Human Activity Recognition (HAR) enables context-aware applications using sensor data, supporting advancements in IoT-based health monitoring, smart environments, and ambient assisted living. With the growing use of smartphone sensors, time-series data from accelerometers and gyroscopes have become a primary modality for activity classification. This study presents a deep learning-based HAR framework employing both raw time-series and image-transformed representations to assess the comparative performance of 1D and 2D Convolutional Neural Networks (CNNs), alongside Transfer Learning (TL) techniques using well-known pretrained models such as ResNet50 and VGG16. A large-scale dataset of 1.5 million tri-axial accelerometer records was collected from five users performing five daily activities. Raw data were used to train a 1D-CNN, while sliding-window transformations enabled 2D-CNN evaluation, confirming the effectiveness of image-based representations for activity recognition. Building on this, TL was applied by fine-tuning pretrained models at varying input image resolutions. Among the evaluated models, VGG16 at 300 \(\times \) 300 resolution achieved the highest classification accuracy of 98.70%, demonstrating the effectiveness of TL in retaining fine-grained discriminative features. These results highlight the strong generalization capability of pretrained CNNs on image-based sensor data, advocating their suitability for robust, reliable, and scalable activity recognition in real-world IoT-driven applications.