Hands emerge quite often in egocentric video, and their structure and disposition reveal a lot as to what individuals have been engaged and showing interest to. Earlier researches in hand detection, on the other hand, have depended on significant effect that only operate in limited situations, such as lab settings or with minimal social interaction. This work investigates the recently developed YOLO architectural for object recognition in egocentric video, which offers a good balance of precision and performance. We performed an experiment of assessments utilizing various different architecture configurations and training methods. Our experiment leads to all sorts of conclusions about the most cost - effective methodologies to our strategic objectives, and leads us to further research to enhance detection of smart glasses videos. We tested these approaches on an existing dataset that contains 48 first individual footage of people interacting in situations in real life as well as the bounding box on a pixel level for more than 15,000 hand gestures.
The rapid increase in Internet of Things (IoT) connected devices and generated data has greatly promoted the development of artificial intelligence. Federated Learning (FL) provides strong support for the data-driven society by training the model locally without leaking the privacy of the client. However, due to the unreliability of IoT devices and the centralized training model of FL, there are still security risks such as privacy leakage in federated learning. Therefore, this paper proposes a decentralized and privacy-preserving aggregation scheme, named DPPA, for FL. DPPA leverages the blockchain to structure a decentralized FL architecture, and leverages Paillier cryptosystem to realize the privacy protection of local data and the safe aggregation of model parameters. Through security analysis, we demonstrate that DPPA resists various security threats and preserve client privacy. Experiments show that in the large-scale IoT device environment, DPPA has great advantages over the existing competing approaches in terms of computational overhead.
Vision-based simultaneous positioning and mapping technology plays a crucial role in many fields including intelligent robots, autonomous driving, etc. However, the visual front-end tracking quality and the final positioning accuracy are still unsatisfactory due to the insufficient number of feature points in scenes with light changes, dynamic objects, or insufficient texture. To address this issue, this paper proposes a vision-based positioning and trajectory mapping method optimized for dynamic scenes. We first remove the characteristic points of the dynamic object, and then fuse motion sensor data and the vision sensor data. We also optimize the proposed system in terms of computation consumption. Extensive experiments have been conducted in three different settings to validate the accuracy and speed. It has been demonstrated that the proposed algorithm can improve the positioning accuracy by 90% and decrease the computation time by 94.1%, making it usable in edge devices with limited computation power in real world applications.
The indoor environment is much more complicated than the outdoor environment, which brings huge challenges to localization. In order to improve the accuracy and reliability of the indoor positioning system, we propose a single-site hybrid positioning (HP) system based on LOS recognition in this paper. The system adopts different positioning methods for different situations by distinguishing whether there is a line-of-sight (LOS) path in the environment. In this system, a LOS recognition algorithm is firstly designed to judge the LOS path depending on the received channel state information (CSI). According to the LOS recognition result, the HP system adaptively selects the optional positioning algorithm. When there is a LOS path, the positioning method based on parameter estimation is applied for localization. When there is no LOS path, a fingerprint-based positioning algorithm is proposed. Mean filtering and priciple component analysis (PCA) are used to do data dimensionality reduction and a combination of neural network (NN) and weighted k-nearest neighbour (WKNN) is used to complete position estimation. Finally the HP system is compared with other positioning systems, and the experimental results show that our proposed HP system can achieve higher positioning accuracy.
As the key to image processing, text localization has a wide range of applications in intelligent detection. Given the complexity of text recognition algorithms, a large amount of data processing will consume too much time and energy. Aiming at the high latency defect of the Histogram of oriented gradients (HOG) algorithm in text localization technology, this paper designs an efficient single-core heterogeneous acceleration scheme. This scheme overcomes the problem of global memory synchronization caused by multiple cores and avoids complicated protocol operations. Experiments show that the scheme saves more than 50% of resources than the multi-core acceleration scheme, and the calculation speed is increased by more than twice. Compared with the central processing unit (CPU), the speedup ratio and energy efficiency ratio of the hog algorithm on the graphics processing unit (GPU) platform are 28 and 22.8, and on the field programmable gate array (FPGA) platform it is 6.9 and 42.5. The acceleration scheme meets the real-time and low energy consumption requirements of text localization and will play an important role in the field of image processing.
