Small-object detection based on YOLOv5 in autonomous driving systems

• We discuss the benefits of accurate detection of small objects like traffic signs and traffic lights in autonomous driving . • We analyze the practical limitations of the original YOLOv5 structure. • We propose novel architectural refinements to the same for improving its performance in the detection of small objects. • We perform extensive experimentation over the BDD100K, TT100K, and DTLD datasets. • We further evaluate the generalization ability of the proposed iS-YOLOv5 model in different road weather conditions. With the rapid advancements in the field of autonomous driving, the need for faster and more accurate object detection frameworks has become a necessity. Many recent deep learning-based object detectors have shown compelling performance for the detection of large objects in a variety of real-time driving applications. However, the detection of small objects such as traffic signs and traffic lights is a challenging task owing to the complex nature of such objects. Additionally, the complexity present in a few images due to the existence of foreground/background imbalance and perspective distortion caused by adverse weather and low-lighting conditions further makes it difficult to detect small objects accurately. In this letter, we investigate how an existing object detector can be adjusted to address specific tasks and how these modifications can impact the detection of small objects. To achieve this, we explore and introduce architectural changes to the popular YOLOv5 model to improve its performance in the detection of small objects without sacrificing the detection accuracy of large objects, particularly in autonomous driving. We will show that our modifications barely increase the computational complexity but significantly improve the detection accuracy and speed. Compared to the conventional YOLOv5, the proposed iS-YOLOv5 model increases the mean Average Precision (mAP) by 3.35% on the BDD100K dataset. Nevertheless, our proposed model improves the detection speed by 2.57 frames per second (FPS) compared to the YOLOv5 model.