Towards Real-Time Detection and Tracking of Spatio-Temporal Features: Rodent Whisker Tracking

The rodent whisker system is becoming a prominent experimental subject for the study of active sensing. As a result of improved video recording hardware/software and progressively better neurophysiological methods, there is now the prospect of precisely analyzing the intact vibrissal sensorimotor system. The vibrissae and snout analyzer (ViSA), as a widely used algorithm based on computer vision and image processing, has been proved successful for tracking and quantifying rodent sensorimotor behavior, but at great cost of processing time and power. In order to accelerate this offline algorithm and eventually apply it in online tracking, optimization is essential on both software and hardware level. We implemented 6 different versions of the algorithm, thereby exploring a wide range of diverse high-performance approaches including GPU parallelization, multi-thread computing, and FPGA dataflow computing on the optimized algorithm. Our experimental results indicate that the optimal solution for an offline implementation of ViSA is the mutli-threading execution. By using 16 CPU threads, we achieved x4,767 speed-up compared to the original Matlab serial version, resulting in a processing speed of 1.2 ms/frame, which is a solid step towards the real-time tracking. Analysis shows that running the algorithm on a 32-thread enabled machine reduces this number to 0.72 ms/frame, thereby enabling 1 KHz real-time tracking.Electrical Engineering, Mathematics and Computer ScienceQuantum EngineeringComputer Engineerin