Embedded device for targeted video recording of sporting activities

Abstract

The invention relates to a device (110) for recording a video of a sporting activity on a sports field (200), the device comprising a camera provided with a lens having a short focal length (111) and an embedded computing unit (112), characterised in that the computing unit (112) is arranged to: for each processed image in a stream of images (100, F1), predict the location of the current action in the sporting activity on the field; determine a region of interest (103; 104) in this image, the position of which depends on the location predicted from the processed image; extract, from the image, a sub-image corresponding to the determined region of interest (103; 104); generate a second stream of images (105, 106) comprising the set of extracted sub-images.

Description

Description Title of the invention: On-board device for targeted video recording of sporting activities Technical field [1] The present invention relates to the technical field of video recording systems for sporting events and in particular to devices and methods for analyzing and processing images to monitor and highlight key actions and moments of sporting activities. [2] In the above field, it is known to use fixed cameras or cameras piloted by an operator to capture sporting events. However, these methods have several major drawbacks. [3] First, fixed cameras cannot effectively track moving actions on the field, which can lead to poor video recording quality and make it difficult to analyze sports performances. In addition, fixed cameras can miss key moments of the match due to their limited field of vision. [4] On the other hand, operator-driven cameras require constant human intervention to track on-field actions. This method can lead to tracking errors and be costly in terms of human resources. In addition, the operator may also miss important moments of the match due to the need to focus on specific areas of the field. [5] Furthermore, existing video recording systems are not always able to automatically and accurately determine regions of interest in captured images, which can hamper the effective analysis and broadcasting of key moments in sporting events. [6] Finally, current solutions may not be suitable to effectively deal with different lighting conditions, terrain specificities or the particularities of amateur sports, particularly team sports, which may affect the overall quality of video recording and analysis of sports performance. Furthermore, existing solutions often require multiple cameras, which significantly increases the final cost of the product and makes the device more complex to use. [7] In the field of video recording systems for sporting events, existing shortcomings in the state of the art pose significant problems. For example, currently established methods for tracking sporting actions do not always manage to provide complete and accurate coverage of events on the field. This can lead to gaps in performance analysis and difficulties for spectators who wish to focus on key moments or specific actions. [8] Furthermore, existing video recording systems are often not flexible enough to adapt to different types of sports, fields, and lighting conditions. This limitation can reduce the quality of captured images and the relevance of the analyses provided, making it difficult for coaches, analysts, and spectators to identify the most important elements. [9] Another issue concerns the automatic determination of regions of interest in captured images. Current systems may lack accuracy and efficiency in this task, which can impact the quality of performance analysis and the broadcasting of key moments in sporting events. In addition, this can lead to a waste of time and resources when reviewing video recordings.

[0010] Additionally, current video recording devices can be expensive and complex to set up and maintain, especially when multiple cameras or operators are required to cover the sporting event. These costs and complexities can be a barrier for sports organizations and teams looking to record and analyze their performances, especially for organizations with limited budgets.

[0011] Finally, existing video recording systems may not provide an optimal user experience for viewers who wish to follow and replay key moments of sporting events. The image streams generated by these systems can be difficult to navigate, analyze, and share, which can harm spectator engagement and the popularity of the sports concerned.

[0012] Currently, there is no affordable and accessible solution that can automatically and accurately capture sports actions and key moments of sporting events without constant human intervention, while adapting to different lighting conditions, specific pitch characteristics, and the peculiarities of amateur sport. Furthermore, there is also no affordable embedded system capable of extracting key moments from a sports match, including selectively recording video sequences from a past moment. Therefore, there is a need for an innovative solution that addresses these issues and fills the gaps in the state of the art.

[0013] Furthermore, major shortcomings of the state of the art include the need to use multiple cameras to capture the stream of an amateur sporting event to be broadcast in other images. These systems are often not well suited to extracting key moments from a match or competition. Furthermore, another significant drawback of these devices is their high cost, making their acquisition difficult for amateur sports clubs. Despite their technological potential, these solutions do not always manage to meet the specific requirements of amateur sport. State of the art

[0014] The invention therefore falls within this context and seeks to resolve all of the aforementioned drawbacks. Thus, the invention seeks to propose an on-board device and a method for video recording of sporting activities, making it possible to extract regions of interest from the continuous flow of images captured by a single camera and to generate an image stream, without aberration and optimized for detailed performance analysis. It also facilitates more efficient and targeted broadcasting of significant moments of sporting events. Presentation of the invention.

[0015] The subject of the invention is a device for video recording a sporting activity on a sports field, comprising a camera equipped with a short focal length lens and a computing unit embedded in the device and capable of receiving at least one stream of images recorded by said camera equipped with a short focal length lens; said computing unit is arranged to: a. for each image, called processed image, of at least part of the received stream of images, predict the location of the current action of the sporting activity on the field; b. for each image of each group of successive images of the received stream of images containing a processed image, determine a region of interest in this image whose position depends on the location predicted from said processed image; c. for each image of each group of images, extract a sub-image from said image corresponding to said determined region of interest; d.generate a second image stream containing the set of extracted sub-images.

[0016] Advantageously, the camera equipped with a short focal length lens makes it possible to capture a large part of the sports field, or even the entire sports field.

[0017] If desired, said camera equipped with a short focal length lens may be in particular a wide-angle camera, preferably a “fisheye” type camera. The short focal length lens may be an integral part of the camera, in particular in the case of an action camera, or alternatively be mounted on electronic equipment incorporating the camera and the computing unit, in particular in the case of a smartphone or a touchscreen tablet, the lens in this case being equipped with a means of mounting on said electronic equipment, such as a shell or a clamp, so as to come opposite the camera when it is mounted on the electronic equipment.

[0018] The device for video recording a sporting activity on a sports field according to the invention allows better analysis and precise monitoring of the current action on the field thanks to the prediction of the location of a region of interest. This allows for optimal coverage of the event by focusing only on relevant areas.

[0019] Advantageously, determining a region of interest in each image makes it possible to target key moments of the action, improving the quality of the recorded content and facilitating subsequent analysis.

[0020] Furthermore, extracting a sub-image corresponding to the region of interest allows reducing the weight of processed and stored data, making the device more efficient and economical in terms of resources and storage space.

[0021] Therefore, generating a second image stream comprising all the extracted sub-images provides a clearer and more concentrated visualization of the sporting action, improving the experience for users of the device.

[0022] By doing so, this invention allows the use of an inexpensive on-board device to record video of sports activities. This makes this technology particularly attractive to sports enthusiasts who have limited means but still want to capture and analyze their performances on the field. Therefore, this invention helps to democratize access to high-quality video analysis tools for the general public. In the context of the present invention, the term "video recording device" means any device capable of capturing and processing digital visual sequences. This may include, but is not limited to, smartphones and digital tablets equipped with at least one camera, possibly equipped with an optical lens arranged on a shell or on a clip attached to the phone or tablet so that the lens faces the camera, or even action cameras or drones equipped with at least one camera.

[0023] If desired, the prediction of the location of the current action on the field can be achieved using different image processing and machine learning algorithms, thus offering alternative solutions to adapt to various sports activities and lighting conditions. These algorithms will also be able to take into account the movements of players and objects, such as balls, to improve the accuracy of the prediction.

[0024] Advantageously, the determination of a region of interest in each image can be carried out by taking into account additional factors, such as the speed of movement of the players, areas of high player density or key moments of the game. Additionally, the size and shape of the region of interest can be dynamically adjusted based on these factors, providing a flexible solution tailored to each situation.

[0025] Advantageously, the extraction of a sub-image corresponding to the region of interest can be achieved using different compression and image processing techniques, thus making it possible to find an optimal balance between image quality and reduction of the amount of processed and stored data. For example, lossless compression algorithms can be used to maintain image quality while reducing data size.

[0026] Advantageously, the generation of a second image stream comprising the set of extracted sub-images can be achieved using different editing and presentation methods, thus offering a variety of solutions to improve the experience of spectators and sports analysts. For example, the sub-images can be assembled into a single panoramic view, or presented in mosaic form, so as to facilitate the comparison and analysis of actions on the field.

[0027] In one embodiment of the invention, the calculation unit is capable of receiving information on the type of sporting activity and the calculation unit is arranged to, upon receipt of said information on the type of sporting activity, implement one or more image processing algorithms on all or part of at least one initial image recorded by the camera equipped with a short focal length lens to detect an outline of a field, the calculation unit selecting said algorithm(s) from a set of given algorithms as a function of said information on the type of sport, and to generate a binary mask, called "of terrain”, the dimensions of which correspond to the dimensions of the images recorded by the camera equipped with a short focal length lens and defining a selection zone corresponding to the outline of the detected terrain.

[0028] According to these features, information about the type of sporting activity allows for automatic tracking of players, where the camera can detect the characteristic movements of each sport and adjust camera parameters, such as zoom level and / or color settings, accordingly. In addition, these features ensure optimal coverage of players, ensuring that key moments of the game are captured with clarity.

[0029] By doing this, the camera can automatically adjust its settings to provide shots more suited to the type of sport being played. For example, during a handball or basketball match, it might choose a wider, wide-angle portion of the image, thus capturing a larger portion of the field. For football or rugby, the camera might take a smaller, wide-angle portion, focusing on specific areas of the field. The information about the type of sport thus makes it possible to retrieve areas adapted to the specific needs of each type of sport, without changing the camera's viewing angle.

[0030] Advantageously, the device allows optimal adaptation to different types of sporting activities by receiving information on the type of sport and implementing specific image processing algorithms. This flexibility improves the detection of field contours, thus ensuring better coverage and analysis of the sporting event. In addition, the selection of algorithms according to the type of sport contributes to more accurate and reliable contour detection, which enhances the quality of the processed images. Finally, the generation of a binary terrain mask, the dimensions of which correspond to the dimensions of the images recorded by the camera equipped with a short focal length lens, makes it possible to define a selection area corresponding to the contour of the detected field, thus facilitating the visualization and analysis of actions on the field by focusing on the relevant areas.

[0031] According to an exemplary embodiment of the invention, the reception of information on the type of sporting activity may be obtained automatically. The computing unit may, for example, be arranged to analyze the images recorded by the camera equipped with a short focal length lens and / or to analyze information from additional sensors, in particular a geolocation device, to detect characteristics specific to the sport practiced. This will allow the device to adapt quickly and autonomously to different sporting activities, without requiring manual intervention.

[0032] According to another exemplary embodiment of the invention, the device is capable of receiving information on a geographical location of the device, in particular provided by the user, so that upon receipt of said geographical location the calculation unit executes an instruction to deploy a menu for selecting the type of sporting activity allowing the user to choose the type of sport recorded.

[0033] Advantageously, the computing unit can be arranged to infer information on the type of sporting activity from the geographical location. This functionality makes it possible to improve the accuracy and efficiency of the image processing algorithms and the entire device, optimizing the tracking and analysis of actions on the field for different sports.

[0034] Alternatively, identification of the geographic location of the device may be performed using other location technologies, such as wireless LAN (Wi-Fi) positioning, triangulation of mobile phone signals or low energy Bluetooth beacons.

[0035] Alternatively, the inference of information on the type of sporting activity from the geographical location could be improved by using enriched geographical databases containing specific information on sports grounds and facilities, enabling the device to better identify the type of sport practiced and to adapt more precisely and quickly to different situations.

[0036] Alternatively, the inference of information on the type of sporting activity from the geographical location could be combined with other sources of information, such as sensors integrated into the device or data from other connected devices.

[0037] Alternatively, the computing unit may be arranged to use algorithms based on machine learning techniques to infer information about the type of sporting activity from geographic location and other contextual data such as calendars.

[0038] In one embodiment of the invention, the calculation unit is arranged to implement at least one colorimetric segmentation algorithm of said initial image to determine a set of regions in said initial image, each region having substantially the same color, and to select from said set of regions a region corresponding to the terrain, said selection zone being defined from the selected region.

[0039] Advantageously, by implementing at least one colorimetric segmentation algorithm of the initial image, the computing unit makes it possible to determine a set of regions having substantially the same color, thus improving the accuracy of the detection of the terrain, particularly when the latter consists mainly of a single color, such as a rugby or football field. This approach facilitates the selection of a region corresponding to the terrain from among all the identified regions, which makes it possible to define the selection area more precisely and to improve the quality of the extracted images for a more detailed and relevant analysis.

[0040] Alternatively, the implementation of a colorimetric segmentation algorithm may be supplemented or replaced by the implementation, by the computing unit, of other image segmentation techniques, such as segmentation based on textures, contours or shapes. These alternatives may offer equivalent or superior performance for certain situations or types of terrain, by adapting better to variations in color, lighting or specific characteristics of the terrain.

[0041] Alternatively, the computing unit may implement multiple color segmentation algorithms in parallel or in series, combining or comparing their results to improve the detection and selection of the region corresponding to the terrain. This approach provides a more robust and reliable solution for changing lighting conditions or terrains with complex colors or patterns.

[0042] Alternatively, the computing unit may implement one or more algorithms based on deep learning techniques, such as convolutional neural networks, to perform color segmentation and selection of the region corresponding to the terrain. These techniques are particularly robust to variations in color, lighting or specific terrain characteristics.

[0043] If desired, the determination of the selection area from the selected region can be carried out by the computing unit taking into account other factors or contextual information, such as the position of the players, the direction of the game or key moments of the action. This approach allows the selection area to be dynamically adjusted and the relevance of the extracted images to be improved.

[0044] In a cumulative embodiment of the invention, the calculation unit is arranged to implement at least one panoptic segmentation algorithm of all or part of said initial image to determine a set of regions in said initial image, each region being associated with an object class, and to select from said set of regions a region whose class corresponds to a “terrain” class, said selection zone being defined from the selected region.

[0045] Advantageously, by implementing at least one panoptic segmentation algorithm of the initial image, the computing unit makes it possible to determine a set of regions associated with different classes of objects. This approach improves the accuracy of field detection by identifying objects relevant for sports analysis in the initial image. By selecting a region whose class corresponds to a "terrain" class among the set of regions, the calculation unit precisely defines the selection zone in the initial image, thus improving the quality and relevance of the extracted sub-images.

[0046] Advantageously, the terrain selection step uses panoptic segmentation to analyze and identify specific areas based on vertical position criteria. More precisely, this technical step involves the calculation of two relevant values, which correspond to specific percentiles of the vertical position in the area recovered by the panoptic segmentation. The first value, associated with a higher percentile, provides an indicator of the maximum height of the terrain, while the second value corresponds to a percentile located approximately at the vertical midpoint of the terrain. In practice, the area selected for subsequent operations is the one that happens to be closest, in terms of absolute value, to this second value.Thus, this method enables optimized terrain selection based on robust and accurate analysis of the vertical attributes of the image, providing a significant improvement in the accuracy and efficiency of the terrain selection process.

[0047] Alternatively, the computing unit may be arranged to implement multiple panoptic segmentation algorithms in parallel or in series, combining or comparing their results to improve the detection and selection of the region corresponding to the terrain. This approach provides a more robust and reliable solution for changing lighting conditions or terrains with complex objects or patterns.

[0048] Alternatively, the determination of the selection area, by the calculation unit, from the selected region may be carried out taking into account other factors or contextual information, thus making it possible to dynamically adjust the selection area.

[0049] Alternatively, the selection of the region corresponding to the terrain may be based on additional or alternative criteria, such as the size, shape or texture of the detected regions. These criteria make it possible to more precisely identify the region of the terrain among all the regions.

[0050] In a cumulative embodiment of the invention, the calculation unit is arranged to implement at least one semantic segmentation algorithm of all or part of said initial image to determine at least one region in said initial image associated with a “terrain” type object class, said selection zone being defined from the region associated with this “terrain” type object class.

[0051] Advantageously, the computing unit may implement one or more algorithms based on deep learning techniques, such as convolutional neural networks, to carry out the semantic segmentation of said region associated with this “terrain” type object class.

[0052] In particular, it will be possible to consider using an algorithm comprising, on the one hand, a basic convolutional neural network, called a "backbone network", such as ResNet50, trained to extract information from the basic image, for example the edges and shape of the field, the positions and contours of specific areas of the field such as goals, or even characteristics related to the colors and textures of the field. The algorithm also comprises a dilated spatial pyramid pooling module, or ASPP, trained to apply in parallel, to the feature map from the basic network, different point convolutions dilated according to different dilation rates, for example 1, 6 and 12, and to concatenate the results of these convolutions, possibly with the result of a global pooling layer applied to the feature map.The enriched feature map resulting from this concatenation can be subject to a new point convolution, then to upsampling, before being classified, for example via a softmax type layer, each pixel of the base image thus being classified as being part of the terrain or not.

[0053] This type of model, for example known as “DeepLabV3”, makes it possible to segment an image at different spatial scales, and in particular to capture local details, such as the contours of goal areas, information at a larger scale, such as the distribution of lines on the terrain and global features, such as terrain zones.

[0054] Alternatively, the computing unit may implement multiple semantic segmentation algorithms in parallel or in series, combining or comparing their results to improve the detection and selection of the region corresponding to the terrain. This approach provides a more robust and reliable solution for changing lighting conditions or terrains with complex colors or patterns.

[0055] It may be provided that the semantic segmentation algorithm is combined with an instance segmentation algorithm, in particular by using the image classified by the semantic segmentation algorithm as input to the instance segmentation algorithm, or by applying an instance segmentation algorithm in parallel and combining, for example by multiplication, the images classified by each of the semantic segmentation and instance segmentation algorithms. The combination of these algorithms thus forms a panoptic segmentation algorithm.

[0056] Advantageously, the computing unit can be arranged to select algorithms based on information on the type of sport by combining several algorithms or by using meta-algorithms to select and adjust the algorithms most suited to each situation. This will further improve contour detection and adapt to more complex or unexpected scenarios.

[0057] Advantageously, the generation of a binary field mask by the computing unit can be supplemented by the creation of additional masks, for example to distinguish different areas of the field, teams or key players. These masks could be used to facilitate the analysis of actions on the field and to customize the presentation of images according to the specific needs of spectators or sports analysts.

[0058] In an alternative or cumulative embodiment of the invention, the calculation unit is arranged to predict the location of the current action of the sporting activity, from each processed image and said terrain binary mask. The prediction may be made directly from the processed image, combined with the terrain binary mask or from a mask obtained from the terrain binary mask, or alternatively be made from data obtained from the processed image, this data being able to be combined with the terrain binary mask or from a mask obtained from the terrain binary mask.

[0059] Advantageously, by arranging the computing unit to detect and estimate the position of people in each processed image, the device improves the quality of sports analysis by taking into account the movements and position of the players. In addition, by predicting the location of the current action of the sporting activity from each processed image, the positions of the detected people and the binary terrain mask, the device offers a more accurate and dynamic representation of the sporting event. This approach makes it possible to improve the relevance of the extracted images and to optimize the analysis of sporting actions for a better understanding and a more accurate evaluation of the players' performances.

[0060] Alternatively, person detection and position estimation could be performed using different object tracking techniques, such as correlation tracking, appearance tracking, or particle filter tracking. These alternatives could provide equivalent or superior performance for certain situations or types of sports, by better adapting to the movements and interactions of players on the field.

[0061] Alternatively, the computing unit could detect and estimate not only the position of people, including players, but also other objects or elements relevant to sports analysis, such as the ball, goals or field lines.

[0062] In a cumulative embodiment of the invention, the calculation unit is arranged to detect foreground objects in each processed image and to generate a binary mask, called "foreground" mask, defining a selection zone for each detected foreground object; to generate a final binary mask by multiplying said terrain binary mask and said mask foreground binary; and to predict the location of the current action of the sporting activity from each processed image and said final binary mask.

[0063] Advantageously, by arranging the computing unit to detect foreground objects in each processed image and generate a foreground binary mask defining a selection area for each detected foreground object, the device enables more accurate identification of important elements of the sporting action. In addition, by generating a final binary mask by multiplying the field binary mask and the foreground binary mask, the device improves the accuracy of localizing sporting events by focusing on relevant areas of the image. Thus, by predicting the localization of the current action of the sporting activity from the final binary mask, the device provides a better representation of the key moments and important actions of the event, which improves the analysis and understanding of the performance of players and teams.

[0064] Advantageously, the computing unit may be arranged to implement a machine learning algorithm to detect foreground objects in each processed image and to generate said foreground binary mask, said machine learning algorithm being trained to detect and estimate the position of a foreground object in an image.

[0065] Alternatively, the computing unit may be arranged to implement a two-step approach to image processing. The first step may rely on the use of a first algorithm, which may be used to segment each processed image. This algorithm will be able to accurately distinguish the foreground from the background of the image, thus making it possible to delimit the areas of interest for subsequent operations.

[0066] After this initial segmentation, the computing unit can call upon a second algorithm. This latter can be used to detect specific objects in the foreground of each processed image. It should be noted that this foreground will have been previously segmented by the first algorithm. This combination of two algorithms can provide a fine-grained image.

[0067] The final phase of image processing may involve feature extraction. This extraction can be done by superimposing the foreground mask onto the original image. By superimposing these two elements, a richer set of information can be obtained from the image, allowing for a more detailed analysis and a better understanding of its content.

[0068] In one embodiment of the invention, said machine learning algorithm trained to detect and estimate the position of a foreground object in an image may employ a neural network, in particular a convolutional neural network.

[0069] Alternatively, foreground object detection may be performed by the computing unit using alternative image segmentation methods, such as texture-based segmentation, region-growing segmentation, or supervised classification segmentation.

[0070] As an alternative to the operation of multiplying the binary terrain and foreground masks to generate a final binary mask, other operations may be used by the calculation unit such as mask combination operations, in particular resulting from the composition of one or more operations such as intersection, union or difference of images.

[0071] In an exemplary embodiment of the invention, the computing unit may be arranged to implement at least one machine learning algorithm, from each processed image and said final binary mask, the machine learning algorithm being trained to predict the location of a current action of a sporting activity from an image. For example, each processed image may be superimposed with said final binary mask, the active pixels resulting from this superposition being highlighted relative to the inactive pixels, in particular by being colored red. The image resulting from this superposition may thus be provided as input to the machine learning algorithm trained to predict the location of the current action of the sporting activity.

[0072] For example, said machine learning algorithm may use a ResNet50 type neural network. In this case, the ResNet50 type neural network used by the algorithm implemented by the computing unit may, for example, include a convolutional input layer, provided in particular with filters of size 7x7 and a step size of 2. This first layer may be followed by a batch normalization function, a ReLU type activation function and finally a maximum pooling layer of size 3x3 and step size 2. This combination of operations makes it possible to extract low-level characteristics from the input image while reducing the spatial size for further processing.

[0073] Preferably, the ResNet50 type neural network employed by the algorithm implemented by the computing unit comprises, following the convolutional input layer, a plurality of main stages, such as four main stages, each composed of several residual blocks. The fundamental principle behind these residual blocks is the concept of residual connection or skip, which makes it possible to combat the problem of gradient disappearance in deep networks.

[0074] The first stage could, for example, include one convolutional block and two identity blocks. The second stage could, for example, include one convolutional block and three identity blocks. The third stage could, for example, include one convolutional block and five identity blocks. The fourth stage could include one convolutional block and two identity blocks. The dimensions of the filters used in these blocks could gradually increase, allowing for the extraction of increasingly complex features.

[0075] Preferably still, the ResNet50 type neural network used by the algorithm implemented by the computing unit comprises, after the last main stage, a global average pooling layer followed by a dense layer, for example having a softmax activation function. The last global pooling layer is used to reduce the spatial dimension to lxl while the output dense layer is used to produce the class probabilities output. This last step allows the final prediction to be made based on the features extracted by the network.

[0076] In an alternative or cumulative embodiment of the invention, the computing unit is arranged to detect and estimate the position of people in each processed image; to exclude the positions of the detected people which are located outside the selection area of ​​said binary terrain mask; and to predict the location of the current action of the sporting activity from the remaining positions of the detected people.

[0077] Advantageously, excluding the positions of detected people located outside the selection area of ​​the binary field mask allows to focus only on the people actually involved in the sporting activity on the field. Thus, this approach improves the accuracy and relevance of the prediction of the location of the current action of the sporting activity, avoiding potential interference or distractions caused by irrelevant people or objects located outside the field. In addition, it also allows to optimize computing resources and reduce processing time, since the computing unit does not have to process unnecessary or irrelevant information for the analysis of the current sporting activity.

[0078] Advantageously, the computing unit may be arranged to implement a machine learning algorithm to detect and estimate the position of people in each processed image, said machine learning algorithm being trained to detect and estimate the position of a person in an image.

[0079] In one embodiment of the invention, said machine learning algorithm trained to detect and estimate the position of a person in an image may use a “Faster R-CNN” type neural network.

[0080] In an exemplary embodiment of the invention, the Faster R-CNN type neural network used by the algorithm implemented by the computing unit comprises a first network of the Region Proposal Network type followed by a second detection network. It may be provided that the neural network comprises a Deep convolutional neural network provided upstream of the first network, allowing feature maps to be extracted from the input images. This deep convolutional neural network could be a pre-trained model, such as ResNet50, to produce a feature map from an image provided as input.

[0081] Preferably, the first region proposal network is arranged to receive the feature map of said deep convolutional neural network and to generate a series of proposals for regions of interest of said image, each likely to contain an object, each associated with an objectivity score or a probability that said region of interest actually contains an object.

[0082] This first network could, for example, move a grid of variable size over the input feature map and, for each position of the grid, generate, from the grid and the feature map, one or more bounding boxes likely to contain a person and, for each box, a probability that the box contains a person rather than a background element. Consideration could be given to using a set of anchor boxes, with various scales and aspect ratios, to accommodate people of various sizes and shapes.

[0083] Preferably, the first region proposal network may include a stage for grouping region proposals into regions of interest. This stage ensures that the proposed regions are of a fixed size before moving on to the subsequent layers. For example, the grouping of regions of interest may be achieved by max pooling on inputs of varying sizes to produce feature maps of fixed size, which are then fed to a dense neural network.

[0084] Advantageously, the second detection network receives as input the fixed-size feature maps of the clustering of regions of interest and comprises a series of fully connected layers, arranged to classify the objects and determine whether a region of interest proposed by the first network comprises a person. It may be provided that the classification layer applies a function softmax to calculate the probability distribution over different object classes, including the background. These layers can also be arranged to optimize the coordinates of the object's bounding box, in order to obtain more precise localization.

[0085] Thus, due to its architecture, the Faster R-CNN network provides a robust mechanism for object detection, in this case, for the detection of players on the field in the context of this invention.

[0086] Alternatively, instead of excluding the positions of detected people located outside the selection area of ​​the field binary mask, the computing unit may implement a person tracking algorithm that assigns a unique identifier to each detected person and tracks their position over time. This allows filtering out people not involved in the sporting activity on the field based on their trajectory and behavior, thus providing an alternative to predict the location of the current action of the sporting activity accurately.

[0087] Alternatively, instead of excluding the positions of detected persons located outside the selection area of ​​said binary terrain mask, the computing unit may implement a weighting algorithm that assigns weights to the positions of detected persons based on their proximity to the terrain and their apparent involvement in the sporting activity. This approach allows for greater importance to be given to the most relevant persons for the analysis of the sporting activity, while considering information from persons located outside the terrain, thus providing an alternative for predicting the location of the current action of the sporting activity.

[0088] Alternatively, instead of excluding the positions of detected persons located outside the selection area of ​​said binary terrain mask, the computing unit may implement a supervised learning algorithm using a trained neural network to predict the position of persons positioned only on the terrain.

[0089] In one embodiment of the invention, the computing unit is arranged to implement at least one machine learning algorithm for predicting the location of the current action of the sporting activity, from the remaining positions of the detected people, the machine learning algorithm being trained to predict the location of a current action of a sporting activity from a set of coordinates.

[0090] Advantageously, the use of at least one machine learning algorithm to predict the location of the current action of the sporting activity from the positions of the people detected and present within the binary field mask, makes it possible to improve the accuracy and efficiency of the prediction. Indeed, the machine learning algorithm can adapt to the different situations encountered in a sporting activity and learn to recognize the patterns and behaviors characteristic of each sport, which makes the device more versatile and efficient in tracking the action taking place on the sports field. In addition, this approach facilitates the consideration of various situations, including those that might be difficult to handle with traditional methods, such as overlapping positions of players or fast and complex movements of athletes.

[0091] In one embodiment of the invention, said machine learning algorithm trained to predict the location of the current action of the sporting activity may employ a neural network architecture specifically designed to predict the central position of the action in a sports game from the locations of the players. This architecture may be adapted to different sports, each having a specific number of input coordinates. For example, for handball, the architecture takes into account 192 input coordinates, while for basketball, this number is reduced to 144.

[0092] Preferably, the neural network may have a perceptron-type architecture comprising at least two dense layers. This type of architecture makes it possible to process the input coordinates to produce the position of the center of the action. Thus, the output of this network is a pair of coordinates indicating the center of the action. This prediction is notably carried out by evaluating the position of the players with different temporalities.

[0093] Alternatively, the computing unit could be configured to implement multiple machine learning algorithms, each specifically designed to handle particular aspects of the sporting activity or tailored to different types of sports. This approach allows for a more personalized solution tailored to the specific needs of each sport, leveraging the strengths of each algorithm for better prediction of the location of the current action.

[0094] Alternatively, the computing unit may implement one or more algorithms based on online learning techniques, allowing the machine learning algorithm to adapt and improve in real time, based on data received during a sporting event. This approach contributes to more accurate prediction and better adaptation to changing conditions during the event, such as changes in lighting, changes in weather or changes in team game strategies.

[0095] If desired, the computing unit could also be configured to implement a combination of machine learning algorithms and traditional image processing techniques to predict the location of the current action. This hybrid approach leverages the advantages of each method, for example, using image processing techniques to quickly detect significant changes in the scene and machine learning algorithms to refine the prediction based on contextual information and player behaviors.

[0096] In a cumulative embodiment of the invention, the calculation unit is arranged to predict the location of the current action of the sporting activity, from at least two successive processed images and said binary terrain mask. For example, the calculation unit may be arranged to implement at least one machine learning algorithm trained to predict the location of the current action of the sporting activity, from at least two successive processed images and said binary terrain mask.

[0097] Advantageously, the use of at least one machine learning algorithm to predict the location of the current action of the sporting activity from at least two successive processed images, and in particular from the positions of the people detected from these successive processed images, makes it possible to obtain a more precise and reliable prediction of the evolution of the action. Indeed, by analyzing the movements and interactions between the players from several successive images, the device can better understand the dynamics of the game and anticipate future actions. Consequently, this contributes to a smoother and more natural monitoring of sporting events, thus improving the viewing experience for spectators and facilitating the analysis of the athletes' performances for coaches and analysts.

[0098] Alternatively, the computing unit may be arranged to implement multiple machine learning algorithms in parallel, each specialized in predicting a specific aspect of the current action, such as player movements, ball trajectories, or player interactions. The predictions of these algorithms can then be combined to obtain a more robust overall prediction of the location of the current action.

[0099] Alternatively, the computing unit may be arranged to implement one or more algorithms based on unsupervised or semi-supervised learning techniques to predict the location of the current action, thus enabling the algorithms to be adapted to specific sports or game situations without requiring labeled training data. This approach could facilitate the implementation of the device in various sporting contexts, particularly for less common sports or amateur levels of competition.

[0100] Alternatively, the algorithms implemented by the computing unit can integrate temporal data, such as the speed and acceleration of players or moving objects, to predict the location of the current action. The use of this additional information makes it possible to refine predictions and improve the accuracy of tracking sports action, particularly in situations where movements are rapid and unpredictable.

[0101] Alternatively, the computing unit may be configured to implement deep neural network algorithms, such as convolutional neural networks (CNNs) or recurrent neural networks (RNNs), to predict the location of the current action. These types of algorithms are particularly suited to processing complex visual and temporal data and could provide improved prediction performance compared to traditional machine learning algorithms.

[0102] In a cumulative embodiment of the invention, the calculation unit is arranged to determine the location of the current sporting activity action on the field in the form of a pair of numbers each representing a horizontal and / or vertical coordinate in the selection zone of the binary field mask and each defining a proportion between a normalized minimum value and a normalized maximum value.

[0103] Advantageously, the use of a pair of numbers to represent the location of the current action of sporting activity on the field allows for a simplified and accurate representation of positions on the field, thus facilitating data processing and analysis. By expressing horizontal and vertical coordinates as proportions between a standardized minimum value and a standardized maximum value, the device is able to easily adapt to different dimensions of sports fields, thus providing greater flexibility and better compatibility with various sporting activities. In addition, this standardized approach facilitates the comparison of action locations between different instances of the device or between different sports, thus allowing for a more consistent and robust analysis of the collected data.

[0104] Alternatively, the computing unit may be arranged to determine the location of the current sporting activity action on the field using a polar rather than Cartesian coordinate system. In this case, the position of the action will be represented by an angle and a distance from a fixed reference point on the field, thus providing an alternative way of expressing the position of actions in a precise manner that is adaptable to different sports and fields.

[0105] Alternatively, the location of the current sporting activity action by the computing unit could be determined using a three-dimensional coordinate system, taking into account not only horizontal and vertical coordinates, but also height. This approach allows for better analysis of actions involving height movements, such as jumps or throws, thus providing a more complete representation of the location of the action for certain sports.

[0106] Alternatively, the computing unit can be configured to determine the location of the current action based on predefined areas of the field rather than continuous coordinates. In this case, the field is divided into sections or grids, and the position of the action would be represented by the index of the corresponding section. This method simplifies data processing and analysis, while allowing a flexible representation adaptable to different sports and fields.

[0107] In a cumulative embodiment of the invention, the calculation unit is arranged to segment so-called "boundary" zones in the initial image; to compare each predicted location with said boundary zones, and to update, as a function of said comparison, the prediction of said location of the action.

[0108] Advantageously, segmenting the so-called "boundary" areas in the initial image allows the device to identify the edges of the field or other areas important for tracking the action. By comparing each predicted location to the boundary areas, the device can adjust and refine the prediction of the action location based on this information. This leads to improved accuracy and reliability of the predictions, allowing for a better understanding of the dynamics of the game and the movements of the players on the field. Furthermore, updating the predictions based on the comparison with the boundary areas ensures that the data generated by the device is more relevant and reliable.

[0109] In a cumulative embodiment of the invention, for each image of each group of successive images of the received image stream containing a processed image, the calculation unit is arranged to determine a region of interest in this image whose position depends on the location predicted from said processed image and whose dimensions depend on the dimensions of the selection area of ​​the terrain binary mask.

[0110] Advantageously, using a region of interest in each image of each successive image group optimizes computational resources by focusing only on relevant areas. By determining the position of the region of interest based on the predicted location from the processed image, the device can track the action more accurately and efficiently. Furthermore, by adapting the dimensions of the region of interest based on the dimensions of the terrain binary mask selection area, the device ensures that the region of interest covers a large enough area to capture the relevant action while avoiding unnecessary data processing. This improves the overall performance of the device and provides more accurate and informative sports analytics.

[0111] Alternatively, the computational unit's determination of the region of interest may be based on other methods, such as motion detection or trajectory analysis, to more accurately track dynamic actions in the image stream. This allows for better targeting of relevant areas without negatively affecting the overall performance of the device.

[0112] Alternatively, rather than using only the dimensions of the bounding box of the terrain binary mask to determine the dimensions of the region of interest, the computation unit can be arranged to also take into account other factors, such as the density of players or objects in the area. This allows for the creation of a region of interest that more accurately adapts to the specific situation on the field and provides more relevant analysis results.

[0113] Alternatively, the computing unit can be arranged to use multiple regions of interest instead of just one to better cover the entire field and capture multiple simultaneous actions. The regions of interest can be adapted based on specific events on the field, such as player duels, passes, or shots. This allows for a more comprehensive and detailed analysis of the ongoing sporting activity.

[0114] In a cumulative embodiment of the invention, the stream of images recorded by said camera equipped with a short focal length lens comprises a succession of groups of successive images. Where appropriate, the calculation unit is arranged to predict the location of the current action of the sporting activity on the field from the first image only of each group of successive images, this image forming said processed image; and the calculation unit is arranged to determine said region of interest in each of the images of the same group from the location predicted from said first image and to extract a sub-image of said image corresponding to said determined region of interest.

[0115] Advantageously, by using only the first image of each group of successive images to predict the location of the current action, the device can reduce the processing load and the resources required while maintaining acceptable accuracy. In addition, by determining the region of interest in each of the images of the same group from the location predicted from the first image, the device can track the evolution of the action without having to process each image individually, thus improving efficiency and speed. Extracting a sub-image corresponding to the determined region of interest also allows the analysis to be focused on the relevant areas of the image, thus improving the quality of the information provided and reducing noise and interference.

[0116] Alternatively, instead of predicting the location of the current action only from the first image of each group of successive images, the computing unit may be arranged to predict the location using a representative subset of images from the group, thereby maintaining a reduction in processing load while potentially improving localization accuracy.

[0117] Alternatively, the region of interest may be determined by the computing unit using an object tracking approach in the group of successive images, thereby enabling dynamic adaptation of the region of interest based on movements of objects of interest, such as players or the ball, within the group of images.

[0118] Alternatively, sub-image extraction can be performed by the computing unit by applying a compression or resolution reduction algorithm to the initial images, thus reducing the amount of data processed without sacrificing the quality of the information extracted from the region of interest.

[0119] In a cumulative embodiment of the invention, the calculation unit is arranged to carry out image processing operations on each extracted sub-image so that the images of said second image stream are free from distortions generated by the camera equipped with a short focal length lens.

[0120] Advantageously, performing image processing operations on each extracted sub-image eliminates distortions caused by the camera equipped with a short focal length lens, thus improving the quality and accuracy of the images in the second stream. This makes it easier for spectators or coaches to analyze and understand ongoing sports actions by providing them with images that are free of distortions and more representative of the reality on the ground. In addition, correcting these distortions can also improve the performance of machine learning algorithms by providing them with more accurate data for analyzing and predicting sports actions.

[0121] Advantageously, the computing unit may comprise a memory in which a camera distortion model is stored. This model may be a non-linear model, such as a polynomial radial distortion model. This model uses a series of coefficients to describe how the radial distance of each point from the center of the image changes as a function of the angle of incidence of the light ray. The computing unit may be arranged to apply, to each pixel of each image of said second image stream, a geometric transformation depending on the coefficients of the distortion model and making it possible to remap the coordinates of each pixel in said image to obtain new coordinates in a rectified image. It may be provided that this geometric transformation is followed by an interpolation, in the case where this geometric transformation does not provide exact pixel coordinates in the rectified image.We can provide that the interpolation is bilinear or bicubic. to determine the pixel values. It may also be possible to follow this geometric transformation with a calibration to adjust the parameters of the distortion model and the geometric transformation according to the specificities of the camera and lens used. This calibration makes it possible to transform an image distorted by a camera equipped with a short focal length lens into a rectified image offering a more precise and realistic representation of the original scene.

[0122] In an exemplary embodiment of the invention, said model may be parameterized by the following parameters: the center of the lens, the radius, or even the vertical radius and the horizontal radius, of the lens, the tilt, roll and pan angles of the lens, indicating the rotations of the image around respectively a horizontal axis, a camera axis and a vertical axis. This model thus makes it possible, for each pixel of an image distorted according to this model, to define a transformation vector towards a pixel of a flattened image, to deform this vector by successive application of rotations according to the tilt, roll and pan angles and according to the position of the pixel with respect to the center and the radius of the lens. The coordinates of each pixel of the flattened image can thus be converted into pixel coordinates of the distorted image.The extracted sub-images can thus be converted to form images free from distortions caused by the camera.

[0123] It may be provided that the computing unit is arranged to apply an anti-aliasing filter during the conversion of each extracted sub-image. In particular, it may be provided that each pixel of the flattened image corresponds to an average of the values ​​of the pixels of a zone of the distorted image centered around the pixel of the distorted image whose coordinates correspond to the conversion, according to the model, of the coordinates of said pixel of the flattened image. Preferably, the conversion operations of each of the pixels may be carried out in parallel by the processor(s) of the computing unit.

[0124] Alternatively, the computing unit may be arranged to apply different types of image processing to correct distortions caused by the camera equipped with a short focal length lens, such as geometric transformations, radial distortion corrections or tangential distortion corrections. These different methods can be chosen depending on the specific characteristics of the camera used or the particular needs of the application.

[0125] Alternatively, the computing unit may be arranged to implement an adaptive approach to image processing, adjusting the correction parameters based on the intensity of distortions observed in the extracted sub-images. This approach allows for more accurate correction of distortions based on the specific conditions of each image.

[0126] Alternatively, the computing unit could be configured to implement deep learning algorithms to correct distortions caused by the camera equipped with a short focal length lens. Convolutional neural networks could be trained to learn how to correct these specific distortions, providing an alternative to conventional image processing methods.

[0127] In a cumulative embodiment of the invention, the calculation unit is arranged to apply a distortion operation to each extracted sub-image, said distortion operation being determined at least from the optical characteristics of the camera equipped with a short focal length lens.

[0128] Advantageously, by applying a distortion operation to each extracted sub-image based on the optical characteristics of the camera equipped with a short focal length lens, the device is able to accurately and appropriately correct the specific distortions caused by that camera. This customized approach allows obtaining images of the second image stream that are free of distortions, thus improving the quality and fidelity of the images for a better analysis and understanding of the ongoing sporting activity. Furthermore, by taking into account the optical characteristics of the camera, the device can be adapted to different types of wide-angle cameras, providing a flexible solution for various situations and needs.

[0129] Alternatively, the computing unit may be arranged to implement one or more machine learning-based distortion correction algorithms, where a pre-trained model is used to estimate and correct the camera-specific distortions with a short focal length lens. This approach allows dynamic adaptation to different wide-angle cameras, providing a more flexible and potentially more accurate solution for distortion correction.

[0130] Alternatively, the computing unit may be arranged to implement one or more distortion correction algorithms based on geometric methods, which take into account the geometry of the camera equipped with a short focal length lens and the placement of objects in the image to correct distortions. This approach provides a robust and efficient alternative for correcting distortions in extracted images.

[0131] Alternatively, the computing unit may be arranged to implement a combination of several distortion correction methods, thus making it possible to exploit the advantages of different approaches for optimal distortion correction. This solution provides better overall distortion correction performance for images extracted from the second image stream.

[0132] In one embodiment of the invention, the distortion operation is determined from one or more pieces of information among: distance from the camera to the terrain, tilt angle of the camera, brightness of the image, current zoom of the camera, the size of the terrain or even the position in the initial image of the determined region of interest.

[0133] Advantageously, by determining the distortion operation from one or more pieces of information such as the distance of the camera to the terrain, the tilt angle of the camera, the brightness of the image, the current zoom of the camera and the size of the terrain, the device can more precisely and adaptively adjust the distortion correction for each extracted sub-image. This makes it possible to obtain images from the second image stream with improved visual quality and a more faithful representation of the real scene. In addition, by taking into account this specific information, the device can dynamically adapt to the changes in recording conditions, thus ensuring optimal correction performance even when conditions vary.

[0134] Alternatively, the distortion operation could be determined from other information such as the camera's altitude relative to the terrain. This alternative approach would also allow distortions to be corrected adaptively based on the specific recording conditions.

[0135] Alternatively, the computing unit may be arranged to implement a machine learning-based distortion correction technique, which would be trained on a dataset representative of typical recording conditions. This approach makes it possible to predict and apply an optimal distortion operation for each sub-image, taking into account the optical characteristics of the camera with a short focal length lens and the specific recording conditions.

[0136] The invention also relates to a system for video recording a sporting activity on a sports field, the system comprising a video recording device, and at least one mobile terminal remote from the device and comprising a control unit, a storage memory and a user interface. Where appropriate, the device and the mobile terminal each comprise wireless communication means so that the device transmits the second video stream to the mobile terminal; and the control unit of the mobile terminal is arranged to, in response to a command entered via the user interface, record in the storage memory of the mobile terminal, at least a portion of the second video stream starting at a time prior to the time of entry of the command.

[0137] Advantageously, the integration of a video recording device into a system for video recording a sporting activity makes it possible to obtain a high-quality recording with adaptive correction of the distortions generated by the camera equipped with a short focal length lens. The use of a remote mobile terminal, equipped with a control unit, a storage memory and a user interface, facilitates management and interaction with the system by users. Thanks to the wireless communication means integrated into the device and the mobile terminal, the second video stream is transmitted in real time to the terminal, allowing users to view and analyze the images live. The control unit of the mobile terminal makes it possible to record in the storage memory of the terminal, at least a part of the second video stream in response to a command entered via the user interface, thus facilitating the saving of important images or sequences at a time prior to the time of entering the command. This configuration considerably improves the flexibility and user-friendliness of the video recording system for users, while ensuring high-quality recordings for subsequent analysis and dissemination.

[0138] Alternatively, instead of using a remote mobile terminal for management and interaction with the system, it may be possible to use a fixed terminal or workstation, providing an alternative solution for situations where mobility is not required or for use in a centralized control center.

[0139] Alternatively, the system may integrate multiple video recording devices, allowing larger areas of the sports field to be covered or different viewing angles to be recorded for more comprehensive monitoring of sporting activity.

[0140] Alternatively, the system may be equipped with wired communications means or a combination of wired and wireless communications to transmit the second video stream to the mobile terminal, thus providing an alternative depending on infrastructure needs and connectivity preferences.

[0141] Alternatively, the control unit of the mobile terminal may allow not only to record parts of the second video stream, but also to edit or annotate them, providing a more complete solution for the analysis and review of captured sequences.

[0142] Alternatively, the video recording system may include sports analysis software integrated into the mobile terminal or video recording device, enabling statistics, reports and visualizations based on recorded sequences, thus providing a more comprehensive solution for coaches and performance analysts.

[0143] In one embodiment of the invention, the mobile terminal may include a pre-recording trigger button forming said user interface. The button may be a mechanical button or a graphic element of an interface.

[0144] In one embodiment of the invention, the mobile terminal may integrate a pre-recording trigger button which will be part of the user interface. This button may be a mechanical button or a graphic element in a digital interface.

[0145] It may be provided that all or part of the second video stream received by the mobile terminal will be temporarily stored in a pre-recording buffer in the memory of the mobile terminal. This buffer may be designed to retain a limited quantity of content, for example the last 15 seconds of the video stream.

[0146] The second video stream will be recorded and played in its entirety, but in parallel, a 15-second circular buffer will also allow selective recording of highlights. When a user activates the shutter button on the mobile terminal, the frames of the second video stream stored in this pre-recording buffer will be retrieved by the control unit of the mobile terminal and added to the beginning of the current recording.

[0147] The stream will be continuously transmitted to the phone and will only be stored while the button is pressed, allowing for the 15 seconds prior to the button being pressed to be recorded specifically. This will allow the user to capture not only real-time moments, but also to revisit important moments that immediately preceded them.

[0148] Of course, the various features, variations and embodiments of the invention may be combined with each other in various combinations to the extent that they are not incompatible with, or mutually exclusive of, each other. Brief description of the figures

[0149] Other advantages and characteristics of the present invention are now described with the aid of examples which are purely illustrative and in no way limitative of the scope of the invention, and from the appended drawings, drawings in which the various figures represent:

[0150] [Fig. 1] represents, schematically and partially, a method for generating a video stream executed by a computing unit comprising regions of interest of a sporting activity according to a first embodiment of the invention.

[0151] [Fig. 2] represents, schematically and partially, a method for generating a video stream executed by a computing unit comprising regions of interest of a sporting activity according to a second embodiment of the invention.

[0152] [Fig. 3] represents, schematically and partially, a first image of a sports field, a second image of a colorimetric segmentation of the same sports field and a third image of a binary mask of the same sports field according to an embodiment of the invention.

[0153] [Fig. 4] represents, schematically and partially, the different stages of generation of a video stream from a succession of extracted sub-images corresponding to regions of interest according to an embodiment of the invention.

[0154] [Fig. 5] represents, schematically and partially, a region of interest of an image of a sporting activity determined from the detection of the players and the ball, according to one embodiment of the invention.

[0155] [Fig. 6] represents, schematically and partially, a device for video recording a sporting activity according to one embodiment of the invention.

[0156] [Fig. 7] represents, schematically and partially, a method of pre-recording and broadcasting a video stream of a region of interest of a sporting activity according to an embodiment of the invention.

[0157] [Fig. 8] schematically and partially represents the architecture of the convolutional neural network ResNet50.

[0158] [Fig. 9] schematically and partially represents the architecture of the Faster R-CNN neural network.

[0159] In the following description, elements which are identical, by structure or by function, appearing in different figures retain, unless otherwise specified, the same references. Description of the embodiments

[0160] In addition, various other characteristics of the invention emerge from the appended description given with reference to the drawings which illustrate non-limiting embodiments of the invention and where:

[0161] [Fig. 1] illustrates a method for generating a video stream 106, according to a first embodiment, executed by a computing unit 112 of a video recording device 110 of a sporting activity on a sports field 200, as represented in [Fig. 6],

[0162] The process begins with step 1 A selection of the type of sport to be recorded from a menu and geolocation of the device.

[0163] Step 1 Aincludes a sub-step 1.1 A of providing said geolocation data by a user of said video recording device 110 through a user interface (not shown).

[0164] Step 1 A includes a sub-step 1.2 A selecting the type of sport from a selection menu accessible from said user interface of the video recording device 110. The options presented by said selection menu being inferred from the geolocation data provided in step 1.1 A .

[0165] Step 1 A of the process is followed by a step 2 A capturing an initial video stream 100 by a camera equipped with a short focal length lens 111 of the video recording device 110 and sampling said initial video stream 100.

[0166] Step 2 A includes a sub-step 2.1 Acapturing the initial video stream by a camera equipped with a short focal length lens 111 of the “fish-eye” type and records 60 images per second. The initial video stream 100 recorded is transmitted to the calculation unit 112, the latter receives said initial video stream and records it in a buffer memory (not shown) of the video recording device 110; said received video stream is a received image stream 100, the images of said received image stream 100 are ordered chronologically.

[0167] Step 2 comprises a sub-step 2.2 of sampling by the calculation unit 112 of said initial video stream received 100 so as to produce two image sub-streams F1 and F2. The first image sub-stream F1 corresponds to the extraction of one image out of five from the received image stream 100 and the second image sub-stream F2 corresponds to the complementary set of the image sub-stream F1 with respect to the initial image stream received 100. The creation of the video streams F1 and F2 is carried out so that the images of each video stream are chronologically indexed as a whole, so that the joining of the two video streams F1 and F2 preserves the temporal order of the events.

[0168] Sub-step 2.2 A also comprises transmitting the second image sub-stream F2 to a processing stack within the calculation unit 112.

[0169] Step 2 A is followed by a step 3 Aof detecting the sports field 200 and generating a binary field mask 203 from the initial video stream 100 or from one of the image sub-streams F1 or F2. The algorithms which are implemented in this step 3A depend in particular on the information from step IA and are thus selected by the calculation unit 112, from a range of algorithms, from this information and in particular from the type of sport obtained at the end of step IA.

[0170] In the example described, for one or more initial images 201 of this image stream 100, F1 or F2, step 3A comprises a first sub-step 3.1A of determining a colorimetric segmentation 202 of this initial image 201 by the calculation unit 112. The colorimetric segmentation 202 produces a set of pixel zones, not necessarily connected, each of said zones being associated with the same color.

[0171] Step 3A comprises a second sub-step 3.2A of determining a panoptic segmentation (not shown) of the initial image 201 by the unit of calculation 112 allowing the identification of the field 200, the players and the spectators present in the initial image 201.

[0172] Step 3 includes a final sub-step 3.3 of generating a binary terrain mask 203 from the geolocation data provided in step 1A and the colorimetric 202 and panoptic segmentations obtained during steps 3.1A and 3.2. A . The binary terrain mask 203 is obtained by first determining the areas of the colorimetrically segmented image whose surface area is larger than a threshold of 25% of the total surface area of ​​the image.

[0173] In this particular case, two key values ​​are calculated from specific percentiles of the vertical position in the area targeted by the panoptic segmentation. The first value, corresponding to a high percentile, gives an indication of the maximum height of the area. The second value, in turn, corresponds to a percentile approximately at the vertical midpoint of the area, thus providing an index of the central position of the area. In this case, the area finally selected for the following operations is the one that is closest, in terms of absolute value, to this second value.

[0174] Finally, the terrain binary mask 203 is created from the selected area. The binary mask 203 represents the precise area of ​​the sports field 200 which will be retained for subsequent processing, in particular in order to exclude elements outside the sports field 200.

[0175] The method also comprises a step 4 A of segmenting the foreground and background of each image from the first image sub-stream F1 and generating a foreground binary mask (not shown). The steps which will thus be described are therefore repeated for each image of the first image sub-stream F1.

[0176] Step 4 A includes a first sub-step 4.1 A segmentation by the computing unit 112 of the foreground and background of each image from the first image stream F1. The computing unit 112 detects the objects in the foreground of each image by means of a machine learning algorithm in the form of a Convolutional neural network trained to detect and estimate the position of a foreground object in an image.

[0177] Step 4 includes a second sub-step 4.2 of generation by the calculation unit 112 of a foreground binary mask 204 (not shown) defining a foreground selection zone.

[0178] The foreground binary mask is created based on the foreground segmentation performed in sub-step 4.1 A. It represents the position area of ​​the main objects in the image, especially the players on the field.

[0179] The method comprises a step 5 A of generating a final binary mask and superimposing said final binary mask with the image sub-stream F1.

[0180] Step 5 A includes a first sub-step 5.1 A of generating the final binary mask 205 obtained by multiplying the terrain binary mask 203 and the foreground mask 204.

[0181] Step 5 A comprises a second sub-step 5.2 A of superposition by the calculation unit 112 of the final binary mask 205 with each image of the first sub-stream of images F1 to obtain, for each of the images, a set of active pixels resulting from this superposition and highlighted with respect to the inactive pixels by a coloring in red (not shown). The resulting image 206 of the superposition of an image of the sub-stream of images F1 and the final binary mask 205 is capable of being provided as input to a machine learning algorithm trained to predict the location of the current action of the sporting activity. All of the images 206 thus generated form a third stream of images F3.

[0182] Step 5 A is followed by a step 6 A of providing the third stream of images F3 as input to a ResNet50 neural network to locate the center of the sporting action. The ResNet50 neural network includes a convolutional input layer, provided in particular with filters of size 7x7 and a step size of 2. This first layer can be followed by a batch normalization function, a ReLU type activation function and finally a maximum pooling layer of size 3x3 and step size 2.

[0183] The ResNet50 neural network used by the algorithm implemented by the computing unit 112 comprises, following the convolutional input layer, four main stages, each composed of several residual blocks.

[0184] The first stage includes one convolutional block and two identity blocks. The second stage includes one convolutional block and three identity blocks. The third stage includes one convolutional block and five identity blocks. The fourth stage includes one convolutional block and two identity blocks.

[0185] Following the last main stage, the neural network has a global average pooling layer followed by a dense layer, featuring a softmax activation function.

[0186] This step also includes the determination and extraction, by the calculation unit 112, of a region of interest 103; 104 from the image stream F3. The determination of the region of interest is carried out by means of the ResNet50 neural network trained to return as output the location of the center of the action and to determine from said location of the center of the action a region of interest, in the form of a set of coordinates on the image defining a rectangle.

[0187] Step 6 is followed by a step 7 of generating an image stream 105 corresponding to the extracted regions of interest 103; 104. The extraction of the region of interest is carried out on each image of the image streams F1 and F2 to recompose a stream corresponding to the initial stream 100. All of the regions of interest thus determined form an image stream 105 composed solely of regions of interest.

[0188] Step 7 A is followed by step 8A of generating a stream of rectified images 106 from the stream 105. The rectification of the distortion generated by the camera equipped with a short focal length lens 111 of the video recording device 110 on all of the images forming the stream of images 105. The rectification of the image is carried out by applying a geometric transformation to each of the sub-images to correct the distortion introduced by the optics of the “fish-eye” camera 111.

[0189] First, a mathematical model of the distortion is determined from the specific characteristics of the camera optics equipped with a short focal length lens 111, such as the focal length, the position of the principal point and the radial and tangential distortion coefficients. This information is obtained by a camera calibration procedure by capturing images of a known standard pattern and estimating camera parameters that minimize the difference between the projected positions of the points of the pattern and their observed positions in the images.

[0190] Then, for each sub-image, a set of corresponding points is identified between the distorted image and an idealized version of the undistorted image. These points are used to calculate the geometric transformation that corrects the distortion. This transformation is usually a non-linear function that moves each pixel of the distorted image to its corrected position in the rectified image. Finally, the transformation is applied to the entire image to produce a rectified sub-image.

[0191] Finally, step 8 thus makes it possible to generate a stream of images 106 corresponding to the rectified sub-images of the regions of interest 103; 104.

[0192] [Fig. 2] illustrates a method for generating a video stream 106, according to a second embodiment, executed by a computing unit 112 of a video recording device 110 of a sporting activity on a sports field 200. The steps identical to the first embodiment retain the same references.

[0193] The method begins with a step 1 of selecting the type of sport to be recorded from a menu and the geolocation of the device 110.

[0194] Step 1A comprises a sub-step 1.1A of providing said geolocation data by a user of said video recording device 110 via a user interface (not shown).

[0195] Step 1A includes a sub-step 1.2A of selecting the type of sport from a selection menu accessible from said user interface of the device video recording 110. The options presented by said selection menu being inferred from the geolocation data provided in step 1.1.

[0196] Step 1 of the method is followed by a step 2 A of capturing an initial video stream 100 by a camera equipped with a short focal length lens 111 of the video recording device 110 and sampling said initial video stream 100.

[0197] Step 2 A includes a sub-step 2.1 Acapturing the initial video stream by a camera equipped with a short focal length lens 111 of the “fish-eye” type and records 60 images per second. The initial video stream 100 recorded is transmitted to the calculation unit 112, the latter receives said video stream and records it in a buffer memory (not shown) of the video recording device 110; said received video stream is a received image stream 100, the images of said received image stream 100 are ordered chronologically.

[0198] Step 2 A comprises a sub-step 2.2A of sampling by the calculation unit 112 of said initial video stream received 100 so as to produce two image sub-streams F1 and F2. The first image sub-stream F1 corresponds to the extraction of one image out of five from the received image stream 100 and the second image sub-stream F2 corresponds to the complementary set of the image sub-stream F1 with respect to the received image stream 100. The creation of the video streams F1 and F2 is carried out so that the images of each video stream are chronologically indexed as a whole, so that the joining of the two video streams F1 and F2 preserves the temporal order of the events.

[0199] Sub-step 2.2 A also comprises the transmission of the second image sub-stream F2 to a processing stack within the calculation unit 112.

[0200] Step 2 A is followed by a step 3 Aof detecting the sports field 200 and generating a binary field mask 203 from the initial video stream 100 or from one of the image sub-streams F1 or F2. The algorithms which are implemented in this step 3A depend in particular on the information from step IA and are thus selected by the calculation unit 112, from a range of algorithms, from this information and in particular from the type of sport obtained at the end of step IA.

[0201] In the example described, for one or more images 201 of this image stream 100, F1 or F2, step 3 comprises a first sub-step 3.1 of determining a colorimetric segmentation 202 of this initial image 201 by the calculation unit 112. The colorimetric segmentation 202 produces a set of pixel zones, not necessarily connected, each of said zones being associated with the same color.

[0202] Step 3A comprises a second sub-step 3.2A of determining a panoptic segmentation (not shown) of the initial image 201 by the calculation unit 112 making it possible to identify the sports field 200, the players and the spectators present in the initial image 201.

[0203] Step 3A comprises a final sub-step 3.3A of generating a binary terrain mask 203 from the geolocation data provided in step 1A and the colorimetric 202 and panoptic segmentations obtained during steps 3.1A and 3.2A. The binary terrain mask 203 is obtained by first determining the areas of the colorimetrically segmented image whose surface area is greater than a threshold of 25% of the total surface area of ​​the image.

[0204] In this particular case, two key values ​​are calculated from specific percentiles of the vertical position in the area targeted by the panoptic segmentation. The first value, corresponding to a high percentile, gives an indication of the maximum height of the area. The second value, in turn, corresponds to a percentile approximately at the vertical midpoint of the area, thus providing an index of the central position of the area. In this case, the area finally selected for the following operations is the one that is closest, in terms of absolute value, to this second value.

[0205] Finally, the terrain binary mask 203 is created from the selected area. The binary mask 203 represents the precise area of ​​the sports field 200 which will be retained for subsequent processing, in particular in order to exclude elements outside the sports field 200.

[0206] In this second embodiment of the invention, step 3 A is followed by a step 4 B of determining the positions of the players on the sports field 200 to from each image of the first image sub-stream Fl. The steps which will thus be described are therefore repeated for each image of the first image sub-stream Fl.

[0207] In a first sub-step 4.1B, the image is provided as input to a neural network whose architecture is based on that of the Faster R-CNN network to detect and determine the coordinates of the players on the sports field. The Faster R-CNN network was previously trained on the COCO dataset, which consists of more than 330,000 images divided into 80 object categories and 91 background element categories, with a total of 1.5 million object instances. Each image also has 5 captions, and there are keypoints for 250,000 people in the dataset.

[0208] Then, in a second sub-step 4.2B, the calculation unit 112 removes possible false positives of the players from the binary field mask 203, by excluding the positions located outside the boundaries of the field defined by this mask. The calculation unit can also remove other positions, depending on the information of the type of sport provided in step 1 by counting the number of players and referees expected in this type of sport. Thus, a set of positions of the players on the field is generated.

[0209] Step 4B is followed by a step 5B of determining a region of interest 103; 104 from the image stream F1, from all the positions of the players determined in step 4B as well as from the binary terrain mask 203. The determination of the region of interest is carried out by means of a perceptron-type neural network (not shown) trained to receive as input the coordinates of the positions of the players, determined in the previous step, and to return as output the center of the action, in the form of a pair of coordinates relative to the dimensions of the image and defining the center of the sporting action. From the determined center of the sporting action, the calculation unit 112 determines regions of interest 103; 104 on all the images of the stream F1.

[0210] Step 5B is followed by a step 7 of generating an image stream 105 corresponding to the regions of interest 103; 104 extracted in step 5B. The extraction of the region of interest is carried out on each image of the image streams F1 and F2 to recompose an image stream corresponding to the initial stream 100. All of the regions of interest thus determined form an image stream 105 formed solely of the regions of interest.

[0211] Step 7 is followed by a step 8 A of generating a stream of rectified images 106 from the stream 105. The rectification of the distortion generated by the camera equipped with a short focal length lens 111 of the video recording device 110 on all of the images forming the stream of images 105. The rectification of the image is carried out by applying a geometric transformation to each of the sub-images to correct the distortion introduced by the optics of the camera equipped with a short focal length lens 111 of the “fish-eye” type.

[0212] First, a mathematical model of the distortion is determined from the specific characteristics of the camera optics equipped with a short focal length lens 111, such as the focal length, the position of the principal point and the radial and tangential distortion coefficients. This information is obtained by a camera calibration procedure by capturing images of a known standard pattern and estimating camera parameters that minimize the difference between the projected positions of the points of the pattern and their observed positions in the images.

[0213] Then, for each sub-image, a set of corresponding points is identified between the distorted image and an idealized version of the undistorted image. These points are used to calculate the geometric transformation that corrects the distortion. This transformation is usually a non-linear function that moves each pixel of the distorted image to its corrected position in the rectified image. Finally, the transformation is applied to the entire image to produce a rectified sub-image.

[0214] Finally, step 8 A thus makes it possible to generate a second stream of images 106 corresponding to the rectified sub-images of the regions of interest 103; 104.

[0215] It should be noted that although the various steps of this method have been presented in a specific order, they may be performed in any order that remains consistent with the objectives of the method. In addition, certain steps may be omitted or combined with others, depending on the specific needs of the user and the capabilities of the video recording apparatus 110.

[0216] [Fig. 3] presents a series of three images illustrating different images used in the determination of the binary mask of a sports field. The first image 201 depicts an initial panorama of the sports field captured in its entirety and the elements around it such as the stadium, the stands and the spectators.

[0217] The second image 202 represents the same image after having been the subject of a colorimetric segmentation allowing the different zones to be displayed distinctly, in particular of the field, thanks to their specific color range. Finally, a third image 203 corresponding to a binary mask of a sports field obtained by joint application of the detection and localization algorithms allowing an image to be displayed by panoptic segmentation of the field on the one hand, and by the information obtained by the colorimetric segmentation on the other hand.

[0218] [Fig. 4], we observe a sequence of drawings illustrating the different sub-steps of a video processing process according to the embodiment of [Fig. 2].

[0219] The first part of the sequence shows the image stream 100 of a sports field obtained by the camera equipped with a short focal length lens 111.

[0220] The second part illustrates the process of sampling said first image stream 100 to form the sub-streams F1 and F2.

[0221] The third part illustrates the step where the terrain mask 203 is generated, by the calculation unit 112, from an initial image 201 of the first sub-stream F1, the step where the calculation unit 112 determines the positions of the people present on each image of the first sub-stream F1 and the step where certain positions are excluded, in particular taking into account mask 203 to arrive at a set of coordinates (xj, yj).

[0222] The fourth part illustrates the step where the calculation unit 112 predicts the location of the current action on the ground, from each image of the first sub-stream Fl and the set of coordinates (xj, yj), and thus determines regions of interest 103; 104.

[0223] The fifth part of the sequence illustrates the formation of an image stream 105 where each image is a sub-image extracted from an image of the sub-stream F1 or F2, corresponding to a region of interest 103; 104, the stream 105 thus corresponding to the initial stream 100.

[0224] Finally, the last part illustrates the step where the calculation unit corrects, in each image of the stream 105, the distortions introduced by the short focal length lens 111 of the “fish-eye” type in each image of the initial stream 100.

[0225] In [Fig. 5], we observe a specific graphical representation of a region of interest 103 of a sporting activity in progress on a sports field 200 from an image taken by a camera equipped with a short focal length lens.

[0226] This image highlights several players in the middle of the game, captured at the most crucial moment of the action. The human forms are distinctly highlighted, resulting from the identification of the players on the field.

[0227] Each player is framed by a rectangular box 300 highlighting their specific position on the field. In addition, each box is accompanied by a label, providing additional information about the type of object identified (not shown).

[0228] Also shown are the players' velocity vectors 302, indicating the direction and direction of movement, as well as the players' movement speed. This information is part of the input data for the machine learning algorithm responsible for predicting the region of interest in the frames of the sampled video stream.

[0229] Within said region of interest 103, the ball, the central object of the sporting activity, can also be located. The ball is also highlighted by a circular box 301 indicating its precise position on the field.

[0230] [Fig.6] is a schematic representation of a system for video recording a sporting activity on a sports field according to an exemplary embodiment of the invention. The system comprises a device 110 for video recording a sporting activity. The device 110 integrates a computing unit 112 and a camera comprising a short focal length lens 111, both embedded in the device.

[0231] In the example described, the device 110 is an action camera and the camera equipped with a short focal length lens 111 is a fish-eye camera and captures a stream of images 100 of the sports field 200 and making it possible to record all the sporting activity on the field 200. It may be provided, as a variant, that the device 110 is a smartphone carrying the computing unit 112 and a camera, a short focal length lens fixed on a shell or a clip being mounted, through the shell or the clip, on the smartphone, so as to come opposite the camera.

[0232] The computing unit 112, in close interaction with the camera 111, processes the captured image stream 100. The computing unit 112 implements the method of [Fig. 1] and segments the images, identifies the regions of interest 103; 104, predicts the actions and locations, and performs all the operations necessary to create a second image stream 106 free of distortions and focused on the regions of interest 103; 104.

[0233] Furthermore, the device integrates a wireless communication element 113, of the Wi-Fi type. This element allows the device 110 to transmit the second image stream 106, generated by the calculation unit 112 at the end of step 8 A of the method of [Fig. 1] or [Fig. 2], to a mobile terminal 116.

[0234] The mobile terminal 116 is provided with a control interface comprising a trigger button 115 for recording in memory the second image stream 105 transmitted by the device 110 via the wireless communication element. thread and this from an instant preceding the instant of pressing the trigger button 115, in this case the 15 seconds preceding said instant of pressing the trigger button 115.

[0235] In [Fig. 7], we have illustrated a procedure that continuously records a sporting activity from the second video stream 105 on the mobile terminal 116, highlighting a specific region of interest. Step 0 refers to the actuation of the trigger button 115. This does not initiate the recording, but rather marks a point of interest in the continuous video stream. More specifically, pressing the trigger button 115 signals an instruction to retain the last fifteen seconds of the second video stream 105 preceding that moment, until the next actuation of the same button.

[0236] After step 0, the process continues with a series of steps, indicated by a block, corresponding to steps 1 to 8 described either in the first embodiment of the invention described in [Fig. 1], or in the second embodiment of the invention described in [Fig. 2]. It should be noted that these steps are constantly being executed, capturing the entire sports match and transmitting it to the mobile terminal 116, independently of the actuation of the trigger button 115. The trigger button 115 is mainly used to mark moments of interest to be reviewed later, although the entire sequence of the match can be retrieved thanks to a specific functionality.

[0237] The process is completed by a step 9 of post-processing the second video stream 105, which makes it possible to add graphic elements, such as for example score tracking and player identification, to the images of the video stream.

[0238] In [Fig. 8] we have represented the architecture of the ResNet50 neural network.

[0239] In [Fig. 9], we have represented the architecture of the Faster R-CNN neural network, used to determine the region of interest.

[0240] The network is composed of a feature extraction layer 400, performed by a pre-trained ResNet50 network, arranged to extract a feature map 401 from an image 101 provided as input to the network.

[0241] Next, a region proposal network 402 is applied to the feature map 401. The region proposal network, or RPN, 402 uses a convolutional network with a filter of size 3x3 to produce two outputs: the objectness scores (not shown) and the regression bounding boxes 403.

[0242] The RPN object proposals then go through a 404 region of interest grouping mechanism.

[0243] Finally, a 405 detection network is applied, which typically comprises two dense layers of size 4096. Two outputs are produced by this 405 detection network: an object classification layer and a bounding box regression layer.

[0244] Dimensions may vary depending on the size of the input image 101, the network used for feature extraction, and the number of object classes.

[0245] The training of the Faster R-CNN follows a process comprising a four-step alternative training regime. The training begins by training the region proposal network as an independent network for generating object proposals. Subsequently, the proposals generated by the region proposal network are used to train the detection network. Next, the region proposal network is trained again, this time initialized with the parameters from the detection network. Finally, the detection network is retrained once more using the common convolutional layers with the region proposal network.

[0246] In terms of training data, for handball, a database comprising 279,656 images was used. For basketball, the training database has 121,000 images. These databases are then augmented to strengthen the robustness of the network training. For validation, 25% of the training database is reserved. These images are annotated with the position of the center of the action.

[0247] Regarding optimization, the ADAM optimizer is used for all models. The cost function used to evaluate the performance of the models is the mean square cost which measures the average of the squares of the errors or deviations between predictions and actual values.

[0248] The foregoing description clearly explains how the invention achieves its stated objectives, namely to improve the quality and efficiency of recording and monitoring sports activities, by proposing an intelligent system which, through the use of a camera equipped with a short focal length lens and a sophisticated computing unit, is capable of predicting and tracking the location of the current action of the sports activity. Furthermore, the invention provides for the use of a mobile terminal allowing convenient user interaction and management of video recordings.

[0249] In any event, the invention cannot be limited to the embodiments specifically described in this document, and extends in particular to any equivalent means and to any technically operative combination of these means. In particular, it is possible to envisage the use of different types of sensors or cameras, the application of other machine learning algorithms, the prediction of location based on additional data such as historical activity data, or even the extension of the use of the system to applications other than sport, such as security monitoring or movement analysis in the context of artistic performances. In addition, it is possible to envisage the integration of additional functionalities in the mobile terminal, such as the possibility of annotating videos in real time, or of instantly sharing video extracts on social networks.

Claims

Claims

1. Device (110) for video recording a sporting activity on a sports field (200) comprising a camera provided with a short focal length lens (111) and a calculation unit (112) embedded in the device (110) and capable of receiving at least one image stream (100) recorded by said camera provided with a short focal length lens (111), characterized in that the calculation unit (112) is arranged to: a. for each image, called processed image, of at least a part of the received image stream (100, F1), predict the location of the current action of the sporting activity on the field; b. for each image of each group of successive images of the received image stream (100) containing a processed image, determine a region of interest (103; 104) in this image whose position depends on the location predicted from said processed image; c.for each image of each group of images, extracting from said image a sub-image corresponding to said determined region of interest (103; 104); d. generating a second image stream (105, 106) comprising the set of extracted sub-images.

2. Device (110) for video recording a sporting activity on a sports field (200) according to claim 1, characterized in that the calculation unit (112) is capable of receiving information on the type of sporting activity and in that the calculation unit (112) is arranged to, upon receipt of said information on the type of sport, implement one or more image processing algorithms on all or part of at least one initial image (201) recorded by the camera equipped with a short focal length lens (111) to detect an outline of a field (200), the calculation unit (112) selecting said algorithm(s) from a set of given algorithms as a function of said information on the type of sport, and to generate a binary mask (203), called a “field mask”, the dimensions of which correspond to the dimensions of the images recorded by the camera equipped with a short focal length lens (111) and defining a selection zone corresponding to the contour of the detected terrain (200).

3. Device (110) for video recording a sporting activity on a sports field (200) according to claim 2, characterized in that the calculation unit (112) is arranged to implement at least one colorimetric segmentation algorithm of said initial image to determine a set of regions in said initial image, each region having substantially the same color, and to select from said set of regions a region corresponding to the field (200), said selection zone being defined from the selected region.

4. Device (110) for video recording a sporting activity on a sports field (200) according to one of claims 2 to 3, characterized in that the calculation unit (112) is arranged to implement at least one panoptic segmentation algorithm of all or part of said initial image to determine a set of regions in said initial image, each region being associated with an object class, and to select from said set of regions a region whose class corresponds to a “ground” class, said selection zone being defined from the selected region.

5. Device (110) for video recording a sporting activity on a sports field (200) according to one of claims 2 to 4, characterized in that the calculation unit (112) is arranged to predict the location of the current action of the sporting activity, from each processed image and said binary field mask (203).

6. Device (110) for video recording a sporting activity on a sports field (200) according to the preceding claim, characterized in that the calculation unit (112) is arranged to detect foreground objects in each processed image and to generate a binary mask (204), called "foreground" mask, defining a selection zone for each detected foreground object; to generate a final binary mask (205) by multiplying said field binary mask (203) and said binary mask foreground (204); and to predict the location of the current action of the sporting activity from each processed image and said final binary mask.

7. Device (110) for video recording a sporting activity on a sports field (200) according to one of claims 5 to 6, characterized in that the calculation unit (112) is arranged to detect and estimate the position of people in each processed image; to exclude the positions of the detected people which are located outside the selection zone of said binary field mask (203); and to predict the location of the current action of the sporting activity from the remaining positions of the detected people.

8. Device (110) for video recording a sporting activity on a sports field (200) according to one of claims 5 to 7, characterized in that the calculation unit (112) is arranged to determine the location of the current action of sporting activity on the field in the form of a pair of numbers each representing a horizontal and / or vertical coordinate in the selection zone of the binary field mask (203) and each defining a proportion between a standardized minimum value and a standardized maximum value.

9. Device (110) for video recording a sporting activity on a sports field (200) according to one of the preceding claims, characterized in that the calculation unit (112) is arranged to carry out image processing operations on each extracted sub-image so that the images of said second image stream (106) are free from distortions generated by the camera equipped with a short focal length lens (111).

10. System for video recording a sporting activity on a sports field, the system being characterized in that it comprises a device (110) for video recording a sporting activity on a sports field (200) according to one of the preceding claims, and at least one mobile terminal (116) remote from the device and comprising a control unit, a storage memory and a user interface, in that the device and the mobile terminal each comprise wireless communication means so that the device transmits the second video stream (106) to the remote mobile terminal (116), and in that the control unit of the mobile terminal is arranged to, in response to a command entered via the user interface, record in the storage memory of the mobile terminal, at least a part of the second video stream (106) starting at a time prior to the time of entry of the command.

Images