Existing multi-object tracking methods encounter significant challenges when dealing with occluded objects. To address the issue of ID switching caused by occlusions, these methods often heavily rely on motion prediction, which can lead to cumulative errors. Additionally, they typically utilize re-identification (reID) networks to extract appearance features for association, but this heavy reliance on training data and computational power is a major limitation. This study presents a fast multi-object tracking method that is robust to occlusions, which employs BEV bounding boxes estimated using geometric constraints. We introduce a novel algorithm that leverages homography transformation and geometric priors to quickly estimate target 3D bounding boxes through optimization. Furthermore, we enhance ByteTrack by integrating the reprojection of 3D bounding boxes into the association process, effectively mitigating the influence of occlusions on MOT. Moreover, this study reduces the frequency of using computationally expensive object detection by employing sparse optical flow for fast tracking. The experimental findings indicate that our method attains enhanced accuracy and faster execution times compared to existing multi-object tracking methods.