MVL (Machine Vision Lab) HUMAN MOTION VIDEO DATABASE Scripts, Queries, Recognition Jezekiel Ben-Arie ECE Department University Of Illinois at Chicago UIC MVL (Machine Vision Lab) Composition of interactive motion queries. Analysis and Recognition of human activities. Human body parts labeling. Human detection. UIC MVL (Machine Vision Lab) UIC HUMAN ACTIVITY CAPTURE AND REGONITION UIC MVL (Machine Vision Lab) Visual Feedback User Motion Query Video Retrieval Videos Video Database Video Analysis and Recognition Retrieved videos MVL (Machine Vision Lab) UIC HUMAN BODY PART LABELING Objective: Identify the roles of parts that appear as bars. Labeling : Using the spatial locations and orientations. Method : Finding maximum conjunction of partial hypotheses. MVL (Machine Vision Lab) HUMAN BODY PART LABELING Theoretical Foundations UIC UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING Illustration of Theoretical Foundations (a) Overlap of Spatial distribution for (a) Correct Labeling (b) Incorrect Labeling (b) UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING (a) Mesh diagram of Overlap of Spatial distribution for (a) Correct Labeling (b) Incorrect Labeling (b) MVL (Machine Vision Lab) HUMAN BODY PART LABELING Experimental Results Silhouette Extraction Bar detection Using Gabor signatures. Parsing silhouettes 90 different human poses 98.7% correct labeling. UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING Experimental Results UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING Experimental Results UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING Silhouette Extraction UIC MVL (Machine Vision Lab) HUMAN BODY PART LABELING Silhouette Extraction Illustration of variation of chromaticity and brightness distortion UIC MVL (Machine Vision Lab) HUMAN ACTIVITY RECOGNITION Introduction Poses indicative of actions taking place Poses involved in walking Indexing based recognition using sparse frames Extends this technique with optimal constrained sequencing based voting UIC MVL (Machine Vision Lab) UIC HUMAN ACTIVITY RECOGNITION Introduction Temporal sequence of pose vectors Multidimensional hash tables for model activities Individual hash tables for each body part Identifying input pose vectors as samples of densely sampled model activity and create vote vectors Vote vectors are temporal depiction of the loglikelihood that indexed pose belongs to a model Dynamic programming based constrained sequencing to recognize activities MVL (Machine Vision Lab) HUMAN ACTIVITY RECOGNITION Creating Vote Vectors Illustration of the entire voting process UIC MVL (Machine Vision Lab) UIC HUMAN ACTIVITY RECOGNITION Experimental Results Videos of sitting action overlaid with skeleton superposed with the help of tracking information Sparse samples of jump activity adequate for robust recognition UIC MVL (Machine Vision Lab) HUMAN ACTIVITY RECOGNITION Experimental Results Average votes for 5 test videos of each activity along with the votes for other activities. Rows – Test Activity Columns – Model Activity Recognition rate under various conditions of occlusion MVL (Machine Vision Lab) UIC HUMAN ACTIVITY RECOGNITION Experimental Results Performance of the approach under conditions of view point variance UIC MVL (Machine Vision Lab) FACE DETECTION Original Image Skin detection Detection by the Gabors Regions passing the ellipse area criterion Detected Faces UIC MVL (Machine Vision Lab) FACE DETECTION Original Image Detected faces with medium threshold (0.7) Detected faces with maximum threshold (0.8) MVL (Machine Vision Lab) UIC GUI for Queries Composition Motion query is composed by using model motion data clips. An example of a model motion data clip is a walk cycle consisting of a sequence of poses in one basic cycle of left-right steps. Model motion data clip can also be non-cyclic such as sitting. Model motion data clip is obtained from a motion capture library or can be interactively composed by the user. UIC MVL (Machine Vision Lab) INTERACTIVE GUI Specify Trajectory Key-points Interpolate by Splines Specify Activities Calculate Segments Calculate Position and Orientations Generate Motion Sequences(Scripts) Display MVL (Machine Vision Lab) UIC Theoretical Foundations • • • • • Parameterization of 3-D rotations (Euler Quaternions) Splines (Catmull Rom) Interpolation (SLERP, Quaternions) Human body model Motion composition techniques (Inverse Kinematics, Mocap) MVL (Machine Vision Lab) Limb Pose Vocabulary UIC MVL (Machine Vision Lab) Example of complete body poses UIC MVL (Machine Vision Lab) Inverse kinematics based key framing tool UIC MVL (Machine Vision Lab) Implementation UIC