Hockey Dataset Research Articles

Player detection method based on scale attention and scale equalization algorithm.

Object detection methods for team ball games players often struggle due to their reliance on dataset scale statistics, resulting in missed detections for players with smaller bounding boxes and reduced accuracy for larger bounding boxes. This study introduces a two-fold approach to address these challenges. Firstly, a novel multi-scale attention mechanism is proposed, aiming to reduce reliance on scale statistics by utilizing a specially created SIoU (Similar to Intersection over Union) label that explicitly represents multi-scale features. This label guides the training of multi-scale attention network modules at two granularity levels. Secondly, an integrated scale equalization algorithm within SIoU labels enhances the detection ability of multi-scale targets in imbalanced samples. Comparative experiments conducted on basketball, volleyball, and ice hockey datasets validate the proposed method. The relative optimal approach demonstrated improvements in the detection accuracy of players with smaller and larger scale bounding boxes by 11%, 7%, 15%, 8%, 9%, and 4%, respectively.

Novel Deep Feature Fusion Framework for Multi-Scenario Violence Detection

Detecting violence in various scenarios is a difficult task that requires a high degree of generalisation. This includes fights in different environments such as schools, streets, and football stadiums. However, most current research on violence detection focuses on a single scenario, limiting its ability to generalise across multiple scenarios. To tackle this issue, this paper offers a new multi-scenario violence detection framework that operates in two environments: fighting in various locations and rugby stadiums. This framework has three main steps. Firstly, it uses transfer learning by employing three pre-trained models from the ImageNet dataset: Xception, Inception, and InceptionResNet. This approach enhances generalisation and prevents overfitting, as these models have already learned valuable features from a large and diverse dataset. Secondly, the framework combines features extracted from the three models through feature fusion, which improves feature representation and enhances performance. Lastly, the concatenation step combines the features of the first violence scenario with the second scenario to train a machine learning classifier, enabling the classifier to generalise across both scenarios. This concatenation framework is highly flexible, as it can incorporate multiple violence scenarios without requiring training from scratch with additional scenarios. The Fusion model, which incorporates feature fusion from multiple models, obtained an accuracy of 97.66% on the RLVS dataset and 92.89% on the Hockey dataset. The Concatenation model accomplished an accuracy of 97.64% on the RLVS and 92.41% on the Hockey datasets with just a single classifier. This is the first framework that allows for the classification of multiple violent scenarios within a single classifier. Furthermore, this framework is not limited to violence detection and can be adapted to different tasks.

OBJECT DETECTION IN HOCKEY SPORT VIDEO VIA PRETRAINED YOLOV3 BASED DEEP LEARNING MODEL

Object detection is the most common task in Sports Video Analysis. This task requires accurate object detection that can handle a variety of objects of different sizes that are partially occluded, have poor lighting, or are presented in complicated surroundings. Object in field sports includes player’s team and ball detection; this is a difficult task resulting from the rapid movement of the player and speed of the object of concern. This paper proposes a pre-trained YOLOv3, deep learning-based object detection model. We have prepared a hockey dataset consisting of four main entities: Team 1 (AUS), Team 2 (BEL), Hockey Ball, and Umpire. We constructed own dataset because there are no existing field hockey datasets available. Experimental results indicate that the pre-trained YOLOV3 deep learning model generates comparative results on this dataset by modifying the hyperparameters of this pre-trained model.

Video violence recognition and localization using a semi-supervised hard attention model

The significant growth of surveillance camera networks necessitates scalable AI solutions to efficiently analyze the large amount of video data produced by these networks. As a typical analysis performed on surveillance footage, video violence detection has recently received considerable attention. The majority of research has focused on improving existing methods using supervised methods, with little, if any, attention to the semi-supervised learning approaches. In this study, a reinforcement learning model is introduced that can outperform existing models through a semi-supervised approach. The main novelty of the proposed method lies in the introduction of a semi-supervised hard attention mechanism. Using hard attention, the essential regions of videos are identified and separated from the non-informative parts of the data. A model’s accuracy is improved by removing redundant data and focusing on useful visual information in a higher resolution. Implementing hard attention mechanisms using semi-supervised reinforcement learning algorithms eliminates the need for attention annotations in video violence datasets, thus making them readily applicable. The proposed model utilizes a pre-trained I3D backbone to accelerate and stabilize the training process. The proposed model achieved state-of-the-art accuracy of 90.4% and 98.7% on RWF and Hockey datasets, respectively.

Mobile Neural Architecture Search Network and Convolutional Long Short-Term Memory-Based Deep Features Toward Detecting Violence from Video

Recently, surveillance cameras are deployed in many public places to monitor human activities. Detecting violence in videos through automatic analysis means significant for law enforcement. But almost many monitoring systems require to manually identify violent scenes in the video which leads to slow response. However, violence detection is a challenging problem because of the broad definition of violence. In this work, we will concern with physical violence that involved two persons or more. This work proposed a novel method to detect violence using automated mobile neural architecture search network and convolution long short-term-memory to extract spatiotemporal features in the video, and then adding two types of pooling layers max and average pooling to capture richer features, standard scaling these features and reducing the dimension using linear discriminative analysis to remove redundant features, and making classifier algorithms working well in low dimension. For classification, we trained and tested various machine learning models which are random forest, support vector machine (SVM), and K-nearest neighbor classifiers. We develop a combined dataset that contains violence and non-violence scenes from public datasets: hockey, movie, and violent flow. The performance of the proposed method is evaluated on a combined dataset in addition to three benchmark datasets, hockey, movie, and violent flow datasets in terms of detection accuracy. The results of our model showed high performance in combined, movie, and violent flow datasets using SVM classifier with accuracies of 97.5%, 100%, and 96%, respectively, whereas in the hockey dataset, we achieve the best result of 99.3% using the random forest classifier.

Long-term recurrent convolutional network violent Behaviour recognition with attention mechanism

Violent behavior recognition is an important direction of behavior recognition research. For traditional violent behavior recognition algorithms, there is too much background information when processing video information, which will cause greater interference in feature extraction, so the recognition accuracy is not high. Improved on the basis of effective recurrent convolutional network, a long-term recurrent convolutional network with attention mechanism is proposed. In the video preprocessing stage, a variety of attention mechanisms are introduced. In the feature extraction stage, the lightweight end-to-side neural network architecture GhostNet and convLSTM are selected to build a long-term recurrent convolutional network. The global average pooling and fully connected layer are used in the classification process. The combined approach realizes the classification of behaviours. The final results show that in the Hockey dataset, the algorithm in this paper has increased by 0.4% compared to LRCN, in the RWF-2000 dataset with more samples, it has increased by 10.5% compared to LRCN, and has increased by 1.75% compared to I3D, indicating that the algorithm in this paper can effectively suppress the background information. Interference, improve the performance of the algorithm.

Violence detection and face recognition based on deep learning

Abstract With the emergence of the concept of “safe city”, security construction has gradually been valued by various cities, and video surveillance technology has also been continuously developed and applied. However, as the functional requirements of actual applications become more and more diverse, video surveillance systems also need to be more intelligent. The purpose of this article is to study methods of brute force detection and face recognition based on deep learning. Aiming at the problem of abnormal behavior detection, especially the low efficiency and low accuracy of brute force detection, a brute force detection method based on the combination of convolutional neural network and trajectory is proposed. This method uses artificial features and depth features to extract the spatiotemporal features of the video through a convolutional neural network and combines them with the trajectory features. In view of the problem that face images in surveillance video cannot be accurately recognized due to low resolution, two models are proposed: the multi-foot input CNN model and the SPP-based CNN model. By testing the performance of the brute force detection method proposed in this paper, the accuracy of the method on the Crow and Hockey datasets is as high as 92% and 97.6%, respectively. Experimental results show that the violence detection method proposed in this paper improves the accuracy of violence detection in video.

Hockey activity recognition using pre-trained deep learning model

Activity recognition in sports is often complex task resulting from the rapid dynamic interaction within players. In this paper, pre-trained VGG-16, deep learning based hockey activity recognition model has been proposed. Own hockey dataset consisting of four main activity includes free hit, goal, penalty corner and long corner was constructed as there are no existing field hockey datasets available. Experimental results indicate that the pre-trained deep learning model generates comparative results on this challenging dataset by tweaking the hyperparameters of this pre-trained model.

Violent scene detection algorithm based on kernel extreme learning machine and three-dimensional histograms of gradient orientation

Most existing feature descriptors for video have limited representation ability. In order to improve the recognition accuracy of method for detecting the videos that include violent scenes and take advantage of the logical structure of video sequences, a novel feature constructing approach based on three dimensional histograms of gradient orientation (HOG3D), the Bag of Visual Words (BoVW) model, and feature pooling technology is proposed. This approach, combined with kernel extreme learning machine (KELM), can be used to detect violent scene. First, the HOG3D feature is extracted on the block level for video, and then the K-Means clustering algorithm is implemented to generate visual words. Then, the bag of visual words framework is used for the quantization of feature. And the feature pooling technology is operated to generate a feature vector for an entire video segment, and feature vectors of training data and testing data were used separately to train the model and evaluate the performance of the proposed approach. The experimental results showed that the proposed feature descriptor had good representation and generalization abilities. The proposed approach is efficient for violent scene detection, and the accuracy matches the best result on Hockey dataset, and it outperforms state-of-the-art on Movies.

Temporal segmentation and recognition of team activities in sports

A method for temporal segmentation and recognition of team activities in sports, based on a new activity feature extraction, is presented. Given the positions of team players from a plan view of the playground at any given time, we generate a smooth distribution on the whole playground, termed the position distribution of the team. Computing the position distribution for each frame provides a sequence of distributions, which we process to extract motion features for activity recognition. We can classify six different team activities in European handball and eight different team activities in field hockey datasets. The field hockey dataset is a new, large and challenging dataset that is presented for the first time for continuous segmentation of team activities. Our approach is different from other trajectory-based methods. These methods extract activity features using the explicitly defined trajectories, where the players have specific positions. In our work, given the specific positions of the team players at a frame, we construct a position distribution for the team on the whole playground and process the sequence of position distribution images to extract activity features. Extensive evaluation and results show that our approach is effective.

Player detection in field sports

We describe a method for player detection in field sports with a fixed camera setup based on a new player feature extraction strategy. The proposed method detects players in static images with a sliding window technique. First, we compute a binary edge image and then the detector window is shifted over the edge regions. Given a set of binary edges in a sliding window, we introduce and solve a particular diffusion equation to generate a shape information image. The proposed diffusion to generate a shape information image is the key stage and the main theoretical contribution in our new algorithm. It removes the appearance variations of an object while preserving the shape information. It also enables the use of polar and Fourier transforms in the next stage to achieve scale- and rotation-invariant feature extraction. A support vector machine classifier is used to assign either player or non-player class inside a detector window. We evaluate our approach on three different field hockey datasets. In general, results show that the proposed feature extraction is effective and performs competitive results compared to the state-of-the-art methods.

BiomechZoo: An open-source toolbox for the processing, analysis, and visualization of biomechanical movement data

It is common for biomechanics data sets to contain numerous dependent variables recorded over time, for many subjects, groups, and/or conditions. These data often require standard sorting, processing, and analysis operations to be performed in order to answer research questions. Visualization of these data is also crucial. This manuscript presents biomechZoo, an open-source toolbox that provides tools and graphical user interfaces to help users achieve these goals. The aims of this manuscript are to (1) introduce the main features of the toolbox, including a virtual three-dimensional environment to animate motion data (Director), a data plotting suite (Ensembler), and functions for the computation of three-dimensional lower-limb joint angles, moments, and power and (2) compare these computations to those of an existing validated system. To these ends, the steps required to process and analyze a sample data set via the toolbox are outlined. The data set comprises three-dimensional marker, ground reaction force (GRF), joint kinematic, and joint kinetic data of subjects performing straight walking and 90° turning manoeuvres. Joint kinematics and kinetics processed within the toolbox were found to be similar to outputs from a commercial system. The biomechZoo toolbox represents the work of several years and multiple contributors to provide a flexible platform to examine time-series data sets typical in the movement sciences. The toolbox has previously been used to process and analyse walking, running, and ice hockey data sets, and can integrate existing routines, such as the KineMat toolbox, for additional analyses. The toolbox can help researchers and clinicians new to programming or biomechanics to process and analyze their data through a customizable workflow, while advanced users are encouraged to contribute additional functionality to the project. Students may benefit from using biomechZoo as a learning and research tool. It is hoped that the toolbox can play a role in advancing research in the movement sciences. The biomechZoo m-files, sample data, and help repositories are available online (http://www.biomechzoo.com) under the Apache 2.0 License. The toolbox is supported for Matlab (r2014b or newer, The Mathworks Inc., Natick, USA) for Windows (Microsoft Corp., Redmond, USA) and Mac OS (Apple Inc., Cupertino, USA).