Motion Parameter Optimization Using Modified Epipolar Geometric Constraints for Efficient Object Tracking

As the practical implementation of autonomous driving technology accelerates, the ability to accurately recognize surrounding objects during driving and to stably track their movements have become essential. In particular, 3D multi-object tracking (3D MOT) is a core technology for safe path planning and collision avoidance in autonomous vehicles, because it is important to not only identify the current positions of objects but to accurately understand their trajectories. Recently, camera-based 3D MOT has been studied as a cost-effective alternative to expensive LiDAR sensors. However, this approach tends to rely heavily on deep learning detection models, resulting in a significant degradation of tracking continuity in environments with limited computational resources, or in situations where detection fails. To solve that degradation problem, this study proposes a geometry-based motion optimization method that can predict object movement to enable continuous tracking without deep learning inferences. The proposed system operates by optimizing object acceleration and yaw angle in the prediction stage of the extended Kalman filter (EKF) based on the 3D bounding box and the constant turn rate and acceleration (CTRA) model, which reflect the physical characteristics of vehicles. In particular, by introducing a line sampling technique that utilizes segments inside the bounding box and corresponding points in the image—based on an extended form of epipolar geometry—the geometric validity of motion prediction is secured. The optimization process is based on quadratic programming, ensuring both numerical stability and computational efficiency, and the generated pseudo-detection can be used as a substitute observation, even in time intervals where actual detection does not occur. Experiments using the nuScenes dataset evaluated the performance of the proposed method under various detection interval conditions. As a result, while fast, poly-based tracking systems show increased tracking fragmentation and error as the detection interval grows longer, the proposed method demonstrates more stable tracking performance in terms of six metrics. In particular, we quantitatively and qualitatively confirm that the proposed method predicts object motion in near real time, even in frames where deep learning detection is absent. This study presents a practical alternative that ensures both real-time performance and reliability from camera-based 3D MOT, and demonstrates its applicability in real-time autonomous driving systems where using high-performance GPUs or continuous deep learning inference is not feasible. Furthermore, it proves that high-performance tracking can be achieved without relying heavily on appearance features, but by using only geometric constraints and a physics-based prediction model. Future research plans will expand the proposed framework into a more generalized 3D MOT system by applying various motion models tailored to different object motion characteristics beyond the currently used CTRA model.