With advances in consumer electronics, demands have increased for greater granularity in differentiating and analyzing human daily activities. Moreover, the potential of machine learning, and especially deep learning, has become apparent as research proceeds in applications such as monitoring the elderly, and surveillance for detection of suspicious people and objects left in public places. Although some techniques have been developed for Human Action Recognition (HAR) using wearable sensors, these devices can place unnecessary mental and physical discomfort on people, especially children and the elderly. Therefore, research has focused on image-based HAR, placing it on the front line of developments in consumer electronics. This paper proposes an intelligent human action recognition system which can automatically recognize the human daily activities from depth sensors using human skeleton information, combining the techniques of image processing and deep learning. Moreover, due to low computational cost and high accuracy outcomes, an approach using skeleton information has proven very promising, and can be utilized without any restrictions on environments or domain structures. Therefore, this paper discusses the development of an effective skeleton information based HAR which can be used as an embedded system. The experiments are performed using two famous public datasets of human daily activities. According to the experimental results, the proposed system outperforms other state-of-the-art methods on both datasets.
In this paper, we propose a hybrid method in which rolling skew histogram and neural network techniques are fused to recognize patterns and identify cows in the milking rotary parlor of dairy farms. Individual cow identification is very important for managing the welfare and health care of an individual cow and developing the body condition scoring system. Although there has some sensor-based cows' identification system, those systems require to attach the sensor devices on each cow which is costly and burden on the cows. Since the proposed method applies a single video camera which is a non-contact device for the identification of many different cow patterns, the proposed system is low cost and no burdens on the cow. In particular, the operation of the system takes place while the cows are in the milking process in rotary milking parlor where the monitoring of individual cow is more effective than some other time and places. The identification process is based on the black and white pattern on the cow's body while moving on the rotary milking parlor. For the detecting and cropping of cows' body region is carried out by using rolling the skew histogram and used for the training process of the deep convolutional neural network. The trained network is employed for the identification of individual cow in the testing process. The experiments are performed on the self-collected cow video dataset which includes around 60 different cow's body patterns that have been taken at the large-scale farm in Oita Prefecture, Japan. The experimental results show that the proposed system is promising with the overall accuracy of 96.3 % and it is very effective and practical for the real-time cow identification system needed for establishing a modern precision dairy farming.
Nowadays, with the emergence of sophisticated electronic devices, human daily activities are becoming more and more complex. On the other hand, research has begun on the use of reliable, cost-effective sensors, patient monitoring systems, and other systems that make daily life more comfortable for the elderly. Moreover, in the field of computer vision, human action recognition (HAR) has drawn much attention as a subject of research because of its potential for numerous cost-effective applications. Although much research has investigated the use of HAR, most has dealt with simple basic actions in a simplified environment; not much work has been done in more complex, real-world environments. Therefore, a need exists for a system that can recognize complex daily activities in a variety of realistic environments. In this paper, we propose a system for recognizing such activities, in which humans interact with various objects, taking into consideration object-oriented activity information, the use of deep convolutional neural networks, and a multi-class support vector machine (multi-class SVM). The experiments are performed on a publicly available cornell activity dataset: CAD-120 which is a dataset of human–object interactions featuring ten high-level daily activities. The outcome results show that the proposed system achieves an accuracy of 93.33%, which is higher than other state-of-the-art methods, and has great potential for applications recognizing complex daily activities.
Nowadays, deep learning is very popular in a variety of research field due to its outperformance over the existing machine learning methods and its high generality over raw inputs. According to recent surveys, deep learning can give high performance in visual object recognition system. Human Action Recognition (HAR) is a promising research area over the computer vision research field due to its enormous applicability. Most of the conventional HAR need to extract the handcrafted features in advance before classifying the actions and the environments are fixed. Those limitations make HAR to depend too much on the problem. In real-world, it is difficult to choose the suitable feature depending on the problem and difficult to fix the environment. In this paper, we applied the deep learning technology over the Skeleton Motion History Image (Skl MHI) of human actions to implement HAR that can work independently on the problem domain. According to the experimental results, the proposed system achieves the high recognition accuracy with low computational cost under the various environments.
31 March 2018 in Journal of Robotics, Networking and Artificial Life
In smart city environments, public safety increasingly depends on intelligent surveillance systems that can be capable of adapting to dynamic and context‐dependent access restrictions. Traditional systems often rely on static and predefined boundaries that fail to respond to rapidly changing environments such as construction sites, public gatherings or emergency situations. This paper introduces a novel deep learning‐driven framework using ground‐plane homography for real‐time proactive intrusion prediction within these dynamically restricted zones (DRZs). Our method first employs deep learning to accurately detect and localise physical restriction markers (e.g., traffic cones). We then utilise ground‐plane homography estimation to accurately map these markers into two‐dimensional ground‐plane perspective, precisely defining the spatial boundaries of the DRZ in real‐time. After the reactive detection of restriction markers region, intrusion prediction is achieved through sophisticated human trajectory analysis and future path extrapolation. By forecasting a person's path and identifying projected future presence within the dynamic ground‐plane zone, the system assists proactive alerts and adaptive security responses before an actual violation. To the best of our knowledge, this is the first system capable of predicting intrusions into areas dynamically demarcated by visual restriction markers. The experimental results on real‐world surveillance datasets demonstrate the system's effectiveness in identifying the presence of humans in DRZ, validating its potential for deployment in smart cities and critical infrastructure.