Prior Knowledge-Based Optimization Method for the Reconstruction Model of Multicamera Optical Tracking System.
Houde Dai,Yadan Zeng,Zengwei Wang,Haijun Lin,Mingqiang Lin,Hui Gao,Shuang Song,Max Q. -H. Meng
DOI: https://doi.org/10.1109/tase.2020.2989194
IF: 6.636
2020-01-01
IEEE Transactions on Automation Science and Engineering
Abstract:The optical tracking system (OTS) plays a vital role in the computer-assisted surgical navigation process, whereas the performance of the commonly used binocular stereo vision is affected by the line-of-sight problem and limited workspace. Thus, this article proposed a prior knowledge-based multicamera reconstruction model (PKRM) to both expand the tracking workspace and improve the tracking robust and computational efficiency of OTS when working in unstructured clinical conditions. This reconstruction model inherits the advantages of the geometrical method, data-driven method, and gating technique (GT). First, we added the geometric principle as the prior knowledge to optimize the training of the multicamera OTS reconstruction model through the Lagrange multiplier method; hence, the prior knowledge feedforward NN (PKFNN) was built. Second, besides the training features, the state of camera (SOC) was extracted in advance to determine the NN structure using GT. According to the SOC feature, the OTS can be self-adaptive to the changing field of view (FOV) caused by optical occlusion, which is frequently occurred in surgery. Furthermore, experiments were carried out to verify the performance of the proposed model, whose accuracy and runtime performed 0.4627 mm and 0.0016 ms, respectively. Results demonstrate that the proposed reconstruction model can achieve higher accuracy and computational efficiency than both the geometrical model and the data-driven model. Especially, by considering SOC as the state prior knowledge, the tracking robustness is enhanced when one or two of the four cameras are not working properly. Note to Practitioners —The original motivation for this article derives from both the line-of-sight limitation and robust demand for optical tracking of surgical instruments. The performance of the multicamera optical tracking system (OTS) depends on its reconstruction model. However, the geometric reconstruction model requires more calculation to obtain high accuracy, which will enlarge the latency and reduce the update rate. In our previous work, the reconstruction model based on the neural network (NN) has achieved accurate tracking in real-time, while the training of the model tends into local optimal values. Hence, we proposed the prior knowledge feedforward NN model to improve the accuracy and computational efficiency. Moreover, to guarantee the line-of-sight in the optical occlusion, the state of camera combining with the gating technique enables the OTS to be self-adaptive for changing the field of view, which greatly ensures the robust tracking process with larger workspace in case of line-of-sight obstructions.