DeepOCSORT Tracking Method

• The paper introduces a robust extension to the Kalman filter by integrating dynamic appearance and adaptive weighting modules to mitigate tracking drift.
• DeepOCSORT achieves superior spatial precision with an ADE of 31.15 pixels compared to 114.3 pixels for ByteTrack on fast-moving racquetballs.
• Real-time efficiency is maintained at 26.8 ms per frame, despite challenges from erratic, non-linear object motion and frequent occlusions.

DeepOCSORT is a state-of-the-art tracking method designed to enhance the capabilities of Kalman filter–based multi-object trackers, specifically for the challenging setting of fast-moving tiny objects such as racquetballs. It operates within the predictive-update framework of Kalman filtering, incorporating advanced appearance modeling and adaptive weighting strategies to address track fragmentation and drift that arise from highly erratic, non-linear object motion. DeepOCSORT is distinguished by its superior trajectory accuracy among tested methods, though significant limitations remain in scenarios characterized by rapid, unpredictable movement and frequent occlusions. These challenges are especially pronounced in sport robotics and similar high-speed perception applications.

1. Methodological Foundations of DeepOCSORT

DeepOCSORT adopts the standard Kalman filter object tracking paradigm, which relies on a two-step process: prediction of the object state based on a linear motion model, followed by correction using new detections. The mathematical foundations for the prediction and update stages are as follows:

• Prediction step:

$\hat{x}_{k|k-1} = F_k \hat{x}_{k-1|k-1} + B_k u_k$

$P_{k|k-1} = F_k P_{k-1|k-1} F_k^T + Q_k$

• Update step:

$$K_k = P_{k|k-1} H_k^T (H_k P_{k|k-1} H_k^T + R_k)^{-1}$$

$$\hat{x}_{k|k} = \hat{x}_{k|k-1} + K_k (z_k - H_k \hat{x}_{k|k-1})$$

$P_{k|k} = (I - K_k H_k) P_{k|k-1}$

DeepOCSORT extends this framework with a Dynamic Appearance (DA) module that leverages feature-based appearance information to distinguish between similar small objects, as well as an Adaptive Weighting (AW) mechanism that modulates reliance on appearance features according to detector confidence. These enhancements address several weaknesses of basic Kalman filtering, including susceptibility to track confusion and rapid drift.

2. Evaluation Metrics and Accuracy

Tracking accuracy for deep object trackers is frequently assessed via the Average Displacement Error (ADE), which quantifies the mean Euclidean distance between predicted and ground-truth trajectory positions over time:

$$\mathrm{ADE} = \frac{1}{N_{\text{traj}}} \sum_{i=1}^{n} \sqrt{(x^{\text{pred}}_i - x^{\text{gt}}_i)^2 + (y^{\text{pred}}_i - y^{\text{gt}}_i)^2}$$

On a custom dataset containing 10,000 annotated frames of fast-moving racquetballs, DeepOCSORT achieved an average ADE of 31.15 pixels—the best among state-of-the-art Kalman-filter-based trackers evaluated. For comparison, ByteTrack recorded an ADE of 114.3 pixels under the same conditions.

Tracker	ADE (pixels)	Avg. Inference Time (ms)
DeepOCSORT	31.15	26.8
ByteTrack	114.3	26.6

This result highlights DeepOCSORT's improved spatial precision, although all tested methods exhibited tracking drift and spatial errors over the order of 3–11 cm—a substantial limitation relative to standard benchmarks on larger, slower-moving objects.

3. Kalman Filtering and Appearance Modeling

Kalman filtering underpins the prediction and correction of state estimates in DeepOCSORT. However, the method augments basic position-velocity state estimation by integrating appearance cues through the DA module. Appearance features, extracted from object detections, are dynamically incorporated to maintain consistent object identities across frames. Adaptive Weighting (AW) further refines the influence of these appearance cues, increasing their weight when detector confidence is high and reducing it in ambiguous or occluded scenarios.

This approach helps mitigate identity switches and provides some robustness to missed detections, though the core challenges associated with abrupt, non-linear object motion persist due to limitations of the underlying linear motion model.

4. Computational Performance and Trade-offs

In terms of real-time efficiency, DeepOCSORT delivers performance comparable to leading competitors, as evidenced by average inference times (26.8 ms per frame versus ByteTrack’s 26.6 ms). A key operational trade-off is between update frequency—how often the tracker incorporates new measurements—and overall inference speed. More frequent updates increase adaptability to sudden object motion but impose greater computational burden and elevate the risk of drift when detections are unstable or erroneous. This challenge is exacerbated for rapid, small objects, where visual ambiguity and missed detections are common.

5. Challenges and Limitations

Despite its enhancements, DeepOCSORT demonstrates fundamental limitations for fast-moving tiny objects. The principal challenge arises from the breakdown of linear assumptions in Kalman filters when objects undergo erratic, non-linear trajectories, sharp acceleration changes, and temporary occlusion. Fragmented trajectories and tracking drift persist, and the reported ADE values (31–114 pixels) translate to spatial errors substantially above those typical for larger or more predictable objects. This suggests that, while DeepOCSORT advances the state-of-the-art for this application domain, its accuracy is still limited by the core characteristics of fast-moving tiny object tracking.

6. Future Research Directions

Analysis indicates that current Kalman filter–based trackers, including DeepOCSORT, require substantial methodological advances to meet the demands of fast-moving tiny object scenarios. Hybrid approaches—combining robust object-specific detectors with tracking strategies that model object physics more realistically—offer a promising direction. For example, tracking algorithms that encode the expected bouncing dynamics of racquetballs could better predict non-linear trajectories.

Further improvements may be achievable through enhanced adaptive update strategies, systematic hyperparameter optimization, and deeper fusion of appearance and motion features—potentially by embedding physical constraints as trainable parameters in the detection architecture. The exploration of physics-inspired tracking frameworks is particularly noted as a prospective avenue for future breakthroughs in this challenging area.

7. Connections to Sport Robotics and Broader Impact

Precise tracking of fast-moving tiny objects has direct relevance for sport robotics, where lightweight, accurate tracking systems are essential for robust robot perception and action planning. The demonstrated limitations of DeepOCSORT and related Kalman filter–based methods highlight a key research gap, guiding future work towards more specialized techniques that can accommodate rapid, unpredictable dynamics and achieve error rates commensurate with industrial and athletic applications. The ongoing evolution of appearance modeling, dynamic weighting, and physics-aware frameworks will be critical to progress in this domain.