Human Activity Recognition (HAR) based on Statistical Estimation.

<< Back to Recent Research Areas & Projects

<< Back to Past Research Areas & Projects

Vision based recognition of human activities is an important and challenging problem in computer vision, machine learning and pattern recognition societies. In the last decade, this human activity recognition (HAR) has gained significant attention due to its immense potential applications, such as video surveillance for security, monitoring of patients or old people, in hospitals or in their homes, human-computer interactions and entertainment systems. Since human body is not a rigid object and may present a multitude of shapes and postures along with self occlusions even for the same person, a robust modeling is difficult to obtain. The main challenges to such models are viewpoint dependence, appearance variability, presence of occlusion, statistical or deterministic representation of a sequence of motion segments and a parsing mechanism that can temporally align the input signal with known activity patterns, as shown in Figs. 1 and 3. In this project we are trying to develop a generic HAR system which can recognize daily human activities.

Fig. 1 Sample images from the Weizmann database for five daily-live activities performed by five persons with five

sample images per activity and their normalized versions.

Fig. 2 The recognition rates plotted against the number of features used in the matching of cosine (-Cos) and Euclidian (-Eucd) distance measures of various approaches on the Weizmann database. 3param-ERE is our proposed approach shown in blue circle.

We have proposed a 3-parameter based algorithm that decomposes the eigenspectrum into reliable, unreliable and null subspaces to alleviate the problems of instability, over-fitting or poor generalization. Eigenfeatures are then regularized differently in these subspaces using the 3-parameter based eigenspectrum model derived from the reliable portion of the real eigenspectrum. The three subspace based eigen-decomposition is not used to limit the discriminant evaluation in one subspace but to enable the evaluation in the whole space. Thus, the proposed method is based on the global optimization that extracts features by searching the most discriminative ones in the whole space. This approach reduces the sensitivity of the extracted features to the dimensionality of activity images, the number of training samples, spatiotemporal dependencies and noise disturbances. The proposed regularization scheme facilitates a discriminative and stable low-dimensional feature representation of the activity images. Experimental results on the Weizmann (shown in Fig. 2) and INRIA-IXMAS (shown in Table 1) databases demonstrate the superiority of our approach over other popular methods.

Fig. 3 Sample images from the INRIA-IXMAS database for four activities performed by four persons with five sample images per activity and their normalized versions.

Table 1 Confusion matrix of the recognition rates on the challenging INRIA-IXMAS database using our proposed 3Param-ERE with dynamic time wrapping (DTW) for various activities. Rows indicate the probe activities and the columns indicate the gallery activities.

Table 3 Results on popular KTH activity database using our method.

[C1] B. Mandal and H. L. Eng, “3-parameter Based Eigenfeature Regularization for Human Activity Recognition,” IEEE 35^th International Conference on Acoustics, Speech and Signal Processing (ICASSP 2010), Dallas, Texas, USA, pp. 954-957, 14-19 March 2010. (Poster) [PDF]

<< Back to Past Research Areas & Projects

<< Back to Recent Research Areas & Projects