9th International Conference on ICT for Sustainable Development

Authors - Srivatsa Kulkarni, Sneha Appasaheb Kale, Shweta Ramesh Sankri, Rohan Abbayya, Preeti Pillai, Nalini Iyer
Abstract - In the realm of autonomous vehicles, object detection assumes a pivotal role. Two predominant methods for object detection are prevalent: one-stage and two-stage object detection. In one-stage object detection, classification and bounding box predictions occur in a single step, while in two-stage object detection, these occur in two steps. The current state-of-the-art for one-stage object detection is YOLO v8, while for two-stage object detection, it is Faster R-CNN. Each architecture carries its own set of advantages and disadvantages. One-stage architectures exhibit quicker processing times but often compromise accuracy. Conversely, two-stage architectures demand more processing time but typically deliver superior accuracy. These distinctions stem from the differing approaches employed in object detection. Evaluation of custom datasets reveals that Faster R-CNN yields higher average precision, recall, and mean Average Precision (mAP), rendering it a more promising choice compared to YOLO v8. Consequently, amalgamating the strengths of both architectures could potentially birth a novel model that achieves enhanced accuracy in a shorter timeframe while maintaining robustness. Such advancements hold the promise of facilitating prompt decision-making by vehicles in traffic scenarios, thereby mitigating the occurrence of accidents.