Motion of Oriented Magnitudes Patterns for Human Action Recognition

Abstract

In this paper, we present a novel descriptor for human action recognition, called Motion of Oriented Magnitudes Patterns (MOMP), which considers the relationships between the local gradient distributions of neighboring patches coming from successive frames in video. The proposed descriptor also characterizes the information changing across different orientations, is therefore very discriminative and robust. The major advantages of MOMP are its very fast computation time and simple implementation. Subsequently, our features are combined with an effective coding scheme VLAD (Vector of locally aggregated descriptors) in the feature representation step, and a SVM (Support Vector Machine) classifier in order to better represent and classify the actions. By experimenting on several common benchmarks, we obtain the state-of-the-art results on the KTH dataset as well as the performance comparable to the literature on the UCF Sport dataset.

Action recognition based on motion of oriented magnitude patterns and feature selection

“…By several hundred test cases where the recognition rates are calculated on 300 KTH action videos with different parameters, we found that the optimal parameters, as follows: Feature extraction – MOMP descriptor: we follow the same experimental settings as in [23]: o Step 1: Number of orientations is used to compute gradient and quantise orientation. o Step 2: We accumulate the magnitude of the pixels from their neighbours by Gaussian filter with the kernel size and the standard deviation . o Step 3: We encode the features based on LTP‐based self‐similarity and the SSD of gradient magnitudes among successive frames. To calculate SSD, we choose a cell size , ; the number of neighbouring cells in each block and a threshold (where is cell size, the threshold is varying from 5 to 7).…”

“…In order to evaluate the contribution of the feature post‐processing – PCA and feature selection – some filter methods, we compare the results on different systems separately in ours 1, ours 2, and ours 3. Ours 1 [23], which integrates our descriptor MOMP to VLAD and rbf‐SVM, obtains 94.4%. In comparison to ours 1, ours 2 has a better result (95.4%).…”

“…The features encode the action across different orientations of consecutive frames by the local gradient distributions. MOMP features are calculated as follows (see the detail of MOMP in [23]): Step 1: We calculate and quantise the gradient of each frame based on Haar features. We use soft assignment technique to decrease the loss of quantisation (Fig.…”

“…98.4 [12] 91.8 Klaser et al [19] 91.4 Yeffet and Wolf (LTP) [44] 90.1 Kovashka and Grauman [45] 94.5 Taylor et al (CNN) [46] 90.0 Le et al (CNN) [2] 93.9 Wang et al [47] 93.8 Action Bank [48] 98.2 Jiang et al [49] 95.8 Kliper-Gross et al (MIP) [50] 93.0 Zhou and Wang [51] 93.3 Wang et al [13] 94.2 Shabani et al [52] 95.3 Ji et al (CNN) [1] 90.2 Iosifidis et al [53] 92.1 Liu et al [54] 93.5 ours 1 (MOMP + VLAD + rbf-SVM) [23] 94. 4 We use a leave-one-out scheme to test the efficiency of our approach in recognition, i.e.…”

Multi-strange hadrons in Pb-Pb collisions at the LHC with ALICE /

“…Based on the taxonomy proposed by Goodfellow et al [26], the approaches to this problem can be categorized into two groups: (i) traditional methods [6,56,77,92], where the action representation is explicitly chosen and the action recognition is defined under conventional machine learning algorithms, and (ii) representation-learning strategies that explore machine learning techniques for both tasks. The latter includes shallow approaches, such as dictionary-based methods [44,55,57,83], and deep learning strategies [33,34,35,50,59,67].…”