Video-to-Model: Unsupervised Trace Extraction from Videos for Process Discovery and Conformance Checking in Manual Assembly

Manual activities are often hidden deep down in discrete manufacturing processes. To analyze and optimize such processes, process discovery techniques allow the mining of the actual process behavior. Those techniques require the availability of complete event logs representing the execution of manual activities. Related works about collecting such information from sensor data unobtrusively for the worker are rare. Papers either address the sensor-based recognition of activities or focus on the process discovery part using process mining-compatible data sets. This paper builds on previous works to provide a solution on how execution-level information can be extracted from videos in manual assembly. The test bed consists of an assembly workstation equipped with a single RGB camera. A neural network-based real-time object detector delivers the input for an algorithm, which generates trajectories reflecting the movement paths of the worker's hands. Those trajectories are automatically assigned to work steps using hierarchical clustering of similar behavior with dynamic time warping. The system has been evaluated in a task-based study with ten participants in a laboratory under realistic conditions. The generated logs have been loaded into the process mining toolkit ProM to discover the underlying process model and to measure the system's performance using conformance checking.