Master thesis: Semantic Communications with Foundation Models for Vehicle and Road User Detection

Ericsson; Stockholms kommun; Sur place; SEK 134 000 - 201 000; Plein temps

Résumé du poste

Ericsson in Sweden, Stockholm is seeking a master’s student to explore semantic communications for next-generation wireless networks. You will work with frozen foundation-model features (CLIP-ViT-B/16 or DINOv2-ViT-B/14), compress them with a learned bottleneck, and transmit over a simulated noisy channel.

A lightweight decoder will reconstruct bounding boxes and labels for six COCO classes. You will review semantic communication and joint source-channel coding, extract patch-level features,

Qualifications

• Pursuing a master's degree in Electrical Engineering, Computer Science, Telecommunications, or related field.
• Strong programming skills in Python and PyTorch.
• Solid understanding of deep-learning concepts (CNNs, transformers, training procedures, loss functions).
• Ability to design experiments, work independently, and critically interpret results.
• Fluency in written and spoken English.
• Familiarity with object detection and pretrained vision models is an advantage.

Responsabilités

• Build a semantic-communication system to extract spatial features from images using frozen CLIP-ViT-B/16 or DINOv2-ViT-B/14 models.
• Compress features through a learned linear bottleneck and transmit over AWGN channel.
• Reconstruct bounding boxes and class labels with a lightweight decoder; evaluate performance.
• Train bottleneck and decoder end-to-end with noise injection across SNR levels.
• Compare foundation-model approaches with task-specific encoder-decoder and baselines like JPEG.
• Analyze results in a communications context (mAP vs SNR, per-class performance).

Connaissances

Python

PyTorch

Deep learning

Independent work

English fluency

Object detection knowledge

Formation

Master's degree in Electrical Engineering, Computer Science, Telecommunications

Outils

YOLO

Faster R-CNN

FCOS

Description du poste

About this opportunity:

Next-generation wireless networks, including 5G-Advanced and 6G, will connect large numbers of bandwidth-constrained devices such as vehicles, surveillance cameras, and drones. Transmitting raw or conventionally compressed images can use valuable spectrum on details irrelevant to a particular task. Semantic communication instead aims to transmit task-relevant information—such as detected objects and their locations—while preserving what is needed for downstream decisions.

Join our Team

Next-generation wireless networks, including 5G-Advanced and 6G, will connect large numbers of bandwidth-constrained devices such as vehicles, surveillance cameras, and drones. Transmitting raw or conventionally compressed images can use valuable spectrum on details irrelevant to a particular task. Semantic communication instead aims to transmit task-relevant information—such as detected objects and their locations—while preserving what is needed for downstream decisions.

Vision foundation models such as CLIP and DINOv2 can provide rich image representations without task-specific pretraining. This thesis will investigate whether their spatial features can be compressed and transmitted over a noisy wireless channel while maintaining object-detection performance.

Research Question

How well do frozen foundation-model spatial features withstand compression and wireless-channel noise when used to detect vehicles and other road users?

What you will do:

Build and evaluate a semantic-communication system that extracts spatial features from images using a frozen CLIP-ViT-B/16 or DINOv2-ViT-B/14 model, compresses them through a learned linear bottleneck, quantizes and transmits them over a simulated additive white Gaussian noise (AWGN) channel, and reconstructs bounding boxes and class labels with a lightweight decoder. The study will use six driving-relevant COCO classes: car, truck, bus, motorcycle, person, and bicycle.

• Review semantic communications, joint source-channel coding (DeepJSCC), foundation models, and JPEG compression.
• Extract patch-level features from the selected COCO classes.
• Implement the learned bottleneck, quantization, AWGN channel simulation, and lightweight anchor-free detection decoder.
• Train the bottleneck and decoder end-to‑end with noise injection at random signal-to-noise ratio (SNR) levels.
• Compare both foundation-model approaches with a task‑specific encoder‑decoder and standard baselines such as JPEG.
• Analyze results in a communications context, including mean average precision (mAP) versus SNR and per‑class performance.

The skills you bring:

• You are pursuing a master’s degree in Electrical Engineering, Computer Science, Telecommunications, or a related field.
• Strong programming skills in Python and PyTorch.
• Good understanding of deep‑learning concepts, including CNNs, transformers, training procedures, and loss functions.
• Ability to work independently, design experiments, and critically interpret results.
• Fluency in written and spoken English.
• Familiarity with object detection (e.g., YOLO, Faster R-CNN, or FCOS), wireless communications (e.g., SNR and AWGN), or pretrained vision models (e.g., CLIP, ViT, or DINOv2) is an advantage.

Project Details and Contacts

• Extent: One student
• Contact: Aneta Vulgarakis
• Supervisors: Cristian Vaca Rubio, Maxim Teslenko

Why join Ericsson?

At Ericsson, you´ll have an outstanding opportunity. The chance to use your skills and imagination to push the boundaries of what´s possible. To build solutions never seen before to some of the world’s toughest problems. You´ll be challenged, but you won’t be alone. You´ll be joining a team of diverse innovators, all driven to go beyond the status quo to craft what comes next.

Primary country and city:

Sweden (SE) || Stockholm