Road object detection: a comparative study of deep learning-based algorithms

• Published in: Multimedia tools and applications 2022-04, Vol.81 (10), p.14247-14282
• Main Authors: Mahaur, Bharat, Singh, Navjot, Mishra, K. K.
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Autonomous vehicles; Comparative studies; Computer Communication Networks; Computer Science; Data Structures and Information Theory; Deep learning; Energy efficiency; Image segmentation; Intelligent transportation systems; Machine learning; Multimedia Information Systems; Object recognition; Pedestrian safety; Pedestrians; Real time; Special Purpose and Application-Based Systems; Target detection; Traffic signals; Traffic signs; Vehicles
• ISSN: 1380-7501; 1573-7721
• DOI: 10.1007/s11042-022-12447-5

Deep learning field has progressed the vision-based surround perception and has become the most trending area in the field of Intelligent Transportation System (ITS). Many deep learning-based algorithms using two-dimensional images have become an essential tool for autonomous vehicles with object detection, tracking, and segmentation for road target detection, primarily including pedestrians, vehicles, traffic lights, and traffic signs. Autonomous vehicles rely heavily on visual data to classify and generalize target objects which can satisfy pedestrians’ and other vehicles’ safety requirements in their environment. In real-time, outstanding results are obtained by deep learning-based algorithms for object detection. While several studies have thoroughly examined different types of deep learning-based object detection methods, there are a few comparable studies that either test the detection speed or accuracy of the object detection algorithms. In addition to speed and accuracy, autonomous driving also depends on model size and energy efficiency. However, there is a lack of comparison on various such metrics among existing deep learning-based methods. This article aims to provide a detailed and systematic comparative analysis of five independent mainstream deep learning-based algorithms for road object detection, namely the R-FCN, Mask R-CNN, SSD, RetinaNet, and YOLOv4 on a large-scale Berkeley DeepDrive (BDD100K) dataset. The experimental results are analyzed using the mean Average Precision (mAP) value and inference time. Additionally, various practical metrics, such as model size, computational complexity, and energy efficiency of deep learning-based models are precisely computed. Furthermore, the performance of each algorithm is evaluated under different road environmental conditions at various times of day and night. The comparison presented in this article helps to gain insight into the strengths and limitations of the popular deep learning-based algorithms under practical constraints with their real-time deployment feasibility. Code is publicly available at: https://github.com/bharatmahaur/ComparativeStudy