Deep Learning Driven Object Detection and Classification for Autonomous Vehicles in Diverse Traffic and Weather Conditions
Jayant Singh Jhala, Chandani Joshi, Darpan Anand · 2024
The rapid development of self-driving vehicles necessitates integrating a sophisticated sensing system to address various obstacles posed by road traffic efficiently. While several datasets support object detection in autonomous vehicles, evaluating their suitability for different weather conditions globally is crucial. In this study, we present deep learning models trained on a novel dataset derived from YouTube videos recorded from Indian car’s dashcams. These videos capture a wide range of conditions, including rain, fog, daytime, hazy and night-time driving scenarios prevalent in India. The dataset comprises a total of 1450 annotated images depicting vehicles and other road assets across six different classes. In this work, performance analysis of the YOLOv8 models trained using an existing dataset was compared with the model trained on an expanded version using the proposed weather-specific dataset. The results demonstrate improved accuracy metrics of 91.3%, 84.5%, and 91.2% for Precision, Recall, and mean Average Precision (mAP) upon integrating the proposed dataset. The model trained on this diverse dataset exhibits heightened robustness, proving highly beneficial for autonomous and conventional vehicle operations in India’s dynamic traffic environments. This research contributes to advancing object detection capabilities crucial for autonomous driving technologies in real-world settings.