Preventing the Impact of False Negatives on Vehicle Object Detection in Autonomous Driving A Thorough Analysis of Calibration, Thresholding, and Fusion Methods

Abstract

The surge in the development and deployment of autonomous vehicles (AVs) in recent years has been underpinned by their ability to effectively use sensors and algorithms to understand and navigate their surroundings. One of the foundational components of this system is object detection, which identifies other vehicles, pedestrians, and obstacles. However, a persistent challenge with these systems is the occurrence of false negatives — scenarios where the system overlooks real objects. This not only undermines the reliability of AVs but can also lead to potential safety hazards. Our research undertook a comprehensive study of methodologies aimed at minimizing the impact of these false negatives. Calibration emerged as a prime solution. Through calibration, we can adjust the system's predictions to align more closely with real-world probabilities. Techniques such as Platt Scaling and Isotonic Regression were evaluated in depth. Their purpose is to finetune the outputs of the detection algorithms, thereby providing more accurate probabilities of object presence. Another pivotal strategy we delved into is thresholding. Here, specific limits or boundaries are set, determining when an object is considered detected by the system. The setting of these boundaries is critical, as they can influence the rate of false detections. Our exploration spanned various techniques of thresholding, especially focusing on their applicability in diverse driving environments, from congested urban settings to open highways. We investigated sensor fusion methods. Given that AVs utilize a myriad of sensors — from cameras to LIDAR — effectively combining their outputs can lead to enhanced detection accuracy. We evaluated methodologies for integrating this multifaceted data. Implementing a combination of these techniques can substantially boost the reliability and safety of autonomous driving systems. The road ahead necessitates continuous refinement of these strategies, adapting to evolving real-world conditions and technological advancements.