Optimizing Convolution Operations for YOLOv4-based Object Detection on GPU

¹ Université Paris-Saclay, ENS Paris-Saclay, CNRS, SATIE, 91190, Gif-sur-Yvette, France
² Systems Analysis, Information Processing and Industrial Management Laboratory, Higher School of Technology of Sale, Mohamed V University, Rabat, Morocco

^* e-mail: fatima-zahra.guerrouj@universite-paris-saclay.fr

Abstract

Real-time object detection is crucial for autonomous vehicles, and YOLO (You Only Look Once) algorithms have demonstrated their effectiveness for this purpose. This study examines the performance of YOLOv4 [3] for real-time object detection on an embedded architecture. We focus on optimizing the computationally intensive convolution operations by employing the cuDNN library to achieve efficient inference. The evaluation assesses critical performance metrics, including object detection accuracy in terms of Mean Average Precision (mAP) and inference latency on the embedded architecture. We conduct a comparative analysis using the publicly available KITTI [7] database. The reported results establish a benchmark between the parallelized YOLOv4 model and the baseline implementation, assessing the advantages of cuDNN acceleration for real-time object detection on resource-constrained devices.