CP-CNN: Computational Parallelization of CNN-Based Object Detectors in Heterogeneous Embedded Systems for Autonomous Driving

• Published in: IEEE access Vol. 11; pp. 52812 - 52823
• Main Authors: Chun, Dayoung, Choi, Jiwoong, Lee, Hyuk-Jae, Kim, Hyun
• Format: Journal Article
• Language: English
• Published: Piscataway IEEE 2023; The Institute of Electrical and Electronics Engineers, Inc. (IEEE)
• Subjects: Accuracy; Algorithms; Artificial neural networks; Autonomous vehicle; Autonomous vehicles; convolutional neural network; Convolutional neural networks; Deep learning; Detectors; embedded platform; Embedded systems; Graphics processing units; Hardware; Low power electronics; low-power design; Machine learning; Object recognition; Optimization; Optimization techniques; Parallel processing; Platforms; Power efficiency; Processing speed; real-time system; Real-time systems
• ISSN: 2169-3536; 2169-3536
• DOI: 10.1109/ACCESS.2023.3280552

The success of research using convolutional neural network (CNN)-based camera sensor processing for autonomous driving has accelerated the development of autonomous driving vehicles. Since autonomous driving algorithms require high-performance computing for fast and accurate perception, a heterogeneous embedded platform consisting of a graphics processing unit (GPU) and a power-efficient dedicated deep learning accelerator (DLA) has been developed to efficiently implement deep learning algorithms in limited hardware environments. However, because the hardware utilization of these platforms remains low, performance differences such as processing speed and power efficiency between the heterogeneous platform and an embedded platform with only GPUs remain insignificant. To address this problem, this paper proposes an optimization technique that fully utilizes the available hardware resources in heterogeneous embedded platforms using parallel processing on DLA and GPU. Our proposed power-efficient network inference method improves processing speed without losing accuracy based on analyzing the problems encountered when dividing the networks between DLA and GPU for parallel processing. Moreover, the high compatibility of the proposed method is demonstrated by applying the proposed method to various CNN-based object detectors. The experimental results show that the proposed method increases the processing speed by 77.8%, 75.6%, and 55.2% and improves the power efficiency by 84%, 75.9%, and 62.3% on YOLOv3, SSD, and YOLOv5 networks, respectively, without an accuracy penalty.