A Survey on Human Activities Recognition Models (Part-1)

 A Survey on Human Activities Recognition Models (Part-1)

Introduction:

Due to the digitalization of technology, Human Activity Recognition (HAR) is in demand now-a-days. With the rapid advancement of electronic devices, human activity-based processing system requires to perform various activities efficiently with a better accuracy rate. HAR includes video frames or sequences of images of human activities automatically. Since decades HAR processing has become one of the prime areas of research. It has applications in the field of digital devices using pattern recognition and computer vision (such as illegal car parking, analysis of athletes’ performances, surveillance, security, diagnostics of orthopaedic patients etc). HAR systems can be of two types, those are:- sequence-based classification and frame-based classification. It identifies human activities performed by analyzing video frames or sequences of images. Huge numbers of extraction techniques are proposed since years for human activities recognition, such as: spatio-temporal interest points, characterize spatio-temporal volumes and feature-based silhouette histogram. The whole HAR system is divided into three sub-modules, such as:- pre-processing, feature extraction, and recognition. In this survey article, I have thoroughly studied, analyzed and compared various human activities recognition approaches. I have also identified the merits and demerits of each approach and their performances with respect to the other pre-existing human activities recognition systems.

Different Human Activities Recognition Models for Video Processing:

H. Su, J. Zou and W. Wang tried to analyze activities in outdoor and identified several issues such as poor resolution, shadows, distances and segmentation issues [1]. In order to overcome these issues, they presented a new algorithm by considering silhouette width and LBP. The authors transformed the activity vectors to gray values. Both static and dynamic features of silhouette width are presented by gray values. Activity width sequence image is represented as texture and merged with LBP (Local Binary Patterns). LBP can be used as a pattern extraction tool in a spatio-temporal environment. The researchers validated their theory by experiments on an outdoor database. Further research can be done on this methodology in order to extend this approach by implementing an envelope of activities on the said database for optimized results. 

K. G. Chathuramali and R. Rodrigo proposed a new SVM classification for human activity recognition to achieve faster activity recognition [2]. They used SVM over metric learning because of better performance and less cost. This approach performs best when implemented in high dimensional vector space. Here the computation time is directly proportional to the numbers of training data (Decrease in training data results reduced computation time). They simulated and identified some features in their presented approach, those are:- 1) Imbalanced training data results very poor performance rate. 2) Decrease in training data results reduced computation time. 3) On increasing training set, accuracy rate increases. They concluded that their method outperforms other existing approaches. 

W. Lin, M. Sun, R. Poovandran and Z. Zhang proposed a new technique for human activity recognition for video processing automatically [3]. They sub-categorized features having high correlations into Category Feature Vectors (CFVs). Numbers of GMMs (Gaussian Mixture Models) are merged together to form activity and GMMs are represented by CFVs. The authors demonstrated that their presented technique is more flexible than that of other pre-existing methods. To enhance the recognition accuracy rate, a new algorithm known as CFR is formed. Here in the process of human activity recognition, video frames with high confident frames are considered. The researchers validated their theory with experiments and showed that their approach is more effective than that of other approaches.

B. Chakraborty, A. D. Bagdanov, J Gonz`alez and X. Roca tried to merge probabilistic optimization with human activity recognition [4]. They used HMM (Hidden Markov Model) for their approach. The authors suggested that their technique is eligible to differentiate similar actions. The researchers used HMMs to monitor stochastic movements of body parts. The above-said algorithm used detectors to find the viewpoint changes and self-occlusions by using sub-classifiers. Every sub-classifier is represented by a SVM. The numbers of detections are merged by a simple geometric constraint model. The researchers experimented with their method on KTH, Weizmann and HumanEva datasets. They achieved robust human activity recognition through their work.

J. Hernández, R. Cabido, A. S. Montemayor and J.J. Pantrigo introduced a new method to identify human activities in terms of 2D sequences [5]. They integrated real-time visual tracking and feature extraction concepts to present their approach. The researchers divided their concept into three important modules, those are:- tracking, feature extraction and activity recognition. Tracking modules merges particle filter and local search procedure in order to fasten the computation. Feature extraction module split silhouette into smaller rectangular boxes. Statistics of these rectangular boxes are monitored in the process of evolution. In the last module, it transfers these statistics to SVM for classification. The researchers demonstrated that their models work in a real-time environment with better performance.