Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences (2103.06028v2)

Abstract: Estimating the states of surrounding traffic participants stays at the core of autonomous driving. In this paper, we study a novel setting of this problem: model-free single-object tracking (SOT), which takes the object state in the first frame as input, and jointly solves state estimation and tracking in subsequent frames. The main purpose for this new setting is to break the strong limitation of the popular "detection and tracking" scheme in multi-object tracking. Moreover, we notice that shape completion by overlaying the point clouds, which is a by-product of our proposed task, not only improves the performance of state estimation but also has numerous applications. As no benchmark for this task is available so far, we construct a new dataset LiDAR-SOT and corresponding evaluation protocols based on the Waymo Open dataset. We then propose an optimization-based algorithm called SOTracker involving point cloud registration, vehicle shapes, correspondence, and motion priors. Our quantitative and qualitative results prove the effectiveness of our SOTracker and reveal the challenging cases for SOT in point clouds, including the sparsity of LiDAR data, abrupt motion variation, etc. Finally, we also explore how the proposed task and algorithm may benefit other autonomous driving applications, including simulating LiDAR scans, generating motion data, and annotating optical flow. The code and protocols for our benchmark and algorithm are available at https://github.com/TuSimple/LiDAR_SOT/. A video demonstration is at https://www.youtube.com/watch?v=BpHixKs91i8.

• The paper presents a novel model-free tracking method that bypasses traditional detection-based approaches.
• It employs an optimization-based algorithm (SOTracker) leveraging LiDAR geometry to accurately estimate vehicle position and orientation.
• Results on the LiDAR-SOT dataset demonstrate superior tracking performance in both simulated and real-world autonomous driving scenarios.

Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences

The paper "Model-free Vehicle Tracking and State Estimation in Point Cloud Sequences" presents a novel approach to single-object tracking (SOT) that eschews traditional model-based methods and instead leverages point cloud data, particularly in autonomous driving contexts. This work offers a distinct methodology focused on model-free tracking, addressing some inherent limitations of the prevalent detection-and-tracking paradigm predominantly adopted in multi-object tracking (MOT).

Overview and Methodology

In the domain of autonomous driving, accurate state estimation—such as determining the position and orientation—of surrounding vehicles is crucial. Traditional techniques heavily rely on detection followed by tracking, a process which inherently depends on model-based recognition systems trained on extensive datasets. While effective, this strategy often succumbs to inaccuracies in sparse point clouds or when faced with novel objects that the model has not been exposed to during training. The authors propose an alternative with model-free SOT, which leverages vehicle state information from an initial frame and progressively resolves both tracking and state estimation in succeeding frames through point cloud sequences.

This approach involves no dedicated model for classification but instead uses an optimization-based algorithm, referred to as SOTracker, which works directly with raw LiDAR data. The algorithm is built around point cloud registration, leveraging the geometric integrity of LiDAR data to ascertain vehicle states and improve tracking fidelity. The paper introduces a new dataset called LiDAR-SOT, drawn from the Waymo Open Dataset, to benchmark model-free tracking performance.

Results and Implications

Quantitative evaluations demonstrate the capability of SOTracker in both simulated and real-world scenarios. The results highlight its robustness and accuracy, especially notable in challenging conditions such as sparse point cloud data or abrupt vehicle motions. The research points out that SOTracker achieves superior tracking accuracy compared to conventional methods, inferring a promising potential for applications requiring high precision, such as in optical flow annotation and motion data generation.

Another significant advantage of this model-free approach is its potential versatility. Beyond immediate tracking tasks, the technique supports a broader range of applications, including generating training datasets for model-based methods, simulating LiDAR scans, and enhancing vehicle shape libraries. Through dynamic point cloud aggregation, SOTracker also facilitates better shape completion—benefiting tasks like optical flow annotation where precise motion capture is paramount.

Future Directions

This work opens several avenues for future exploration. Extending this model-free framework to encompass a wider variety of object types, including non-rigid entities like pedestrians and cyclists, remains a key challenge. There is also potential in refining the integration between model-free and model-based techniques to enhance overall tracking solutions by combining the strengths of both methodologies.

Moreover, addressing limitations such as point cloud sparsity and reducing drift due to prolonged exposure are critical for real-world applications. As autonomous systems advance, the ability to accurately infer motion under varied environmental conditions will be critical.

Overall, this paper's contribution lies in its innovative model-free tracking approach which, through precise state estimation and point cloud utilization, offers a robust alternative to traditional object tracking methodologies in autonomous vehicle systems. The results are instrumental in advancing vehicular automation technologies and hold promise for further research into more adaptable and scalable tracking systems.