Object Detection in Autonomous Vehicles: A Performance Analysis

Automotive manufacturers are investing in smart cars, driverless processes, and pre-collision technologies to deliver safe and fuel-efficient mobility solutions. This paper aims to develop an advanced object detection model which incorporate more sophisticated techniques or architectures to identify commonly found objects on roads using one-stage (Single-Shot Detector) and two-stage object detectors (Faster R-CNN). The models are trained using TensorFlow and the Open Image Dataset V7, with ResNet50 architecture as the backbone. The Faster R-CNN models outperformed the SSD models, achieving a mean average precision of 92.5% at IoU 0.50. This paper presents the theory, methodology, and performance results of the work, which focuses on five classes of objects commonly found on roads (bus, car, motorcycle, stop sign, truck). The detection model exhibited an overall good performance in terms of precision and recall in object classification and localization, as evaluated from TensorBoard. The average precision value is 0.925, indicating accurate object classification. The recall value is 0.727, demonstrating successful object localization.