Haoyang Li

Particle track reconstruction is an essential part of elementary particle experiments and is traditionally conducted by sophisticated statistical methods such as the combinational Kalman filter. Inspired by the successful performance of deep learning in computer vision, I propose a deep-learning approach which decomposes the track reconstruction mission into object detection and object segmentation. In this approach, particle hits in a detector are projected onto a plane. The object detection module samples areas from the projection plane, in which each area contains a track and segments from extraneous tracks. The object segmentation module is then applied to each sampled area to extract the intact track from other tracks’ segments. We evaluated the performance of this approach on the Mu2e experiment’s simulation data. We used a Faster Region-based Convolutional Neural Network (RCNN) as our object detection module, and we evaluated its performance along with several object segmentation models such as ResNet, U-Net, FC-DenseNet. We will discuss its strengths and limitations for particle track reconstruction.