.955.

For the modern multimedia devices, display resolution and image quality are actively improved nowadays. Although such improvement can produce the high visual perception for the observer, the power consumption becomes an inevasible problem as it is rising progressively. In order to achieve a good balance between visual perception and power consumption, we propose a histogram-based power saving algorithm to improve the image contrast for OLED display panels. The proposed algorithm modifies the empty bins of the image histogram as a pre-process of power reduction. Furthermore, the visual effect is compensated using the power saving histogram equalization algorithm. Experimental results show that the proposed algorithm not only decreases the display power to be lower than that of compared algorithms, but also generates the highly perceptual contrast of the images.

Contrast enhancement is usually applied to those images captured in poor lighting conditions for improving the visual quality. Using interpixel contextual information, a 2-D histogram based contrast enhancement (CE) was proposed to improve image contrast and preserve more details as well. In order to maintain the balance between contrast enhancement and detail preservation, the window size of a 2-D histogram-based contrast enhancement should be adjustable based on the original image contrast and details. In addition, the computation intensive 2-D histogram based CE should be accelerated for real-time applications. Thus, we propose an efficient dynamic window size 2-D histogram construction algorithm in this paper. The proposed algorithm divides the input image into sub-blocks and assigns them appropriate window sizes, which depend upon the standard deviation and the number of distinct intensity values of each individual sub-block. Furthermore, the integral histogram is employed to be able to compute the dynamic range 2-D histogram in constant time while fluctuant window size is adopted dynamically. Experimental results demonstrate the efficacy and efficiency of the proposed algorithm.

The large-scale surveillance videos analysis becomes important as the development of the intelligent city; however, the heavy computational resources necessary for the state-of-the-art deep learning model makes real-time processing hard to be implemented. As the characteristic of high scene similarity generally existing in surveillance videos, we propose an effective compression architecture called dynamic convolution, which can reuse the previous feature maps to reduce the calculation amount; and combine with filter pruning to further speed up the performance. In this paper, we tested the presented method on 45 surveillance videos with various scenes. The experimental results show that the hybrid pruning architecture can reduce up to 80.4% of FLOPs while preserving the precision within 1.34% mAP; furthermore, the method can improve the processing speed up to 2.8 times compared to the traditional Single Shot MultiBox Detection.

This paper discusses the current state of project “OurPuppet” which is a system to support elderly people suffering from dementia. In contrast to other systems, OurPuppet not only focuses on the patient but also on caregiving relatives. If the caregiving relatives aren't present, many dependent persons show high insecurity. In turn, this limits the freedom of caregiving persons and hence leads to an emotional burden or even to an isolation of the caregiving relatives. Hence, in the project “OurPuppet” we are developing a sensor based interactive puppet, which is supposed to support the caregiving relatives, relieve the burden from them. The puppet acts as a companion for the dependent persons as well as a communication interface between the caregiving relatives and them. To this end, the puppet is equipped with suitable sensors to detect and monitor the emotional condition of the depending person and provide the caregiving relatives with the appropriate information. The puppet can also show emotional expressions in order to provide human like interaction capabilities. The puppet can intervene in situation where the dependent person needs assistance. This shall prevent isolation and decrease emotional burden of the caregiving relatives. Finally it will improve the situation for the dependent people as well.

Yield loss caused by via failures is unacceptably high in many semiconductor manufacturing processes. Redundant via insertion (RVI) is a typical approach for improving manufacturing competitiveness. The double-via insertion in concurrent routing or post-routing stage is introduced to improve yield and reliability. However, RVI with double-via patterns requires a considerable routing resource, which limits the enhancement of RVI rate, especially in congested designs. In this paper, an efficient and effective post-layout RVI approach is proposed. The proposed RVI method is the first to overcome the limitation of double-via insertion by combining double-via and rectangle-via patterns. The proposed RVI method determines the via configuration used in a two-phase flow which can effectively increase the insertion rate of total via which is dominated by via1. Experimental results show that the proposed method uses multi-via mechanisms to improve the total via insertion rate in RVI designs from 86.8 to 92.7 % and to improve the via1 insertion rate from 73.5 to 85.1 %.

As an emissive display, the organic light emitting diode (OLED) endure an indispensable role on the market growth of consumer electronics. Despite the preferable power efficiency, the active matrix OLED (AMOLED) displays still consume large energy. To adaptively increase the power efficiency of AMOLED displays, we propose an adjustable pixel dimming algorithm based on the structural similarity metric on the HSI color-space with a weighted pixel suppressor that selectively reduces the intensity of the high-frequency spectrum. Experimental results show that the proposed method offers higher power reduction rate compared with the existing algorithms, but also achieve high visual saliency index.

This paper presents high performance kinect depth video compression with low computational cost. In the proposed scheme, down-sample is placed at the first part of pre-processing to reduce computational cost. Therefore, down-sampled image is refined in the middle of pre-processing part. At last stage of preprocessing, up-sample scheme is located in order to return into the original size then feed into H.264/AVC encoder. The proposed method can gain 3–4 dB and reduce computational time by 44 %–50 %.