19 June 2024 in IEEE Transactions on Intelligent Vehicles
Automated Driving Systems (ADS) rely on a variety of sensors and algorithms to perceive their environment and make proper and timely driving decisions to ensure safety and efficiency. One of their critical tasks is object detection, which detects and classifies object presence with a specific emphasis on cars, pedestrians, and cyclists, in 3D space in their surroundings with the use of sensing data from those sensors, including LiDAR. 3D object detection in Autonomous Vehicles (AVs) is a challenging problem due to the complexity and variability of real-world scenarios, such as occlusions, varying lighting conditions, and diverse object shapes and sizes. Moreover, the limitations of computing power and the availability of information about the surrounding environment, combined with the quality and age of sensors installed in standalone AVs, often lead to poor perception performance. In this context, Connected and Autonomous Vehicles (CAVs) have become a promising solution that can be employed to harness connectivity and external information and enhance their perception capabilities. In this paper, we provide a review of existing 3D object detection techniques based on deep learning for ADS using data from LiDAR and other sensors, as well as the architecture of CAVs and their communications, which facilitate heterogeneous networks of mobile and satellite communication technologies. We also discuss real-world and synthetic datasets for both single-vehicle and multi-view vehicles, such as vehicle-to-vehicle (V2V) and vehicle-to-everything (V2X). Finally, we review the applications and future outlook of 3D object detection.
11 November 2024 in IEEE Access
RJRungroj Jintamethasawat
Pedestrian detection is fundamental in the realm of autonomous driving, relying on comprehensive datasets for accurate deep-learning methods. Detecting diverse pedestrian behaviors, including rare and near-accident cases, is necessary, however, there were limitations in the development of such specific datasets. Furthermore, involving humans in data collection is risky and raises privacy concerns since recording and storing data necessitates explicit consent from all participants. To tackle this challenge, we propose a synthetic dataset generated from the Metaverse, where the avatars can represent a wide range of pedestrian behaviors without exposing real individuals to potential harm or privacy violations. Our Metaverse includes both programmed pedestrian behaviors and controllable avatar movements and interactions, ensuring a comprehensive and realistic representation of pedestrian actions. To validate and ensure the generalization capabilities of our synthetic dataset’s performance in real-world scenarios, we perform both cross-dataset and intra-dataset evaluations on comprehensive real-world datasets, including KITTI, Citypersons, and INRIA Person. Our experiments demonstrate that the network trained on our synthetic dataset exhibits robust generalization capabilities for unseen real-world situations and our dataset’s inclusion significantly enhanced the Average Precision (AP) and Average Recall (AR) of the pedestrian detection model. For cross-dataset validation, we generally observe improvements in both AP and AR with models trained on all datasets. The inclusion of our dataset also improves the intra-dataset validation of the INRIA Person dataset, where the AP and AR values increase from 0.958 to 0.988 and AR from 0.697 to 0.725, respectively. For Citypersons and KITTI datasets, the AP values increase from 0.630 to 0.639 and 0.745 to 0.752, respectively. The model training on our dataset achieved the highest AP and AR values across all test datasets and the second-highest AP observed when tested on Citypersons.