Bandwidth-adaptive Cloud-Assisted 360-Degree 3D Perception for Autonomous Vehicles

Faisal Hawladera; Rui Meireles; Gamal Elghazaly; Ana Aguiar; Raphaël Frank

cs.CVFeb 27, 2026

Bandwidth-adaptive Cloud-Assisted 360-Degree 3D Perception for Autonomous Vehicles

Authors: Faisal Hawladera, Rui Meireles, Gamal Elghazaly, Ana Aguiar, Raphaël Frank

Organizations: Interdisciplinary Centre for Security, Reliability, and Trust (SnT), University of Luxembourg, L-1855, Luxembourg · Computer Science Department, Vassar College, Poughkeepsie, NY 12604, USA · Instituto de Telecomunicações, Faculdade de Engenharia, Universidade do Porto, Rua Dr. Roberto Frias, 4200–465 Porto, Portugal

Abstract

A key challenge for autonomous driving lies in maintaining real-time situational awareness regarding surrounding obstacles under strict latency constraints. The high processing requirements coupled with limited onboard computational resources can cause delay issues, particularly in complex urban settings. To address this, we propose leveraging Vehicle-to-Everything (V2X) communication to partially offload processing to the cloud, where compute resources are abundant, thus reducing overall latency. Our approach utilizes transformer-based models to fuse multi-camera sensor data into a comprehensive Bird's-Eye View (BEV) representation, enabling accurate 360-degree 3D object detection. The computation is dynamically split between the vehicle and the cloud based on the number of layers processed locally and the quantization level of the features. To further reduce network load, we apply feature vector clipping and compression prior to transmission. In a real-world experimental evaluation, our hybrid strategy achieved a 72 % reduction in end-to-end latency compared to a traditional onboard solution. To adapt to fluctuating network conditions, we introduce a dynamic optimization algorithm that selects the split point and quantization level to maximize detection accuracy while satisfying real-time latency constraints. Trace-based evaluation under realistic bandwidth variability shows that this adaptive approach improves accuracy by up to 20 % over static parameterization with the same latency performance.

Figures & tables

Panel of Figure 1 — Figure 1: In the Onboard Computing scenario, the BEVFormer model runs locally, transmitting detection results as CPMs over ITS-G5. In the Hybrid Computing scenario, a compressed feature vector is sent via C-V2X to the cloud for intensive processing, with detection results broadcast to nearby vehicles.

Platform	Hardware Configuration
Local ( $\approx$ 30W)	NVIDIA Jetson Orin
	2048 CUDA Cores, 131.4 TOPS (INT8)
	8-core ARM Cortex-A78AE
Cloud ( $\approx$ 3000W)	2x Intel Xeon Skylake CPUs (56 cores total)
	4x NVIDIA Tesla V100 GPUs (16 GB each)
	20480 CUDA Cores, 500 TFLOPS (FP16)

Table 1: Hardware setups for the onboard and cloud computing platforms described in Section 3.1 .

Figure 2: A basic overview of different containers included in the CPM message format as defined by the ETSI standard [ 11 ] .

Quantization	Inference (ms)	NDS	CPM (ms)	End-to-end delay (ms)
FP32	486	0.52	5.9 $(\pm 1.8)$	491.9 $(\pm 2.7)$
FP16	257	0.52	5.7 $(\pm 1.7)$	262.7 $(\pm 2.3)$
FP8	194	0.51	4.9 $(\pm 1.6)$	198.9 $(\pm 2.3)$

Table 2: Performance metrics for the BEVFormer model with ResNet101 backbone, evaluated using TensorRT optimization at different quantization levels (FP32, FP16, FP8) on the onboard vehicle platform, as detailed in Section 3.2 . The table includes inference time and CPM transmission latency, which together form the end-to-end delay. Standard deviations

(\pm)

are provided to reflect measurement variability.

Parameter Name	Value
Transmission Power (Tx)	23 dBm
Energy threshold	-85 dBm
Channel bandwidth / carrier frequency	$10\text{\,}\mathrm{MHz}$ / $5.9\text{\,}\mathrm{GHz}$
Radio Configuration	Single Channel (CCH)
Data rate	$7\text{\,}\mathrm{Mbit}\text{/}\mathrm{s}$
Number of CPM Transmitted / loss ratio	6000 / 0.09

Table 3: Important network parameters for V2X.

Figure 3: CPM transmission latency versus distance between a moving vehicle (

25\text{\,}\mathrm{km}\text{/}\mathrm{h}

avg.) and a stationary receiver at fixed coordinates (longitude: 6.161993, latitude: 49.626478). Dashed line shows average latency

4.10\text{\,}\mathrm{ms}

Figure 4: Feature vector size and extraction time versus split depth. Solid lines represent feature extraction time (left y-axis), while dashed lines indicate feature size (right y-axis). Lower-precision quantization (e.g., FP16, FP8) reduces both extraction time and feature size.

Figure 5: Transmission latency of feature vectors from vehicle to cloud across five split layers for FP32, FP16, and FP8 over a 5G network using C-V2X. FP8 demonstrates the lowest and most stable latency, suitable for real-time transmission. FP32 exhibits the highest latency and variability, especially at earlier split layers due to larger feature size.

FP	Split layer	Onboard Processing Time (ms)		Transmission latency (ms)		Cloud Processing Time (ms)		End-to-end Delay (ms)	NDS	Bandwidth Usage ( $\text{\,}\mathrm{Mbit}\text{/}\mathrm{s}$ )
FP	Split layer	Backbone	Compression	V2C	C2V	Decompression	Head	End-to-end Delay (ms)	NDS
32	1	17.2 $(\pm 2.10)$	10.7 $(\pm 1.85)$	65.8 $(\pm 4.00)$	11.6 $(\pm 1.15)$	2.6 $(\pm 0.60)$	20.8 $(\pm 1.35)$	128.7 $(\pm 4.20)$	0.52	10.5
	2	22.3 $(\pm 1.75)$	8.6 $(\pm 1.55)$	58.0 $(\pm 3.90)$	9.8 $(\pm 0.95)$	2.4 $(\pm 0.55)$	18.4 $(\pm 1.25)$	119.6 $(\pm 3.90)$	0.50	8.4
	3	30.5 $(\pm 1.90)$	7.3 $(\pm 1.50)$	48.9 $(\pm 3.75)$	8.4 $(\pm 0.85)$	2.5 $(\pm 0.50)$	15.9 $(\pm 1.30)$	113.5 $(\pm 3.80)$	0.48	6.8
	4	39.8 $(\pm 1.65)$	6.4 $(\pm 1.30)$	54.5 $(\pm 3.50)$	7.0 $(\pm 0.75)$	2.3 $(\pm 0.45)$	14.7 $(\pm 1.10)$	124.7 $(\pm 3.60)$	0.47	5.9
	5	55.4 $(\pm 1.45)$	5.1 $(\pm 1.10)$	56.3 $(\pm 3.20)$	5.8 $(\pm 0.70)$	2.5 $(\pm 0.40)$	12.6 $(\pm 0.95)$	137.7 $(\pm 3.40)$	0.46	5.4

Table 4: Performance metrics for various split layers and quantization levels, including onboard processing time, V2C and C2V transmission latency, cloud processing time, and total end-to-end delay. The standard deviations

(\pm)

reflect variability. The last column shows bandwidth utilization for offloading feature vectors from vehicle to cloud. Rows with end-to-end delay below 100 ms are highlighted in green.

Figure 6: End-to-end delay vs. detection accuracy ( NDS ) across split layers and quantization levels. Earlier layers result in higher accuracy but increased delay, while intermediate layers provide a balance between delay and accuracy.

Figure 7: Qualitative results at split layer 3 with FP16 on the nuScenes validation set (SceneID: n008-2018-05-21-11-06-59-0400), an NDS score of 0.47. We show 3D bounding box predictions in multi-view (6x) camera images and the 360-degree BEV .

Figure 8: Upload bandwidth distribution used to evaluate the dynamic hybrid-computing parameter selection algorithm. The green triangle indicates the mean, and the orange line marks the median.

Figure 9: Distribution of parameter tuple usage by dynamic selection algorithm, as a function of the amount of bandwidth dedicated to the perception task.

Figure 10: Detection accuracy ( NDS ) and number of latency bound violations, as a function of bandwidth dedicated to the perception task, for the dynamic parameter selection algorithm, the highest-accuracy static algorithm

(split=1,FP32)

, and the latency minimizing static algorithm

(split=5,FP8)

Figure 11: Detection accuracy ( NDS ) gain offered by dynamic selection algorithm, as a function of the amount of bandwidth assigned to the perception task.

Computation Offloading

See all 15 papers

Apr 6, 2026

Adversarial Robustness Analysis of Cloud-Assisted Autonomous Driving Systems

Autonomous vehicles increasingly rely on deep learning-based perception and control, which impose substantial computational demands. Cloud-assisted architectures offload these functions to remote servers, enabling enhanced perception and coordinated decision-making through the Internet of Vehicles (IoV). However, this paradigm introduces cross-layer vulnerabilities, where adversarial manipulation of perception models and network impairments in the vehicle-cloud link can jointly undermine safety-critical autonomy. This paper presents a hardware-in-the-loop IoV testbed that integrates real-time perception, control, and communication to evaluate such vulnerabilities in cloud-assisted autonomous driving. A YOLOv8-based object detector deployed on the cloud is subjected to whitebox adversarial attacks using the Fast Gradient Sign Method (FGSM) and Projected Gradient Descent (PGD), while network adversaries induce delay and packet loss in the vehicle-cloud loop. Results show that adversarial perturbations significantly degrade perception performance, with PGD reducing detection precision and recall from 0.73 and 0.68 in the clean baseline to 0.22 and 0.15 at epsilon= 0.04. Network delays of 150-250 ms, corresponding to transient losses of approximately 3-4 frames, and packet loss rates of 0.5-5 % further destabilize closed-loop control, leading to delayed actuation and rule violations. These findings highlight the need for cross-layer resilience in cloud-assisted autonomous driving systems.

Maher Al Islam, Amr S. El-Wakeel

Lane Department of Computer Science and Electrical Engineering, West Virginia University, WV, USA

Adversarial Attacks on Object Detectors Adversarial Robustness Autonomous Driving Autonomous Driving Safety Evaluation Adversarial Attacks Autonomous Driving Perception

Ask PDF

May 27, 2026

Multi-Resolution End-to-End Deep Neural Network for Optimizing Latency-Accuracy Tradeoff in Autonomous Driving

Latency-accuracy tradeoffs are fundamental in real-time applications of deep neural networks (DNNs) for cyber-physical systems. In autonomous driving, in particular, safety depends on both prediction quality and the end-to-end delay from sensing to actuation. We observe that (1) when latency is accounted for, the latency-optimal network configuration varies with scene context and compute availability; and (2) a single fixed-resolution model becomes suboptimal as conditions change. We present a multi-resolution, end-to-end deep neural network for the CARLA urban driving challenge using monocular camera input. Our approach employs a convolutional neural network (CNN) that supports multiple input resolutions through per-resolution batch normalization, enabling runtime selection of an ideal input scale under a latency budget, as well as resolution retargeting, which allows multi-resolution training without access to the original training dataset. We implement and evaluate our multi-resolution end-to-end CNN in CARLA to explore the latency-safety frontier. Results show consistent improvements in per-route safety metrics - lane invasions, red-light infractions, and collisions - relative to fixed-resolution baselines.

Qitao Weng, Heechul Yun

University of Kansas Lawrence, KS, USA

Efficient Neural Network Inference Adaptive Inference Dynamic Neural Networks End-to-End Autonomous Driving Convolutional Neural Networks

Ask PDF

Aug 15, 2026

Risk-Adaptive Edge--Cloud Visual Reasoning for Communication-Efficient Autonomous Driving

Cloud-hosted vision-language models (VLMs) offer greater contextual reasoning capabilities than smaller onboard models, but frequent visual uploads increase communication overhead and add network and inference latency to tactical decisions. We present a risk-adaptive edge-cloud architecture in which onboard traffic assessment determines when cloud reasoning is requested. An onboard VLM and a lightweight detector capture temporal traffic conditions and path-relative hazards for conservative local response and selective cloud access. The cloud model provides tactical advice, while validation, vehicle control, and automatic emergency braking remain local. In CARLA experiments, our method matched the task success rate of periodic cloud access while reducing cloud requests by 54.1% and recording fewer automatic emergency braking (AEB) activations. In a delayed-roadwork ablation, semantic events triggered requests before the next scheduled audit. Across three emulated network profiles, the method continued to reduce cloud traffic, although lane changes took longer than with periodic access. Onboard traffic assessment therefore served as a practical trigger for selective VLM inference in these experiments.

Meng Ma, Shuyang Li, Naigang Wang +1

Rensselaer Polytechnic Institute, Troy, NY, USA · IBM T. J. Watson Research Center, USA

VLMs for Autonomous Driving Edge Computing Efficient VLM Inference Autonomous Driving Planning

Ask PDF

Abstract

Computation Offloading

Explore similar work

Adversarial Robustness Analysis of Cloud-Assisted Autonomous Driving Systems

Multi-Resolution End-to-End Deep Neural Network for Optimizing Latency-Accuracy Tradeoff in Autonomous Driving

Risk-Adaptive Edge--Cloud Visual Reasoning for Communication-Efficient Autonomous Driving