Invited Paper: Design of an RRAM In-Memory Computing Scheme for Target Tracking Applications

In this paper, we present a resistive random-access memory (RRAM) in-memory computing system for object tracking applications. The system includes a $\mathrm{TaO}_{\mathrm{x}}$ RRAM array for storing algorithm data, capable of supporting both binary and multi-level weight storage. We have also designed digital control circuitry based on the RRAM array to interface with a vision sensor chip and the in-memory computing module. By performing multiply-accumulate (MAC) operations under two different weight storage modes, we obtain the standard deviation from the ideal MAC values and introduce Gaussian noise into the SiamFC object tracking algorithm. The OTB2015 dataset is used to evaluate the two schemes. The results indicate that the algorithm, without error introduction at the software end, achieves an precision of 0.783 and a success rate of $\mathbf{0. 5 8 3}$. When the errors from hardware MAC are introduced, the precision and success rate are 0.781 and 0.582 in the best case, respectively, indicating negligible impact on the algorithm’s performance.