Real-time Object Detection and Associated Hardware Accelerators Targeting Autonomous Vehicles: A Review

2025-09-04

AI Summary

Real-time object detection for autonomous driving faces core challenges: the trade-off among accuracy, latency, and resource efficiency; hardware accelerator (GPU/ASIC) bottlenecks in multi-camera deployments; and a widening gap between academic research and industrial practice. This paper presents the first systematic survey of co-optimization strategies between AI detection models and edge hardware, integrating both publicly available algorithms and proprietary industry practices. We establish a comprehensive evaluation framework covering mainstream CNN architectures and heterogeneous acceleration platforms. A cross-layer co-design methodology is proposed to identify critical bottlenecks in multi-stream video processing and deliver transferable optimization strategies. Experimental results demonstrate that our approach enables inference at hundreds of frames per second, achieving high detection accuracy while significantly reducing end-to-end latency. The work provides a reproducible technical reference and a practical deployment bridge toward highly reliable, ultra-low-latency fully autonomous driving systems.

Technology Category

• Machine Learning: Hardware-aware ML
• Computer Vision: Learning & Optimization for CV
• Search and Optimization: Learning to Search

Application Category

• Systems and Infrastructure for Web, Mobile and WoT: Applied ML and AI for Web-based mobile applications
• Responsible Web: Human-perceived consequences of algorithmic deployment on the web
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

The efficiency of object detectors depends on factors like detection accuracy, processing time, and computational resources. Processing time is crucial for real-time applications, particularly for autonomous vehicles (AVs), where instantaneous responses are vital for safety. This review paper provides a concise yet comprehensive survey of real-time object detection (OD) algorithms for autonomous cars delving into their hardware accelerators (HAs). Non-neural network-based algorithms, which use statistical image processing, have been entirely substituted by AI algorithms, such as different models of convolutional neural networks (CNNs). Their intrinsically parallel features led them to be deployable into edge-based HAs of various types, where GPUs and, to a lesser extent, ASIC (application-specific integrated circuit) remain the most widely used. Throughputs of hundreds of frames/s (fps) could be reached; however, handling object detection for all the cameras available in a typical AV requires further hardware and algorithmic improvements. The intensive competition between AV providers has limited the disclosure of algorithms, firmware, and even hardware platform details. This remains a hurdle for researchers, as commercial systems provide valuable insights while academics undergo lengthy training and testing on restricted datasets and road scenarios. Consequently, many AV research papers may not be reflected in end products, being developed under limited conditions. This paper surveys state-of-the-art OD algorithms and aims to bridge the gap with technologies in commercial AVs. To our knowledge, this aspect has not been addressed in earlier surveys. Hence, the paper serves as a tangible reference for researchers designing future generations of vehicles, expected to be fully autonomous for comfort and safety.

Problem

• Reviewing real-time object detection algorithms for autonomous vehicles' safety
• Analyzing hardware accelerators like GPUs and ASICs for efficient processing
• Bridging the gap between academic research and commercial AV technologies

Innovation

• AI algorithms replace traditional image processing methods
• Convolutional neural networks enable edge-based hardware accelerators
• GPUs and ASICs widely used for real-time detection