Shape-based multi-view human action recognition using distance-based-matrix-regardless-of-row-priority classifier

The recognition of human activity (or action) in videos has elicited significant attention in recent years given its potential use in many real-life applications. Human Action Recognition (HAR) is typically applied in fields such as human–computer interaction, surveillance, content-based video retrieval, and sports event analysis. HAR is a complex process because characteristics such as gender, height, body shape, and age considerably affect the visual reproduction/representation of captured actions. In practical applications, changes in viewpoint are common and fundamentally unavoidable given the inherent limitations of camera technology or the inevitable dynamism of human motion. When such changes are implemented, the recognition rate of current HAR approaches dramatically decreases. This problem is typically mitigated by the use of cameras equipped with multiple fields of views, which provide richer information than that derived from single-view cameras. Even with such innovations, nonetheless, ensuring accurate correlation and acquiring multi-view learning data remain complicated challenges. This work proposes four methods to advance the field of HAR. The Shape-based features are extracted from frames silhouette by using proposed Global Silhouette Shape Representation (GSSR) method. This GSSR is suitable given that silhouettes present spatial information on actions over time. Concatenation, as a data fusion technique, is also applied to create a multi-view feature vector from a combination of single-view feature vectors. In other words, a matrix of multi-view features is generated for each action. Maximum-Distance-among-Feature-Vectors (MDFV) technique, as a frame selection method, is employed to choose a subset of frames (or feature vectors) with the maximum difference among them. This strategy is based on the removal of frames with mostly similar features. Relevant and suitable features are selected using Binary Particle Swarm Optimization (BPSO) technique. This research likewise develops a Distance-based-Matrix-Regardless-of-Row-Priority (DMRRP) classifier, which is driven by the idea that if two action sequences depict motion performed by the same or different individuals, then the sum (or mean) of the minimum distances between each individual frame of sequence 1 and all the frames of sequence 2 reflects the similarity between the two actions. This classifier can recognize actions captured from different views. Finally, this study evaluates the performance of a proposed Multi-View Human Action Recognition Based On Shape-Based Feature Extraction and Distance-Based Classifier (MHARSD) in single- and multi-view HAR. To evaluate this approach, an experiment that involves two publicly available multi-view HAR datasets (i.e., MuHAVi and IXMAS) is conducted to determine the quality of recognition that the method produces for different actions. MHARSD supports the recognition of a wide range of human actions. In all evaluations, it exhibits a recognition accuracy higher than that achieved by 2D multi-view HAR state-of-the-art methods.

Preview Text FSKTM 2016 5 IR.pdf Download (873kB) | Preview

Actions (login required)

View Item