Literature Survey

Last year has witnessed several studies related to cricket. This includes research related to batting bowling, fielding, winning analysis and performance analysis . Shotnet proposes a 13 layered Convolutional Neural Network based architecture to classify the cricket shot images into six categories. They achieve an accuracy of 80% in the dataset they developed with 3600 cricket shot images. Later, the problem was adopted to classifying cricket shot videos. Author proposes an algorithm to identify motion vectors from cricket shot video and classify them using an existing filter developed for action recognition. However, they could only achieve an accuracy level of nearly 62%. Following this approach developed CNN based models and show that CNN is very effective in classifying cricket shot videos. Similar research has been developed to identify eighteen types of bowlers from bowling action images using the transfer learning approach. They use pre-trained CNN models available to retrain them on the bowling images to build the bowling classifier. Existing Solution proposes a multi-valued automated decision to predict whether a ball is a no-ball or wide ball from both the top and side view of the pitch. Similar to the players, even analysing the umpire's poses could help automation tasks like generating highlights. Develops two classifiers, one to differentiate the image frames with an umpire from without umpire and the other one to classify the umpire's pose into five different events, Six, No ball, Wide, Out and No action. Support vector machines are used to develop the two classifiers. A research was conducted to detect the type of bowlers based on their bowling actions. This research was based in CNN to identify eighteen different types of bowlers based on their bowling actions. A fresh dataset of 8100 images bowlers has been introduced to do the research also they have used pre trained CNN models to train their model. They have managed to get an accuracy of 93% on their test dataset. In another research conducted in 2010 to detect a cricket ball they have used videos in their training process. They used region segmentation method to separate the ball and the other objects. Here they have extracted the ball-candidates to detect the ball and eliminate the false alarm detection. Another research conducted was conducted to identify the foot overstep no ball detection. In this research they have proposed a system with 4 cameras placed on either sides of the pitch as soon as the umpire signals for a no ball review, the third umpire will select the respective camera and the taken video clip will be compared with the data set. Many researchers have suggested that they cannot classify the shot only from one angle. Various techniques are used in cricket for visualization and coaching. Still, they could not get proper results . Earlier researchers have used 3600 images of data set. They took 80% images of the dataset that means 2880 images to train the model and 20% image of the dataset that means 720 images for testing By doing this they have got an accurate result of 80 %. We can use CNN and deep learning technologies to classify the shots In research titled "Event detection in cricket video using intensity projection profile of umpire gesture" they have proposed an approach to detect events happening in a cricket match using umpire hand gestures. Through the introduced method umpire signals such as Four, six, wide, No-Ball and out can be detected. They have initially performed scene segmentation of a cricket video. Further the umpire frames have been identified from each scene and umpire gesture frames have been used for feature training in Random Forest model to extract events.