I have used CNN - Alex net algorithm along with Media pipe. The first step is the data obtaining process, where the images are captured to define the region of interest (ROI) in the frame. Lots of variables are present in the image which can result in unwanted outcome. So the data that needs to be processed is reduced to a large extent using Media pipe. The Media pipe contains only the landmark of human. The CNN - Alex's net along with Media pipe packages used in the project to recognize the human action in real time scenario. Alex's net is part of CNN model. The model consists of five layers with a combination of max pooling followed by 3 fully connected layers and they use Re Lu activation in each of these layers except the output layers. The convolutional layer contains filters and parameters. The filter is used to resize the image and reduce the dimension of the image. Max Pooling layer is used for segmentation. The Dense layer is used to classify the image based on the output from the convolutional layer. The output layer is the soft max layer. The soft max layer, which take decisions.
Capturing and pre-processing
In this module, how the input is taken from the user. The input for this system will be taken from images, videos and web camera. The video input will be converted into frames and then preprocessed. During preprocessing, the input will be padded and smart cropped. The input images will be preprocessed during the training process.
Detection
In this module, the pre-processed image will be given as an input to a pre-trained model named “Alex net”. The preprocessed model takes the input and marks the landmarks of human. The detection of the position of a still image or sequence of images (moving images) is known as Action detection. Machines cannot detect objects, whereas human eyes can. The human pose in the image varies due to its changing position whether it be sitting, standing, clapping, drinking, running, yoga pose like plank, goddess, tree, warrior and down dog.
Classification
The detected co-ordinates along with the confidence score will be given as an input to the classifier. Our classification model is a multi-layered perceptron. It is an eight layer perceptron. This model gives how accurate a human is towards a pose after classification. The majority of the poses will be displayed out of the given poses.
I have planned, designed and implemented the system for Human activity recognition system for controlling UI which is a standalone application for controlling the various user interface controls and/or programs like Flak. In the analysis phase we gathered four major actions done by a human on his daily basis whether it be standing, sitting, sleeping and bending and the techniques and algorithms they employ and the success/failure rate of these systems. Accordingly, we made a detailed comparison of these systems and analyzed their efficiency. In the design phase we designed the system architecture diagrams and also the data flow diagram of the system. We studied and analyzed the different phases involved and accordingly designed and studied the algorithms to be used for the same. The image we used is jpg extension to keep things straight forward and code to be clean and easily usable because jpg/png images are easily available.The future scope of our project can be achieved through CCTV cameras fixed on the pillars of a bank which can be a human tracking device that tracks the movements of each person in the bank and captures their actions and if something goes way past the limits it can send data to the manager.
You must be logged in to post a comment.