Design, Implementation of CNN for Human Activity Recognition using CHaiDNN and SDx Tool on Zynq Ultrascale+ MPSoC

Lalitha, Ramaraju and Acharyya, Amit (2019) Design, Implementation of CNN for Human Activity Recognition using CHaiDNN and SDx Tool on Zynq Ultrascale+ MPSoC. Masters thesis, Indian institute of technology Hyderabad.

Text
Mtech_Thesis_TD1483_2019.pdf
Restricted to Repository staff only until 9 July 2022.
Download (6MB) | Request a copy

Abstract

Human activity recognition has a variety of applications . It can be used in Medical domain for recognizing activities such as cycling, running, walking which can be used as a feedback for the treatment of the people with diabetes, heart stroke, obesity. It can also be used in Smart hospitals and to detect abnormal activities of patients with mental pathologies and in security it can be used for capturing suspicious behavior from videos (vision based AR), also for intrusion detection using cameras as sensors and in intelligent homes to recognize complex activities such as eating, washing dishes etc. It finds its application for Ambient assisted living(AAL) where it Exploits activity monitoring, recognition and assistance to support independent living. HAR can also be used in Military to monitor soldier’s activities along with the location, health condition. Challenging task of HAR is to use the sensor data to classify the people’s activity. This can be done by a good classifier. Different classifiers such as support Vector Machine (SVM), K-Nearest Neighbor (KNN), Decision Trees (DT), Multi-Layer Perceptron (MLP), Naive Bayes classifier, were applied for Human Activity Recognition using wearable sensors. Most of them were used to classsify activities such as walking, running, gestures etc. Most of them focus on extracting the informative features, since feeding raw data to a classifer doesn’t give the appropriate results without the feature extraction step. Convolutional Neural Networks has reduced the effort of feature selection and also increased the accuracy levels when compared to other classifiers.Data driven approach used by CNN enables the user to feed the raw data and get the final scores of the predicted classes.Convnets are used for classification, similarity matching, segmentation, Face recognition, video classification, Interpretation and diagnosis of medical images. As a proof of concept Healthy subjects were explored to recognize three arm movements which are involved with the daily living activities with the help of CNN([1]). This will help in monitoring the arm rehabilitation in pathologies that are asssociated with neurodegenerative diseases such as stroke or cerebral palsy. This study was later extended to stroke data and a personalized, low complex CNN model named Rehab-Net was formulated, for which the hardware implementation should be done([2]). Main Challenge in Inference is Rate of AI innovation, new models are getting created which makes Fixed silicon architectures (ASIC) no more a solution because even after creating a state of art model by the time we get the silicon back it is outdated. Hence we need to move towards Adaptable Hardware (FPGA) because of its major advantage of Time to market. But ASIC has advantages over FPGA’s in terms of speed , Area and Power which should be taken care in order to use FPGA. To get the advantages of speed, size, and power, DSA (Domain Specific Architectures) is a solution. DSA’s Created and tuned for a class of networks and has the features of Custom dataflow and Custom memory hierarchy. Low latency, High performance, Low power consumption are also critical in Inference. Low precision arithmetic is also incorporated because of its advantages. some of them are that it results in less energy usage (32b FP add consumes 0.9 pJ where as 8b add consumes only 0.03 pJ). Also storage capacity of the cache increases, computation becomes faster by extracting more parallelism, Memory throughput also increases. CHaiDNN which is a deep Neural Network library for acceleration of deep neural networks on Xilinx UltraScale+ MPSoCs by Xilinx is one such library tuned for Convolutional Networks which is customized in terms of memory and power consumption. It makes use of the Powerful tool SDSoC. CHaiDNN allows a user who has minimal knowledge on Machine learning and also not aware of knowing how to design a CNN on Hardware to get that feel of running an inference of a ML model. CHaiDNN supports classification, segmentation and detection.CHaiDNN supports only caffe. The goal behind this project is to implement the Rehabnet architecture using CHaiDNN. As a proof of concept, Healthy data is taken into consideration and Rehabnet architecute is trained in Caffe and the inference is done using CHaiDNN on Zynq Ultrascale+ MPSoC. An overview of the Literature Survey is given in the first chapter. A brief introduction on CNN and how to train an architecture is given in second and third chapters. In the fourth chapter, how to use CHaiDNN for inference is explained in detail. Future Scope of the work is given in the last chapter.

[error in script]

IITH Creators: