Human Pose Recognition for Real-Time Applications in Surveillance Systems

Human pose recognition is a basic issue in computer vision that is used in analyzing activities, human-computer interaction as well as surveillance systems. Nevertheless, both training pose models and testing small classes often have been assessed on the basis of large volumes of manually labelled records, which is exasperating and shared. To overcome this shortcoming, the present paper gives a weakly monitored human pose classification model, which makes use of automatic estimation of poses and manual labeling avoiding the use of manual annotation. Initially, lightweight pose estimation model is used to extract human body key points out of an image. The set of heuristic rules that will automatically form pseudo-labels of common human poses is based on geometrical relations between joints that are landmarks. These pose features are subsequently again taken to train a multi-layer perceptron (MLP) pose recognizer. The dataset on which the proposed approach is tested is MPII Human Pose which consists of 24,984 images split into train-validation-test subsets of 70%, 20% and 10% respectively. Because of the lack of reliable samples produced by the automatic labeling process, two categories of poses were not for quantitative assessment. The last assessment is based on four well-represented classes of poses, including Standing, Sitting, Walking, and Bending. The experimental outcomes indicate high-performance on all the classes that were tested with macro-averaged classification accuracy of 94.51%. Standing, Sitting, Walking, and Bending pose provide results of class-wise F1-scores 0.9450, 0.8950, 0.9298, and 0.8120, respectively. The efficacy of the proposed method is also confirmed by the high-confidence visual predictions with the help of qualitative results. The findings point to the fact that despite lacking manual labeling, the suggested weakly-supervised framework is capable of efficiently categorizing human poses and gives a viable option that can be used in practice as an alternative to pose-based activities recognition.