Random Finite Set Based Bayesian Filtering with OpenCL in a Heterogeneous Platform

Biao Hu, Uzair Sharif, Rajat Koner, Guang Chen, Kai Huang, Feihu Zhang, Walter Stechele, Alois C. Knoll

TL;DR

This paper tackles the computational bottleneck in advanced multi-target tracking using Random Finite Sets (RFS) by leveraging heterogeneous computing platforms with OpenCL. They demonstrate a significant speedup (up to 18x in worst-case scenarios) for their implemented SMC-LMB tracker, enabling real-time pedestrian tracking with maintained accuracy.

What Problem Does This Paper Address?

Multi-target tracking (MTT) is crucial for applications like autonomous driving and surveillance, aiming to simultaneously track multiple objects whose number is unknown and changes over time. Traditional methods struggle with data association uncertainty, detection noise, and false alarms. Random Finite Set (RFS) based filters, particularly Probability Hypothesis Density (PHD) and Labeled Multi-Bernoulli (LMB) filters, offer a principled probabilistic framework to handle these challenges. However, these advanced filters are computationally intensive, often hindering their application in real-time scenarios. This paper addresses this gap by exploring how to accelerate RFS-based filtering on heterogeneous computing platforms (CPUs and GPUs) using OpenCL to achieve real-time performance without sacrificing tracking accuracy.

Methods

The authors first implemented and evaluated two RFS-based filters: the Gaussian Mixture PHD (GM-PHD) filter and the Sequential Monte Carlo LMB (SMC-LMB) filter. They found the SMC-LMB filter to be more suitable for general pedestrian tracking due to its better accuracy, especially when dealing with non-linear/non-Gaussian motion.

The core of their contribution lies in accelerating the SMC-LMB filter using OpenCL. They identified the "update LMB" computation as the primary performance bottleneck, accounting for approximately 75% of the total computation time. To address this, they:

  • Transformed sequential code to OpenCL kernels: Key computationally intensive parts of the update LMB function were rewritten as parallelizable kernels for the GPU.
  • Leveraged GPU for random number generation: Used the AMD CLRNG compute library for efficient GPU-based random number generation.
  • Parallelized particle computations: Broke down for-loops within the update block to the level of individual particle computations.
  • Implemented efficient parallel scan: Utilized a prefix-sum algorithm on cumulative weights for particle resampling.
  • Optimized memory management: Focused on efficient data transfer between the CPU and GPU to minimize overhead.
  • Exploited vectorization and parallel threading: Used vector operations and assigned each Gaussian component to a separate thread to leverage parallel architectures.

Datasets

The paper utilizes two types of data for evaluation:

  • Simulated data: A 2D surveillance point-target simulation scenario was created to thoroughly evaluate tracking accuracy under various controlled conditions (number of targets, birth/death times, detection/survival probabilities, clutter, noise).
  • Publicly available dataset videos: The Multiple Object Tracking Challenge (MOT) benchmark [3] was used, specifically:
  • KITTI-17: Features a static camera with mostly linear target motion.
  • PETS09-S2L1: Features a static camera with targets moving in irregular patterns.

Key Insights and Ideas

  • SMC-LMB outperforms GM-PHD for general pedestrian tracking: While GM-PHD is efficient for linear/Gaussian systems, SMC-LMB demonstrates superior robustness and accuracy when dealing with the non-linear and non-Gaussian motion patterns typical of real-world pedestrian scenarios.
  • OpenCL is crucial for real-time RFS filtering: The significant computational demands of RFS filters necessitate parallelization. OpenCL provides an effective means to harness the power of heterogeneous platforms (CPUs + GPUs) to achieve real-time performance.
  • Targeting specific bottlenecks is key: Instead of a complete rewrite, focusing OpenCL acceleration on the most computationally intensive parts (like the "update LMB" step) yields substantial performance gains.
  • Memory management is critical for GPU acceleration: Efficient data transfer between the CPU and GPU is paramount; otherwise, the overhead can negate the benefits of GPU computation.
  • Scalability matters: The OpenCL implementation shows better scalability compared to the pure C++ version, with execution times rising less steeply as configurations change.

Key Results

  • Significant Speedup: The OpenCL accelerated SMC-LMB tracker achieved an average frame rate improvement from 15 fps to 100 fps (a 6x increase) on typical MOT benchmark videos.
  • Worst-Case Performance Improvement: The worst-case frame processing time saw an 18x advantage, improving from nearly 2 fps to 36 fps.
  • Accuracy Maintained: The acceleration was achieved without a significant degradation in tracking accuracy, as demonstrated by OSPA distance and cardinality estimates on both simulated and real-world datasets.
  • Comparison with GM-PHD: In simulations, the SMC-LMB tracker showed comparable or better performance than the GM-PHD tracker, especially in more challenging scenarios.
  • MOT Benchmark Performance: The tracker performed well on the KITTI-17 dataset (linear motion) and showed reasonable, though somewhat degraded, performance on the PETS09-S2L1 dataset (irregular motion), highlighting the limitations of its internal linear state-space models.

Limitations and Future Work

  • Linear State-Space Model Limitation: The current SMC-LMB implementation uses linear state-space models internally, which leads to performance degradation when targets exhibit highly irregular or non-linear motion, as observed in the PETS09-S2L1 dataset.
  • Potential for Further Improvement: The authors note that the separation of prediction and update steps in their current implementation reduces computational efficiency. They suggest that combining these steps into a single step, as proposed in recent work [5], could lead to further significant speedups (estimated 10x to 100x).
  • Future Work: The paper implies that exploring non-linear models within the SMC-LMB framework and further optimizing the prediction/update integration would be valuable future directions.

Summary updated

BibTeX
@inproceedings{hu2017random,
  author = {Biao Hu and Uzair Sharif and Rajat Koner and Guang Chen and Kai Huang and Feihu Zhang and Walter Stechele and Alois C. Knoll},
  title = {{Random Finite Set Based Bayesian Filtering with OpenCL in a Heterogeneous Platform}},
  booktitle = {Sensors},
  year = {2017},
  url = {https://openreview.net/forum?id=MZCdiZX3keB},
}
Figure