A CNN Based Transfer Learning Model for Automatic Activity Recognition from Accelerometer Sensors Belkacem Chikhaoui 1 , Frank Gouineau 2 , Martin Sotir 2 1 Department of Science and Technology, TELUQ University, Canada belkacem.chikhaoui@teluq.ca 2 Computer Research Institute of Montreal, Montreal, Canada {Frank.Gouineau, Martin.Sotir}@crim.ca Abstract. Accelerometers are become ubiquitous and available in sev- eral devices such as smartphones, smartwaches, fitness trackers, and wearable devices. Accelerometers are increasingly used to monitor hu- man activities of daily living in different contexts such as monitoring activities of persons with cognitive deficits in smart homes, and mon- itoring physical and fitness activities. Activity recognition is the most important core component in monitoring applications. Activity recogni- tion algorithms require substantial amount of labeled data to produce satisfactory results under diverse circumstances. Several methods have been proposed for activity recognition from accelerometer data. However, very little work has been done on identifying connections and relation- ships between existing labeled datasets to perform transfer learning for new datasets. In this paper, we investigate deep learning based transfer learning algorithm based on convolutional neural networks (CNNs) that takes advantage of learned representations of activities of daily living from one dataset to recognize these activities in different other datasets characterized by different features including sensor modality, sampling rate, activity duration and environment. We experimentally validated our proposed algorithm on several existing datasets and demonstrated its performance and suitability for activity recognition. Transfer learning, deep learning, accelerometer data, activity recognition, CNN. 1 Introduction Human activity recognition is a challenging and well-researched problem [1]. With the emergence of wearable devices and accelerometers, activity recogni- tion is applied in different domains including assisted living, healthcare, sport, human-computer interaction, smart cities, and security [2, 3]. Most researchers use machine learning algorithms for activity recognition, which require substan- tial amount of labeled data to produce satisfactory results under diverse cir- cumstances, and significant efforts are required to apply the learned models to