Detection plays an important role in improving the performance of multi-object tracking (MOT), but most recently MOT works mainly focus on association algorithm and usually ignore the detections. To assist in associating object detections and to overcome detection failures, in this paper, we explore the low-rank-based foreground detection method to refine the detections and show it can significantly lead a better tracking result in online multi-object tracking. Firstly, the low-level pixel information from low-rank foreground segmentation and high-level detection responses from object detector are combined to form an overcomplete detections set, which serves as input for the tracking-by-detection-based multi-object tracking. Then, the predicted object location in online tracking as a prior to feedback for the foreground segmentation in sparse approximation for future frames can improve the foreground detection performance. Finally, to effectively solve the data association problem in online MOT, two-step data association relies on tracklet confidence is used to associate the detections and generate long trajectories since the existing trajectories provide a reliable history to support their presence in current frame. The experimental results in public pedestrian tracking datasets show that our detection optimization strategy can help to improve the tracking performance compared with several state-of-the-art multi-object trackers, with improved recall, precision, FP, FN and MOTA, MOTP results.