Search results for: Swin transformer

Authors: Jia Li, Xing Wei, Yuchen Hong, Yang Lu

Abstract:

Detecting vehicle behavior has always been the focus of intelligent transportation, but with the explosive growth of the number of vehicles and the complexity of the road environment, the vehicle behavior videos captured by traditional surveillance have been unable to satisfy the study of vehicle behavior. The traditional method of manually labeling vehicle behavior is too time-consuming and labor-intensive, but the existing object detection and tracking algorithms have poor practicability and low behavioral location detection rate. This paper proposes a vehicle behavior detection algorithm based on the dual-stream convolution network and the multi-dimensional video dynamic detection network. In the videos, the straight-line behavior of the vehicle will default to the background behavior. The Changing lanes, turning and turning around are set as target behaviors. The purpose of this model is to automatically mark the target behavior of the vehicle from the untrimmed videos. First, the target behavior proposals in the long video are extracted through the dual-stream convolution network. The model uses a dual-stream convolutional network to generate a one-dimensional action score waveform, and then extract segments with scores above a given threshold M into preliminary vehicle behavior proposals. Second, the preliminary proposals are pruned and identified using the multi-dimensional video dynamic detection network. Referring to the hierarchical reinforcement learning, the multi-dimensional network includes a Timer module and a Spacer module, where the Timer module mines time information in the video stream and the Spacer module extracts spatial information in the video frame. The Timer and Spacer module are implemented by Long Short-Term Memory (LSTM) and start from an all-zero hidden state. The Timer module uses the Transformer mechanism to extract timing information from the video stream and extract features by linear mapping and other methods. Finally, the model fuses time information and spatial information and obtains the location and category of the behavior through the softmax layer. This paper uses recall and precision to measure the performance of the model. Extensive experiments show that based on the dataset of this paper, the proposed model has obvious advantages compared with the existing state-of-the-art behavior detection algorithms. When the Time Intersection over Union (TIoU) threshold is 0.5, the Average-Precision (MP) reaches 36.3% (the MP of baselines is 21.5%). In summary, this paper proposes a vehicle behavior detection model based on multi-dimensional dynamic detection network. This paper introduces spatial information and temporal information to extract vehicle behaviors in long videos. Experiments show that the proposed algorithm is advanced and accurate in-vehicle timing behavior detection. In the future, the focus will be on simultaneously detecting the timing behavior of multiple vehicles in complex traffic scenes (such as a busy street) while ensuring accuracy.

Keywords: vehicle behavior detection, convolutional neural network, long short-term memory, deep learning

Procedia PDF Downloads 130

Authors: George Zhou, Yunchan Chen, Candace Chien

Abstract:

Kidney replacement therapy is the current standard of care for end-stage renal diseases. In-center or home hemodialysis remains an integral component of the therapeutic regimen. Arteriovenous fistulas (AVF) make up the vascular circuit through which blood is filtered and returned. Naturally, AVF patency determines whether adequate clearance and filtration can be achieved and directly influences clinical outcomes. Our aim was to build a deep learning model for automated AVF stenosis screening based on the sound of blood flow through the AVF. A total of 311 patients with AVF were enrolled in this study. Blood flow sounds were collected using a digital stethoscope. For each patient, blood flow sounds were collected at 6 different locations along the patient’s AVF. The 6 locations are artery, anastomosis, distal vein, middle vein, proximal vein, and venous arch. A total of 1866 sounds were collected. The blood flow sounds are labeled as “patent” (normal) or “stenotic” (abnormal). The labels are validated from concurrent ultrasound. Our dataset included 1527 “patent” and 339 “stenotic” sounds. We show that blood flow sounds vary significantly along the AVF. For example, the blood flow sound is loudest at the anastomosis site and softest at the cephalic arch. Contextualizing the sound with location metadata significantly improves classification performance. How to encode and incorporate categorical metadata is an active area of research1. Herein, we study ordinal (i.e., integer) encoding schemes. The numerical representation is concatenated to the flattened feature vector. We train a vision transformer (ViT) on spectrogram image representations of the sound and demonstrate that using scalar multiples of our integer encodings improves classification performance. Models are evaluated using a 10-fold cross-validation procedure. The baseline performance of our ViT without any location metadata achieves an AuROC and AuPRC of 0.68 ± 0.05 and 0.28 ± 0.09, respectively. Using the following encodings of Artery:0; Arch: 1; Proximal: 2; Middle: 3; Distal 4: Anastomosis: 5, the ViT achieves an AuROC and AuPRC of 0.69 ± 0.06 and 0.30 ± 0.10, respectively. Using the following encodings of Artery:0; Arch: 10; Proximal: 20; Middle: 30; Distal 40: Anastomosis: 50, the ViT achieves an AuROC and AuPRC of 0.74 ± 0.06 and 0.38 ± 0.10, respectively. Using the following encodings of Artery:0; Arch: 100; Proximal: 200; Middle: 300; Distal 400: Anastomosis: 500, the ViT achieves an AuROC and AuPRC of 0.78 ± 0.06 and 0.43 ± 0.11. respectively. Interestingly, we see that using increasing scalar multiples of our integer encoding scheme (i.e., encoding “venous arch” as 1,10,100) results in progressively improved performance. In theory, the integer values do not matter since we are optimizing the same loss function; the model can learn to increase or decrease the weights associated with location encodings and converge on the same solution. However, in the setting of limited data and computation resources, increasing the importance at initialization either leads to faster convergence or helps the model escape a local minimum.

Keywords: arteriovenous fistula, blood flow sounds, metadata encoding, deep learning

Procedia PDF Downloads 87

Authors: P. Ramesh, Aby Joseph, Arya G. Lal, U. S. Aji

Abstract:

Integration of distributed green renewable energy sources in addition with battery energy storage is an inevitable requirement in a smart grid environment. To achieve this, an Intelligent Green Energy Management System (i-GEMS) needs to be incorporated to ensure coordinated operation between supply and load demand based on the hierarchy of Renewable Energy Sources (RES), battery energy storage and distribution grid. A bi-directional energy harvester is an integral component facilitating Intelligent Green Energy Management System (i-GEMS) and it is required to meet the technical challenges mentioned as follows: (1) capability for bi-directional mode of operation (buck/boost) (2) reduction of circuit parasitic to suppress voltage spikes (3) converter startup problem (4) high frequency magnetics (5) higher power density (6) mode transition issues during battery charging and discharging. This paper is focused to address the above mentioned issues and targeted to design, develop and implement a bi-directional energy harvester with galvanic isolation. In this work, the hardware architecture for bi-directional energy harvester rated 3.5 kW is developed with Isolated Full Bridge Boost Converter (IFBBC) as well as Dual Active Bridge (DAB) Converter configuration using modular power electronics hardware which is identical for both solar PV array and battery energy storage. In IFBBC converter, the current fed full bridge circuit is enabled and voltage fed full bridge circuit is disabled through Pulse Width Modulation (PWM) pulses for boost mode of operation and vice-versa for buck mode of operation. In DAB converter, all the switches are in active state so as to adjust the phase shift angle between primary full bridge and secondary full bridge which in turn decides the power flow directions depending on modes (boost/buck) of operation. Here, the control algorithm is developed to ensure the regulation of the common DC link voltage and maximum power extraction from the renewable energy sources depending on the selected mode (buck/boost) of operation. The circuit analysis and simulation study are conducted using PSIM 9.0 in three scenarios which are - 1.IFBBC with passive clamp, 2. IFBBC with active clamp, 3. DAB converter. In this work, a common hardware prototype for bi-directional energy harvester with 3.5 kW rating is built for IFBBC and DAB converter configurations. The power circuit is equipped with right choice of MOSFETs, gate drivers with galvanic isolation, high frequency transformer, filter capacitors, and filter boost inductor. The experiment was conducted for IFBBC converter with passive clamp under boost mode and the prototype confirmed the simulation results showing the measured efficiency as 88% at 2.5 kW output power. The digital controller hardware platform is developed using floating point microcontroller TMS320F2806x from Texas Instruments. The firmware governing the operation of the bi-directional energy harvester is written in C language and developed using code composer studio. The comprehensive analyses of the power circuit design, control strategy for battery charging/discharging under buck/boost modes and comparative performance evaluation using simulation and experimental results will be presented.

Keywords: bi-directional energy harvester, dual active bridge, isolated full bridge boost converter, intelligent green energy management system, maximum power point tracking, renewable energy sources

Procedia PDF Downloads 140

Authors: Reza Nematirad, Anil Pahwa, Balasubramaniam Natarajan

Abstract:

Time series data consist of successive data points collected over a period of time. Accurate prediction of future values is essential for informed decision-making in several real-world applications, including electricity load demand forecasting, lifetime estimation of industrial machinery, traffic planning, weather prediction, and the stock market. Due to their critical relevance and wide application, there has been considerable interest in time series forecasting in recent years. However, the proliferation of sensors and IoT devices, real-time monitoring systems, and high-frequency trading data introduce significant intricate temporal variations, rapid changes, noise, and non-linearities, making time series forecasting more challenging. Classical methods such as Autoregressive integrated moving average (ARIMA) and Exponential Smoothing aim to extract pre-defined temporal variations, such as trends and seasonality. While these methods are effective for capturing well-defined seasonal patterns and trends, they often struggle with more complex, non-linear patterns present in real-world time series data. In recent years, deep learning has made significant contributions to time series forecasting. Recurrent Neural Networks (RNNs) and their variants, such as Long short-term memory (LSTMs) and Gated Recurrent Units (GRUs), have been widely adopted for modeling sequential data. However, they often suffer from the locality, making it difficult to capture local trends and rapid fluctuations. Convolutional Neural Networks (CNNs), particularly Temporal Convolutional Networks (TCNs), leverage convolutional layers to capture temporal dependencies by applying convolutional filters along the temporal dimension. Despite their advantages, TCNs struggle with capturing relationships between distant time points due to the locality of one-dimensional convolution kernels. Transformers have revolutionized time series forecasting with their powerful attention mechanisms, effectively capturing long-term dependencies and relationships between distant time points. However, the attention mechanism may struggle to discern dependencies directly from scattered time points due to intricate temporal patterns. Lastly, Multi-Layer Perceptrons (MLPs) have also been employed, with models like N-BEATS and LightTS demonstrating success. Despite this, MLPs often face high volatility and computational complexity challenges in long-horizon forecasting. To address intricate temporal variations in time series data, this study introduces Times2D, a novel framework that parallelly integrates 2D spectrogram and derivative heatmap techniques. The spectrogram focuses on the frequency domain, capturing periodicity, while the derivative patterns emphasize the time domain, highlighting sharp fluctuations and turning points. This 2D transformation enables the utilization of powerful computer vision techniques to capture various intricate temporal variations. To evaluate the performance of Times2D, extensive experiments were conducted on standard time series datasets and compared with various state-of-the-art algorithms, including DLinear (2023), TimesNet (2023), Non-stationary Transformer (2022), PatchTST (2023), N-HiTS (2023), Crossformer (2023), MICN (2023), LightTS (2022), FEDformer (2022), FiLM (2022), SCINet (2022a), Autoformer (2021), and Informer (2021) under the same modeling conditions. The initial results demonstrated that Times2D achieves consistent state-of-the-art performance in both short-term and long-term forecasting tasks. Furthermore, the generality of the Times2D framework allows it to be applied to various tasks such as time series imputation, clustering, classification, and anomaly detection, offering potential benefits in any domain that involves sequential data analysis.

Keywords: derivative patterns, spectrogram, time series forecasting, times2D, 2D representation

Procedia PDF Downloads 42