Page - 202 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

manipulation planning framework presented in [3], [9], [16], [5], and [1] considered the trajectory planning problem from the normal operation of a manipulation task. A time-series classification algorithm was presented in [3] to perform the online prediction of human reaching motion by applying a motion capture camera system. Partial segments of actual motion variables are compared with the subset of motion variables which represent the optimally time aligned human motion demonstrations. In [9], the predicted motion trajectories are represented as 3D voxels which infers the workspace occupancy information. Similar approaches were adopted in [16], [5] for human motion prediction. In [6], human-object interactions in combination with human motion trajectories were used to build temporal conditional random fields for anticipating human activities. In [1], a human worker’s intent was estimated by computing the probabilistic representation of workspace segmented areas to which the human is heading. Task and motion planners were integrated in [13] and [4] to identify and handle low-level process deviations such as collision-prone trajectories with the neighbouring objects. Here, the process addressed is a pick and place operation performed by a robot on a cluttered table and a payload carried by two robots respectively. During the process execution, the interface layer in between the task and motion planners determines the presence/absence of obstructions by identifying the collision-prone trajectories from the trajectory planner as low-level process deviations. Based on these deviations, the task planner is updated with a new state and sends a variation of the initial task plan to the trajectory planner. An alternative way to handle these kinds of deviations is to replace object grasping with multiple push-grasps in a cluttered environment [10]. With our work we intend to enhance the state of the art by cascading activity recognition and task recognition to identify low- level process deviations and perform task level trajectory planning. In this work, we also intend to realize activity recognition by estimating the skeletal joint positions with a higher sampling rate. 1.2. Paper Organization Section 2 deals with the methodology proposed for trajectory planning based on activity recognition and identification of low-level process deviations. Section 3 will present the experimental setup including a static process plan where the human worker and robot performs process steps/tasks within their shared workspace. Section 4 will detail the expected contributions. 2. Methodology In this section, the methodology behind the trajectory planning based on activity recognition and identification of low-level process deviations will be described along with the system architecture. Figure 2 depicts the system architecture which consists of 7 major building blocks 1) Object tracking 2) Skeletal joints estimation 3) Action recognition 4) Activity recognition 5) Task recognition 6) Trajectory planner and 7) High-level planner. The algorithms applied for object tracking and action recognition components have already been realized and evaluated in [14] and [15] respectively and will not be mentioned in this research work. Therefore, the methods required for the remaining major blocks will be mentioned here. 202