Removing ghost markers in trinocular optical tracking using convolutional neural network

Surgical instruments are tracked by the near-infrared binocular optical tracking system through the locating of reflective markers. However, during binocular optical tracking, misidentified ghost markers may appear, significantly affecting the accuracy of surgical tracking. In this paper, a trinocular optical tracking system (TOTS) is constructed, and an intelligent method is proposed to remove these misidentified ghost markers by utilizing the color image features of reflective markers. A convolutional neural network is designed to recognize markers, achieving a recognition accuracy of 97.9% through training and testing on a homemade dataset. Subsequently, a data transmission architecture is designed to deploy the network on the TOTS's field programmable gate array, optimized through 12-bit integer quantization, resource multiplexing, and parallel processing. The final recognition accuracy reaches 95.0%, with an average recognition time of 0.231 ms and an on-chip power consumption of 3.327 W. The CNN's processing time is 0.110 ms, which is 2.45 times faster than that on a graphics processing unit and 9.31 times faster than that on a central processing unit. Compared to other methods presented in the relevant literature, this method offers a faster rate and broader applicability in the removal of ghost markers.