In recent years, the number of proposed fall-detection systems that have been developed has increased dramatically. A threshold-based algorithm utilizing an accelerometer has been used to detect low-complexity falling activities. In this study, we defined activities in which the body’s center of gravity quickly declines as falling activities of daily life (ADLs). In the non-falling ADLs, we also focused on the body’s center of gravity. A hyperplane of the support vector machine (SVM) was used as the separating plane to replace the traditional threshold method for the detection of falling ADLs. The scripted and continuous unscripted activities were performed by two groups of young volunteers (20 subjects) and one group of elderly volunteers (five subjects). The results showed that the four parameters of the input vector had the best accuracy with 99.1% and 98.4% in the training and testing, respectively. For the continuous unscripted test of one hour, there were two and one false positive events among young volunteers and elderly volunteers, respectively.

In this paper we propose a robust lane detection and tracking method by combining particle filters with the particle swarm optimization method. This method mainly uses the particle filters to detect and track the local optimum of the lane model in the input image and then seeks the global optimal solution of the lane model by a particle swarm optimization method. The particle filter can effectively complete lane detection and tracking in complicated or variable lane environments. However, the result obtained is usually a local optimal system status rather than the global optimal system status. Thus, the particle swarm optimization method is used to further refine the global optimal system status in all system statuses. Since the particle swarm optimization method is a global optimization algorithm based on iterative computing, it can find the global optimal lane model by simulating the food finding way of fish school or insects under the mutual cooperation of all particles. In verification testing, the test environments included highways and ordinary roads as well as straight and curved lanes, uphill and downhill lanes, lane changes, etc. Our proposed method can complete the lane detection and tracking more accurately and effectively then existing options.

Face recognition (FR) has matured with deep learning, but due to the COVID-19 epidemic, people need to wear masks outside to reduce the risk of infection, making FR a challenge. This study uses the FaceNet approach combined with transfer learning using three different sizes of validated CNN architectures: InceptionResNetV2, InceptionV3, and MobileNetV2. With the addition of the cosine annealing (CA) mechanism, the optimizer can automatically adjust the learning rate (LR) during the model training process to improve the efficiency of the model in finding the best solution in the global domain. The mask face recognition (MFR) method is accomplished without increasing the computational complexity using existing methods. Experimentally, the three models of different sizes using the CA mechanism have a better performance than the fixed LR, step and exponential methods. The accuracy of the three models of different sizes using the CA mechanism can reach a practical level at about 93%.

This research proposes a single network model architecture for mask face recognition using the FaceNet training method. Three pre-trained convolutional neural networks of different sizes are combined, namely InceptionResNetV2, InceptionV3, and MobileNetV2. The models are augmented by connecting an otherwise fully connected network with a SoftMax output layer. We combine triplet loss and categorical cross-entropy loss to optimize the training process. In addition, the learning rate of the optimizer is dynamically updated using the cosine annealing mechanism, which improves the convergence of the model during training. Mask face recognition (MFR) experimental results on a custom MASK600 dataset show that proposed InceptionResNetV2 and InceptionV3 use only 20 training epochs, and MobileNetV2 uses only 50 training epochs, but to achieve more than 93% accuracy than the previous works of MFR with annealing. In addition to reaching a practical level, it saves time for training models and effectively reduces energy costs.

We developed an autonomous navigation system for a solar-panel-cleaning vehicle. The system utilizes the YOLOv4-Tiny object detection model to detect white lines on the solar panels and combines the model with a proportional–integral–derivative (PID) controller to achieve autonomous navigation functionality. The main system platform was built on Raspberry Pi, and the Intel Neural Compute Stick 2 (NCS2) was used for hardware acceleration, which boosted the model’s inference speed from 2 to 8 frames per second (FPS), significantly enhancing the system’s real-time performance. By tuning the PID controller parameters, the system achieved an optimal performance, with $K_P$ = 11, $K_i$ = 0.01, and $K_d$ = 30, maintaining the average value of the error e(t) at −0.0412 and the standard deviation at 0.1826 and improving the inference speed. The system autonomously followed the white lines on the solar panels and automatically turned when reaching the boundaries. The system also autonomously cleaned itself. The developed autonomous navigation system effectively improved the efficiency and convenience of solar panel cleaning.