VOTA: A 2.45TFLOPS/W Heterogeneous Multi-Core Visual Object Tracking Accelerator Based on Correlation Filters

Abstract

VOTA is a domain-specific accelerator for correlation filter (CF)-based visual object tracking (VOT). It encompasses a Winograd convolution core, a FFT core and a vector core in a high-bandwidth starring topology. VOTA’s frame-based instructions and execution enable a 537GFLOPS performance and reduce the code size. An instruction-chaining mechanism permits inter-core pipelining to improve the utilization to 84.2%. A 10.2mm2 28nm FP16 VOTA prototype incorporating a RISC-V host CPU is measured to achieve 2.45TFLOPS/W at 0.72V. Running OPCF, a CF-based VOT enhanced by adaptive boosting and particle filters, the chip achieves 1157FPS on 640×480 input frames at 0.9V and 175MHz, consuming 296mW.