With the general trend of informatization, the importance of handwritten Chinese characters is gradually being ignored. Due to the differences in the education levels and the writing habits of the public, problems such as confusion of stroke order, distorted fonts, and wrong postures for holding the pen always exist. Whether from the perspective of cultural inheritance or practicality, the learning of handwritten Chinese characters is an indispensable and important link in contemporary education. In this project, we hope that by introducing artificial intelligence and computer vision technology, we can monitor and guide students’ writing ability from the two aspects of pen grip posture and stroke order detection. The VGG16 Model was utilized to implement the function of pen grip posture classification. For the recognition of stroke order, it is hoped that a single Chinese character can be recognized first through real-time pen writing trajectories, so as to further perform the stroke extraction. CRNN (Convolutional Recurrent Neural Network) + CTPN (Connectionist Text Proposal Network) were applied for Handwriting Chinese characters recognition. SiamRPN++ model was implemented to track the pen tip region in the writing video, and through pen ballpoint detection and pen up/down classification, to achieve stroke segmentation of a single character. While in the last algorithm, the strokes would be classified as horizontal, vertical, oblique upwards, and oblique downwards according to the absolute value of the slope, and a length threshold and finally achieve the correctness judgment.
Person re-identification is presented to retrieve specific pedestrian targets across monitoring devices. In this paper, we propose an attribute feature fusion network (AFFNet) for pedestrian detection and re-identification, which integrates pedestrian detection, person re-identification and attribute recognition into a deep network for end-to-end multi-task learning. In AFFNet, we design the identity and attribute branches to separately learn the identity and attribute fusion features of pedestrians. The identity features focus on the overall pedestrian appearance, while the attribute fusion features pay attention to local regions of the pedestrian. These two types of features are concatenated for the person re-identification task, which complement each other to provide a more comprehensive description of pedestrians. Experimental results on two commonly used person re-identification datasets including DukeMTMC-reID and Market-1501 demonstrate the effectiveness of proposed AFFNet.
6D pose estimation is the committed step of autonomous docking of underwater vehicles using executive guidance, and it is also an important research topic in recent years. This method provides RGB images and CAD models and calculates the pose of the object. The prediction model based on keypoints performs well in 6D object pose estimation. The method of this model is to select keypoints first, and then estimate 6D pose through keyoint estimation. Therefore, the position of selected keypoints profoundly affects the accuracy of pose estimation results. To reduce the positional errors of keypoints introduced in previous studies, we propose a new framework for using keypoints to determine pose. This framework includes a vector projection based estimation method (VPE) and a two-stage pose calculation based on reprojection to improve the accuracy of keypoint position estimation required for pose solution. The two-stage pose calculation based on reprojection first obtains the initial pose value, and then iteratively optimizes the final pose based on the reprojection distance. Compared with the original method based on RANSAC vector voting on the LINEMOD test set, the final ADD(-S) metric and 2Dprojection perform $4.2 \%$ and $3.3 \%$ better after adding our VPE.
In order to improve the classification accuracy of test samples from different datasets in multi-domain image classification, this paper proposes a multi-domain sample classification method based on Baidu AI open platform image recognition and multi-model fusion. The classification task is divided into two steps: first, the domain to which the sample belongs is determined based on the predicted label and its probability output by the Baidu API, and then the deep learning model trained on that domain is used to predict the label of the sample. The effectiveness of the proposed method is verified through extensive experiments on three public datasets: Cifar-10, Cifar100, and Mini-ImageNet. In addition, compared with existing algorithms, the proposed method shows significant improvements in classification performance.
This article proposed a fast and accurate tire model for vehicle stability control simulation under all working conditions compared to nowadays tire models. It consists of 2 parts, which are offline analysis and online use. Offline analysis includes tire finite element model and fitting algorithm, while online use includes tire prior knowledge, Treadsim, and tire force and limits. This model can reflect tire mechanics at different vertical loads, inflation pressures, cambers, longitudinal slips, sideslip angles and different vehicle speeds without massive experiments for parameter calibration, therefore it is suitable for tire companies and researchers to develop stability control algorithm under all tire possible working conditions such as handling limits, complex road environments, tire wear, tire flat or tire burst.
