BoostTrack++: using tracklet information to detect more objects in multiple object tracking

Vukašin Stanojević,Branimir Todorović

2024-08-23

Abstract:Multiple object tracking (MOT) depends heavily on selection of true positive detected bounding boxes. However, this aspect of the problem is mostly overlooked or mitigated by employing two-stage association and utilizing low confidence detections in the second stage. Recently proposed BoostTrack attempts to avoid the drawbacks of multiple stage association approach and use low-confidence detections by applying detection confidence boosting. In this paper, we identify the limitations of the confidence boost used in BoostTrack and propose a method to improve its performance. To construct a richer similarity measure and enable a better selection of true positive detections, we propose to use a combination of shape, Mahalanobis distance and novel soft BIoU similarity. We propose a soft detection confidence boost technique which calculates new confidence scores based on the similarity measure and the previous confidence scores, and we introduce varying similarity threshold to account for lower similarity measure between detections and tracklets which are not regularly updated. The proposed additions are mutually independent and can be used in any MOT algorithm. Combined with the BoostTrack+ baseline, our method achieves near state of the art results on the MOT17 dataset and new state of the art HOTA and IDF1 scores on the MOT20 dataset. The source code is available at: <a class="link-external link-https" href="https://github.com/vukasin-stanojevic/BoostTrack" rel="external noopener nofollow">this https URL</a> .

Computer Vision and Pattern Recognition,Artificial Intelligence

What problem does this paper attempt to address?

This paper attempts to solve the problem of handling low - confidence detections in multi - object tracking (MOT), especially how to select true - positive detections more effectively in a single - stage association framework. Specifically, the paper proposes an improvement scheme for the limitations of the confidence - boosting technique in the BoostTrack method to improve the performance of multi - object tracking and reduce the number of identity switches (IDSWs) and new IDs.

### Background and Problem Description of the Paper

In multi - object tracking tasks, selecting the correct true - positive detection bounding boxes is crucial for the tracking effect. However, many existing methods have deficiencies in handling low - confidence detections. For example, although the two - stage association method can utilize low - confidence detections to a certain extent, it is also prone to introducing the identity - switching problem. BoostTrack attempts to avoid these problems through single - stage association and detection confidence - boosting, but there are still some limitations in practical applications:

1. **Limitations of Confidence - Boosting**: The DLO (Detection Likelihood Object) confidence - boosting technique in BoostTrack depends on the IoU threshold, but this method may incorrectly increase the confidence of false positives and requires hyper - parameter adjustment for different datasets.
2. **Singularity of Similarity Measurement**: Using only IoU as the similarity measurement standard may lead to mis - matching and identity - switching problems.
3. **Irrationality of Fixed Threshold**: The fixed similarity threshold cannot adapt to the tracking requirements in different scenarios. Especially when the tracklet has not been updated for a long time, its prediction quality will decline, resulting in a decrease in the IoU value.

### Solutions

To overcome the above problems, the paper proposes the following improvement measures:

1. **Soft - Buffered IoU (Soft BIoU)**:
- A new soft - buffered IoU (Soft BIoU) similarity measurement method is introduced. By dynamically adjusting the expansion ratio of the bounding box according to the confidence of the tracklet, the problem that the BIoU method can only be used for two - stage association is solved.
- The formula is:
\[
SBIoU(D_i, T_j) = IoU\left(D_i \rightarrow 1 - \frac{c_{tj}}{4}, T_j \rightarrow 1 - \frac{c_{tj}}{2}\right)
\]
where \( c_{tj} \) is the confidence of tracklet \( T_j \).

2. **Improved Similarity Measurement**:
- A combination of Mahalanobis distance, shape similarity and soft BIoU is used to calculate a more complex similarity measurement, so as to better select true - positive detections with low confidence.
- The similarity formula is:
\[
S(D_i, T_j) = \frac{SBIoU(D_i, T_j) + SMhD(D_i, T_j) + Sshape(D_i, T_j)}{3}
\]

3. **Soft Detection Confidence - Boosting**:
- A soft detection confidence - boosting method is proposed, which takes into account the original detection confidence score and avoids treating low - confidence and high - confidence detections equally.
- The new confidence score calculation formula is:
\[
\hat{c}_{di} = \max\left(c_{di}, \alpha \cdot c_{di}+(1 - \alpha) \cdot \left(\max_j(S(D_i, T_j))^q\right)\right)
\]
where \( \alpha \in [0, 1] \) and \( q \geq 1 \) are hyper - parameters.

4. **Variable Similarity Threshold**:
- A variable similarity threshold based on the number of frames since the tracklet was last updated is introduced to adapt to the changes in the prediction quality of tracklets in different situations.
- As the time that the tracklet has not been updated increases, the IoU value will gradually decrease, so it is necessary to adjust the similarity threshold to ensure accurate matching.

Through these improvements, the method proposed in the paper has achieved close or...