In order to train a classifier for self checkout system in convenience store at low cost, a few-shot domain adaptation problem has to be solved. Since the system treats a classifier for large number of classes, conventional method of few-shot domain adaptation should be extended for many classes. This paper propose to exploit meta-class information by executing the adaptation on the normal-class level and the meta-class level simultaneously. The proposed method are shown to be effective for improving adaptation accuracy of a classifier for many classes. The results of our ablation study implies that i) the meta-class should be decided by using k-means clustering method rather than clustering manually, and that ii) the ratio between the number of normal-class and the number of meta-class should be fixed.

水泳中継においてカメラ映像から選手をリアルタイムトラッキング（追跡）し，その位置や速度を計算・表示する“スイマートラッキングシステム”をパナソニックと共同開発，実現した．プールでのトラッキングは，水面や水しぶきとの切り分け，泳ぎ方の違いなどさまざまな課題があったが，本開発ではそれらを解決し安定したデータ検出を実現した．水泳競技は映っている選手が誰かを判別しづらいが，検出された選手位置データを用いて，泳いでいる選手へのマーカ表示であったり，コースごとの選手のスピード表示であったりなど新しい演出の実現に至った．例えば先頭の泳者と二番手の泳者のスピードを比較することで，「ゴールまでには追いつきそうだ」，「全然追いつかない」といった新たな目線で競技を楽しむことができ，次世代の水泳中継を実現した．

In recent years, there has been a demand in the sports industry to reduce the burden of data collection and video editing for tactical analysis. To achieve these, a system that can recognize the game context is needed. In this study, we proposed a method to identify the player's shot timing at the frame level during a ball-striking sport. In this study, players' shots were detected in video of a tennis match. It was shown that shots could be detected with an F-score value of 87% or more within an error range of 1 frame (0.033 sec) by considering time-series information using a recurrent neural network. This technology is expected to be applied not only to tennis, but also to other sports that involve ball shots, such as table tennis, baseball, and volleyball. At the same time, it can be used to detect moments of a specific action (for example, touching or hitting an object).

Many methods of sports video analysis have been proposed in the computer vision field. However, the analysis of swimming videos is a challenging task. This is because there is a lot of noise, such as water splashes, making it difficult to see the swimmer’s motion and detect body parts. Thus, it is difficult to automatically estimate a swimmer’s motion, especially the stroke. In this paper, we introduce a novel approach to automatically estimating the stroke in such situations. Firstly, we detect the swimmer from a swimming video using a projective transformation, background subtraction, and a Kalman filter. We next create a model that learns a mapping from a window of frames to a point on a onedimensioned(1D) target signal, which represents a swimmer’s stroke (we call a ’stroke signal’). We use a convolutional neural network (CNN) and multi long short-term memory (MultiLSTM) which is an expanded model of LSTM. Finally , we estimate swimmer’s stroke from the stroke signal. In a dataset including various environments, the outputs of our system showed higher accuracy than previous ones.

In this paper we propose a novel approach for video-based person re-identification that exploits convolutional neural networks to learn the similarity of persons observed from video camera. We take 3-dimensional convolutional neural networks (3D CNN) to extract fine-grained spatiotemporal features from the video sequence of a person. Unlike recurrent neural networks, 3D CNN preserves the spatial patterns of the input, which works well on re-identification problem. The network maps each video sequence of a person to a Euclidean space where distances between feature embeddings directly correspond to measures of person similarity. By our improved parameter learning method called entire triplet loss, all possible triplets in the mini-batch are taken into account to update network parameters. This parameter updating method significantly improves training, enabling the embeddings to be more discriminative. Experimental results show that our model achieves new state of the art identification rate on iLIDS-VID dataset and PRID-2011 dataset with 82.0%, 83.3% at rank 1, respectively.

This paper proposes a novel approach for video-based person re-identification that exploits deep convolutional neural networks to learn the similarity of persons observed from video camera. By Convolutional Neural Networks (CNN), each video sequence of a person is mapped to a Euclidean space where distances between feature embeddings directly correspond to measures of person similarity. By improved parameter learning method called Entire Triplet Loss, all possible triplets in the mini-batch are taken into account to update network parameters at once. This simple change of parameter updating method significantly improves network training, enabling the embeddings to be further discriminative. Experimental results show that proposed model achieves new state of the art identification rate on iLIDS-VID dataset and PRID-2011 dataset with 78.3%, 83.9% at rank 1, respectively.

When team sports coaches instruct their team members, a sample video is needed for an offensive instruction. However, it is very difficult to search a video for the parts of interest, and it takes a long time to compile a database, for example, ball trajectories and player locations, because sports videos have many periods with no play such as timeouts. In particular, information on the ball trajectory is very important for tactical analysis. However, it is currently inputted manually, which is time-consuming. Therefore, we focus on American football, where the ball can rarely be seen. The contribution of this paper is the submission of a method of automatic ball trajectory extraction. First, we remove the no-play periods in team sports videos to leave only play time and reduce the processing time. Second, we extract the ball trajectory to enable tactical analysis. We propose a new approach to ball tracking by focusing on the ball holder prediction. Our method can be applied to situations with heavy occlusion. Finally, we construct a play search and 3D virtual display system using information on the ball trajectory and the Unity, which is the development environment of the 3D virtual visualization. We used it to construct a Virtual 3D map for watching the play from an arbitrary viewpoint.

Human tracking in surveillance camera has been challenging task in the field of computer vision. Tracking objects have large variations such as pose, body shape, clothes and so on. Especially in parts-based methods, postural change is big problem since appearnce of human changes drastically. We deal with this problem to use statistical shape model for tracking and detection. It represents the variations of postural change and body shape with low dimensions. Our trakcing result includes more detailed the position and shape of body parts. So we recognize rough pose and body direction to analyze it. These data is useful for seculity system or marketing decision in surveillance.

We propose a per-frame upper body pose estimation method for sports players captured in low-resolution team sports videos. Using the head-center-aligned upper body region appearance in each frame from the head tracker, our framework estimates (1) 2D spine pose, composed of the head center and the pelvis center locations, and (2) the orientation of the upper body in each frame. Our framework is composed of three steps. In the first step, the head region of the subject player is tracked with a standard tracking-by-detection technique for upper body appearance alignment. In the second step, the relative pelvis center location from the head center is estimated by our newly proposed poselet-regressor in each frame to obtain spine angle priors. In the last step, the body orientation is estimated by the upper body orientation classifier selected by the spine angle range. Owing to the alignment of the body appearance and the usage of multiple body orientation classifiers conditioned by the spine angle prior, our method can robustly estimate the body orientation of a player with a large variation of visual appearances during a game, even during side-poses or self-occluded poses. We tested the performance of our method in both American football and soccer videos.

We propose a human lower body pose estimation method for team sport videos, which is integrated with tracking-by-detection technique. The proposed Label-Grid classifier uses the grid histogram feature of the tracked window from the tracker and estimates the lower body joint position of a specific joint as the class label of the multi-class classifiers, whose classes correspond to the candidate joint positions on the grid. By learning various types of player poses and scales of Histogram-of-Oriented Gradients features within one team sport, our method can estimate poses even if the players are motion-blurred and low-resolution images without requiring a motion-model regression or part-based model, which are popular vision-based human pose estimation techniques. Moreover, our method can estimate poses with part-occlusions and non-upright side poses, which part-detector-based methods find it difficult to estimate with only one model. Experimental results show the advantage of our method for side running poses and non-walking poses. The results also show the robustness of our method for a large variety of poses and scales in team sports videos.

We propose a head and upper body pose estimation method in low-resolution team sports videos such as for American Football or Hockey, where all players wear helmets and often lean forward. Compared to the pedestrian cases in surveillance videos, head pose estimation technique for team sports videos has to deal with various types of activities (poses) and image scales according to the position of the player in the field. Using both the pelvis aligned player tracker and the head tracker, our system tracks the player's pelvis and head positions, which results in estimation of player's 2D spine. Then, we estimate the head and upper body orientations independently with random decision forest classifiers learned from a dataset including multiple-scale images. Integrating upper body direction and 2D spine pose, we also estimate the 3D spine pose of the player. Experiments show our method can estimate head and upper body pose accurately for sports players with intensive movement even without any temporal filtering techniques by focusing on the upper body region.