FAuto: An Efficient GMM-HMM FPGA Implementation for Behavior Estimation in Autonomous Systems

Driving behavior estimation in car-following scenario based on contextual traffic information is an essential capability for autonomous driving systems. Real-time motion planning based on incomplete environment perception requires complicated probabilistic model for interactions with surrounding objects and road conditions. Hidden Markov Model (HMM) with Gaussian emissions has been used to model driving behaviors for its ability of inferring unobserved states. While the high-dimensional contextual data is continuously processed, the system should be high-performance and power-efficient to make real-time decisions for safe operations. Field Programmable Gate Array (FPGA) is being increasingly used on embedded System-on-Chip (SoC) for mobile applications mainly because of its parallel computation and low-power consumption. This paper implements FAuto: the framework of HMM coupled with GMM algorithm on a Xilinx PYNQ-Z2 board for autonomous systems. We design the hybrid GMM-HMM model in python, and train the model using Next Generation SIMulation (NGSIM) trajectory data on a CPU platform. The hardware accelerator is designed through Vivado HLS 2018.2, and verified with Jupiter notebook. FAuto achieves 2.59 TOPS/W power efficiency, and 10.39x speedup compared to Python software implementation running on quad-core i7-7500U CPU.

[1]  Kamal Haddad,et al.  An FPGA-Accelerated Design for Deep Learning Pedestrian Detection in Self-Driving Vehicles , 2018, ArXiv.

[2]  Saeed Setayeshi,et al.  FPGA implementation of a linear systolic array for speech recognition based on HMM , 2010, 2010 The 2nd International Conference on Computer and Automation Engineering (ICCAE).

[3]  Wei Zhan,et al.  Probabilistic Prediction of Vehicle Semantic Intention and Motion , 2018, 2018 IEEE Intelligent Vehicles Symposium (IV).

[4]  Mark Horowitz,et al.  1.1 Computing's energy problem (and what we can do about it) , 2014, 2014 IEEE International Solid-State Circuits Conference Digest of Technical Papers (ISSCC).

[5]  Hiren M. Mandalia,et al.  Using Support Vector Machines for Lane-Change Detection , 2005 .

[6]  Mohamed Tarbouchi,et al.  FPGA Implementation of Genetic Algorithm for UAV Real-Time Path Planning , 2009, J. Intell. Robotic Syst..

[7]  Abbes Amira,et al.  An Efficient FPGA Implementation of Gaussian Mixture Models-Based Classifier Using Distributed Arithmetic , 2006, 2006 13th IEEE International Conference on Electronics, Circuits and Systems.

[8]  Morgan McGuire,et al.  Introduction to real-time ray tracing , 2019, SIGGRAPH Courses.

[9]  Douglas A. Reynolds,et al.  Robust text-independent speaker identification using Gaussian mixture speaker models , 1995, IEEE Trans. Speech Audio Process..

[10]  Moriyoshi Ohara,et al.  A power-efficient FPGA accelerator: Systolic array with cache-coherent interface for pair-HMM algorithm , 2016, 2016 IEEE Symposium in Low-Power and High-Speed Chips (COOL CHIPS XIX).

[11]  James H. Martin,et al.  Speech and Language Processing: An Introduction to Natural Language Processing, Computational Linguistics, and Speech Recognition , 2000 .

[12]  Karl-Heinz Hoffmann,et al.  Prediction of driver intended path at intersections , 2014, 2014 IEEE Intelligent Vehicles Symposium Proceedings.

[13]  Alejandro Duran,et al.  Heterogeneous Streaming , 2016, 2016 IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW).

[14]  Tobi Delbrück,et al.  DeltaRNN: A Power-efficient Recurrent Neural Network Accelerator , 2018, FPGA.

[15]  Qin Lin,et al.  Lane-Change Intention Estimation for Car-Following Control in Autonomous Driving , 2018, IEEE Transactions on Intelligent Vehicles.

[16]  Ronald G. Dreslinski,et al.  Sirius: An Open End-to-End Voice and Vision Personal Assistant and Its Implications for Future Warehouse Scale Computers , 2015, ASPLOS.