No abstract

No abstract

No abstract

The full search motion estimation algorithm for video coding is a procedure of high computational cost. For this reason, in real-time low-power applications, low-cost motion estimation algorithms are viable solutions. A novel reduced complexity motion estimation algorithm is presented. It conjugates the reduction of computational load with good encoding efficiency. It exploits the past history of the motion field to predict the current motion field. A successive refinement phase gives the final motion field. This approach leads to a sensible reduction in the number of motion vector that have to be tested. The complexity is lower than any other algorithm algorithms known to the authors, in the literature, it is constant as there is no recursivity in the algorithm and independent of any search window area size. Experimental evaluations have shown the robustness of the algorithm when applied on a wide set of video sequences--a good performance compared to other reduced complexity algorithms and negligible loss of efficiency versus the full search algorithm.

Lithium-ion (Li-Ion) batteries are rechargeable batteries which can maximize battery lifespan thanks to their chemical abilities, at the same time increasing power energy density. For these reasons, Li-Ion batteries have earned considerable popularity, and they are widely used both in mobile computing devices (e.g. smartphones and smartwatches) and automotive systems (e.g. hybrid and electric vehicles). A fundamental parameter for battery health monitoring is the State of Health (SoH), which is computed from the maximum releasable capacity, and which represents battery functionality in energy storage and delivery. Among the most used data-driven approaches are Machine Learning (ML) algorithms, such as Support Vector Machines (SVMs), Random Forest (RF) regressions, and Artificial Neural Networks (ANNs). This article presents a comparison of different ML algorithms for estimating maximum releasable capacity of Li-Ion batteries, with a special focus on the implementation of both Forward and Recurrent ANNs (FNNs and RNNs, respectively), using prognostic Li-Ion battery data sets provided by the National Aeronautics and Space Administration (NASA). After an evaluation of models performances in terms of RMSE and MAE, STM32Cube.AI tool was used to convert pre-trained ANNs to optimized ANSI C code for STM32 microcontrollers (MCUs), and to profile their complexity automatically. Finally, in order to decrease models size with minimal accuracy loss, the implemented ANNs were quantized via STM32Cube.AI, converting weights and activations from 32-bit floating-point to 8-bit integer precision. TensorFlow Lite for Microcontrollers (TFLM) was used as benchmark in the analysis and validation of both non-quantized and quantized models, and the performances obtained via STM32Cube.AI and TFLM were compared.