Viewpoint Independent Action Recognition Based On Projective Invariants

Human action recognition plays a critical role in various computer vision applications such as human computer interaction, video annotation, and content-based retrieval. Extensive efforts have been devoted, but this issue still remains challenging due to action alignment, key frames, and viewpoint distortions. This paper focuses on view-independent action recognition, using projective invariants. There are two approaches:key joints-based and silhouette-based action recognition.For key joints-based action recognition, this paper employs the characteristic number invariant, and proposes two action representations on temporal and spatial domains:temporal and spatial characteristic number. The new invariant named characteristic number, which can construct a loop in non-coplanar conditions, preserving the view invariance. Temporal characteristic number is fit for long action sequences with few tracking joints. But spatial characteristic number does well in short sequences with more joints positions. Therefor, they are complementary for the other.Silhouette-based action recognition extracts the contour of human body actions that casts the action recognition to shape recognition. This paper proposes a new framework named hierarchical projective invariant contexts. Each contour point are represented by a from-coarse-to-fine approach, computing its projective invariants, which preserves the view invariance. With different invariants, two shape representations are constructed, called hierarchical cross ratio contexts and hierarchical characteristic number contexts. In addition, this paper proposes a novel discriminative descriptor called contour segment contexts. This way is divided based on contour segments, and compares the similarity of pair-segment. Based on the advantages of the above two methods, the synergy can remedy defects of the other, not only the discriminative power, but also the view invariance.Finally, experimental results on public datasets validate and analyze the efficiency of the proposed algorithms. Temporal and spatial characteristic number are superior for other invariants, but they are not robust to noise effect, hierarchical projective invariant contexts stabilizes under varying viewpoints, and has good robustness to various noise. Moreover, combining with contour segment contexts, the performance can be improved further.