Neural Network with Early Exits and Split Points

Kód výsledku v IS VaVaI

<a href="https://www.isvavai.cz/riv?ss=detail&h=RIV%2F68407700%3A21230%2F25%3A00383599" target="_blank" >RIV/68407700:21230/25:00383599 - isvavai.cz</a>

Výsledek na webu

<a href="https://gitlab.fel.cvut.cz/mobile-and-wireless/codes/public-software/nn-split-and-early-exit" target="_blank" >https://gitlab.fel.cvut.cz/mobile-and-wireless/codes/public-software/nn-split-and-early-exit</a>

DOI - Digital Object Identifier

—

Jazyk výsledku

angličtina

Název v původním jazyce

Neural Network with Early Exits and Split Points

Popis výsledku v původním jazyce

Our software constructs modular process pipelines for road sign recognition in real-time systems. In the first stage, a road sign is detected using the You Only Look Once algorithm, version 3 (YOLOv3). Once road sign is detected, the sign is cropped and passed to a custom convolutional neural network that supports early exit points and dynamic splitting of computation between the local system and a multi-access edge computing (MEC) server. The split points and early exits reduce the time required to classify the road sign at the cost of lower classification accuracy. The application has been tested on real-world data from autonomous vehicles.

Název v anglickém jazyce

Neural Network with Early Exits and Split Points

Popis výsledku anglicky

Our software constructs modular process pipelines for road sign recognition in real-time systems. In the first stage, a road sign is detected using the You Only Look Once algorithm, version 3 (YOLOv3). Once road sign is detected, the sign is cropped and passed to a custom convolutional neural network that supports early exit points and dynamic splitting of computation between the local system and a multi-access edge computing (MEC) server. The split points and early exits reduce the time required to classify the road sign at the cost of lower classification accuracy. The application has been tested on real-world data from autonomous vehicles.

Druh

R - Software

CEP obor

—

OECD FORD obor

20203 - Telecommunications

Projekt

<a href="/cs/project/LUASK22064" target="_blank" >LUASK22064: Prediktivní alokace výpočetních prostředků pro autonomní řízení na hraně sítě</a><br>

Návaznosti

P - Projekt vyzkumu a vyvoje financovany z verejnych zdroju (s odkazem do CEP)

Rok uplatnění

2025

Kód důvěrnosti údajů

S - Úplné a pravdivé údaje o projektu nepodléhají ochraně podle zvláštních právních předpisů

Interní identifikační kód produktu

nn-split-and-early-exit

Technické parametry

Our software starts by capturing an image of the vehicle's surroundings and detecting potential road signs using a lightweight object detection module, in current version is detection module You Only Look Once, version 3 (YOLOv3). Once a road sign is detected, road sign is cropped and forwarded to a custom convolutional neural network (CNN) designed with support for early exits and split processing. The classification task can be processed either locally on the autonomous vehicle (AV) or partially/fully offloaded to a multi-access edge computing (MEC) server, depending on the selected split point. The excange of information between MEC server and AV is obtained by the structured messages via Transmission Control Protocol (TCP). Early exits enable faster classification and reduced computational load by allowing the inference to terminate at earlier layers. However, this comes at the cost of lower classification accuracy. The input to the software is image data captured from a camera, which is processed internally by the software. The output consists either of a terminal printout showing the detected and classified road signs, or a structured message containing the classification result sent via TCP. For further details and usage instructions, please refer to the associated GitHub project page.

Ekonomické parametry

This software enables a reduction in the cost of computing resources on autonomous vehicles (AVs) by at least 50%. By offloading intensive computation, the AV no longer requires a dedicated GPU or high-performance CPU, which significantly lowers hardware costs. Additionally, this reduces demands on the power supply system, as it no longer needs to be dimensioned for GPU. The usage of MEC server also reduce the energy consumption of the CPU up to 2.6 and decrease the processing time of the CNN up to 2.5, which also reduce the cost of the AV especially the power system. Additionally, the use of a MEC server reduces the CPU's energy consumption by up to 2.6 times and decreases the CNN processing time by up to 2.5 times. These improvements lead to further cost savings, particularly in the vehicle’s power supply system.

IČO vlastníka výsledku

68407700

Název vlastníka

České vysoké učení technické v Praze / FEL / katedra telekomunikační techniky