Action Recognition of the Hand Holding the Pot in Cooking

To determine the action types performed by the cooker's hand holding the pot and the starting and ending time of each action, the paper formed a 2D curve using two components of the 6D time sequences composed by the pot's accelerations and angular velocities along three mutually orthogonal axes of the body frame fixed on the pot, and then convert the above-mentioned problem to a process that searching a curve segment( called curve type) on the 2D curve, where the selection of the two components rests with the to-be-recognized action type. To realize the searching process, the paper proposes an algorithm named slipping time window searching algorithm to segment the 2D curve into series of curve segments, with the curve type included in them. The last work of the searching process is the match of the curve segments and the curve type with the aim to determine which curve segments are similar with the curve type. After the pattern of the curve segments and the curve type are selected, the paper designs and compares two pattern matching methods that based on different types of hidden Markov models, respectively, that is PHMM and EHMM. The experiment shows that the performance of the recognition method based on PHMM is equal to that of EHMM, and the former has lower recognition and position-fixing precision, however faster recognition time, while the EHMM has the opposite results.