Deep Reinforcement Learning (DRL)-based Dynamic Voltage Frequency Scaling (DVFS) has shown great promise for energy conservation in embedded systems. While many works were devoted to validating its efficacy or improving its performance, few discuss the feasibility of the DRL agent deployment for embedded computing. State-of-the-art approaches focus on the miniaturization of agents’ inferential networks, such as pruning and quantization, to minimize their energy and resource consumption. However, this spatial-based paradigm still proves inadequate for resource-stringent systems. In this paper, we address the feasibility from a temporal perspective, where FiDRL, a flexible invocation-based DRL model is proposed to judiciously invoke itself to minimize the overall system energy consumption, given that the DRL agent incurs non-negligible energy overhead during invocations. Our approach is three-fold: (1) FiDRL that extends DRL by incorporating the agent's invocation interval into the action space to achieve invocation flexibility; (2) a FiDRL-based DVFS approach for both inter- and intra-task scheduling that minimizes the overall execution energy consumption; and (3) a FiDRL-based DVFS platform design and an on/off-chip hybrid algorithm specialized for training the DRL agent for embedded systems. Experiment results show that FiDRL achieves 55.1% agent invocation cost reduction, under 23.3% overall energy reduction, compared to state-of-the-art approaches.

Abstract—Auto-registration video synthetic aperture radar (ViSAR), which requires real time pixel index unifying and resolution matching, is of great significance due to its applicability in multi-aspect observation and continuous monitoring. The phase error induced by the wavefront planar assumption, however, varies with different ViSAR frames, which limits the size of auto-registration imaging scene. To enlarge the auto-registration imaging swath, a parameter-adjusting auto-registration overlapped subaperture algorithm (PAAR-OSA) is proposed in this paper. By collaboratively designing the subapertures within each frame and among different frames in the stabilized-scene coordinate and cooperatively compensating the phase error of all frames, auto-registration with larger imaging swath can be achieved. Both the point targets and distributed targets validation results verify the superiority of the proposed method compared with existing algorithms.

Video synthetic aperture radar (ViSAR) is gaining increasing attention in the remote sensing area due to its advantages in continuous observation compared with conventional synthetic aperture radar (SAR). Generally, the polar format algorithm (PFA) is utilized to generate the ViSAR frames considering both processing efficiency and imaging quality. However, the changes in squint angles for different ViSAR frames will cause variations in target positions when the classic fixed-parameter PFA is utilized. Hence, an image registration step is typically needed, which greatly increases the computational load. Meanwhile, the azimuth resolution also deteriorates rapidly for the ViSAR frames as the squint angle becomes larger. In order to solve the aforementioned problems, a parameter-adjusting autoregistration PFA (PAAR-PFA) is proposed for the ViSAR. By adjusting system parameters, such as carrier frequency, pulsewidth, and sampling frequency, according to the azimuth sampling positions, PAAR-PFA can achieve autoregistration for ViSAR frames without range interpolation and geometric correction in fixed-parameter PFA, which significantly improves the processing efficiency. At the same time, since the azimuth sampling number of frame data is adjusted with the squint angle, the variation range of the azimuth resolution under different squint angles is considerably reduced. Point target and extended target simulations confirm the feasibility of the proposed algorithm.

Due to the complexity and inaccuracy of the force model of low earth orbit(LEO) satellite, the ephemeris data is not accurate enough for translation compensation, which will cause the residual translation component error, then leading to the dissatisfaction of the range and azimuth resolution for imaging. To face the problem above, the secondary correction of translational component is carried out by updating the ephemeris data fitted by the ephemeris data calculated by orbit two lines elements(TLE) and center slant which is estimated with the real echo data of the space station through correlation method, after a coarse correction-the registration of ephemeris data and echo data-and then the Inverse synthetic aperture radar(ISAR) image of the space station is obtained by combining the polar format algorithm(PFA), providing the foundation for monitoring, identification and tracking of space targets. The imaging results prove the validity of the method, which can be referred to in the imaging of LEO satellite and the real data processing based on ephemeris data.

The design and implementation of an integrated wearable face recognition and training system for prosopagnosia patients are presented. The purpose of this assistive technology is to provide real-time memory assistance and long-term rehabilitation. The real-time face recognition mode provides audio and visual notification of people who interact with the subject, while the at-home training mode combines features of mnemonic and perceptual training to help with prosopagnosia rehabilitation. In addition, a custom eye tracker is developed to determine the person whom the subject is making eye contact with within a crowd. Using the inverted face effect to mimic the difficulties of prosopagnosia patients, clinically healthy participants have shown improvements in their face-naming abilities. Early results indicate the system's potential to enrich the well-being of prosopagnosia patients.

A novel food volume measurement technique is proposed in this paper for accurate quantification of the daily dietary intake of the user. The technique is based on simultaneous localisation and mapping (SLAM), a modified version of convex hull algorithm, and a 3D mesh object reconstruction technique. This paper explores the feasibility of applying SLAM techniques for continuous food volume measurement with a monocular wearable camera. A sparse map will be generated by SLAM after capturing the images of the food item with the camera and the multiple convex hull algorithm is applied to form a 3D mesh object. The volume of the target object can then be computed based on the mesh object. Compared to previous volume measurement techniques, the proposed method can measure the food volume continuously with no prior information such as pre-defined food shape model. Experiments have been carried out to evaluate this new technique and showed the feasibility and accuracy of the proposed algorithm in measuring food volume.