DetTrack: Detect Target from Local Region for 3D Single Object Tracking in Point Clouds

Abstract

3D single object tracking (SOT) is an essential task in computer vision, widely applied in autonomous driving and robotics. Current Siamese-based and motion-centric methods often require additional parameters and specialized architectures, resulting in high computational demands and failing to fully utilize advancements in 3D object detection. This work proposes that 3D SOT can be reformulated as sequential detection within localized search regions. Based on this idea, a novel framework, DetTrack, is introduced with two core modules: a search region estimator, which identifies candidate regions containing the target, and a modular 3D detector switcher, integrating pretrained detectors for precise localization. DetTrack simplifies the tracking process, avoids specialized tracking architectures, and significantly reduces computational demands. Its modularity enables seamless integration of state-of-the-art 3D detectors, enhancing adaptability and future applicability.