Autonomous vehicle research has been rapidly developing because it improves transportation efficiency and reduces vehicle crash accidents due to human errors as well. Among all the studies in autonomous vehicles, object detection and perception of the surrounding environment is one of the most critical study areas for autonomous vehicles’ driving safety. Numerous high-precision neural network-based models have been developed for object detection tasks in computer vision with the availability of increasing computational speed. Nevertheless, autonomous vehicles' requirements for object detection algorithms are far more than detection accuracy and precision. Real-world driving scenarios such as bad weather, sunlight reflection, vehicle vibration, among others, can adversely affect the detection algorithms' robustness and reliability. This paper presents a comprehensive comparison among four popular algorithms: (a). Single Shot Multibox Detector (SSD), (b). Deconvolutional Single Shot Multibox Detector (DSSD, (c). faster regions with convolutional neural network (R-CNN), and (d). Yolo-V4 for object detection under real driving conditions. The reasons to select these four algorithms are because they are (a) popular state-of-art algorithms (SOTA) that achieve high multiple objects detection accuracies and precisions, (b) able to be processed at real-time speed (at least faster than 30 fps), and (c) widely applied at real-world conditions.Each neural network model is initially trained with the same data sets, the 80-classes COCO dataset, then fine-tuned by the 10-classes BDD100K dataset, to compare the algorithms equally. The former dataset contains more object types (human, animal, electric appliance, etc.), and is commonly used for general object detection training, while the latter is specifically designed for real-world naturalistic driving studies with only traffic-related object types (car, bicycle, traffic lights, etc.). When the model is trained by the COCO dataset, all neural network model parameters can be updated while the model is trained by the BDD dataset, the feature extraction layers’ parameters are set to be fixed, and only the classification layers’ parameters are allowed to be updated. After training all networks, the overall mean average precision (mAP), video running speed (fps), mAPs on different weather conditions (snow, rain, and fog, etc.), mAPs at a different time (dawn, afternoon, and evening, etc.), and small object detection accuracies are evaluated and compared. Finally, all the networks' performance and future object detection algorithm prospects are illustrated with conclusions on preferred models. The study will guide future research on the evaluated NN-based objection detection algorithms' pros and cons.Keywords: Autonomous Vehicle, Neural Network, Object Detection, Deep Learning

A Comparative Analysis of Object Detection Algorithms in Naturalistic Driving Videos_x000B_

Paper Proceedings: https://asmedigitalcollection.asme.org/IMECE/IMECE2021/volume/85543

ASME 2021 International Mechanical Engineering Congress and Exposition (IMECE2021)

Supporting the need among researchers worldwide for additional online and multimedia content, ASME now offers the ASME Conference Video Collection. This platform will collect and provide long-term access to conference presentation videos, encompassing all technology areas within ASME's technical conference and events program. About 30,000-40,000 engineers across all mechanical engineering and related disciplines participate in ASME conferences each year. This includes authors of approximately 6,000 to 8,000 technical papers published by ASME, the majority of whom also submit video presentations for our virtual conferences. The video collection will be available as an add-on to existing product packages or as a separate subscription.