This paper presents a study on the application of convolutional neural networks (CNNs) for recognizing fitness activities using inertial measurement units (IMUs). The study evaluates the performance of CNNs and traditional machine learning (ML) methods on a newly collected dataset of running exercises performed by 20 participants. The dataset includes IMU data from four sensors (two ankles and two wrists) and consists of seven different running exercises. The study compares the performance of four CNN architectures—Deep-CNN, FCN, ResNet, and Scaling-FCN—with three traditional ML methods (Random Forest, SVM, and K-NN) under various input conditions, including different sensor configurations and data sizes. The results show that CNNs achieve high accuracy in recognizing fitness activities, with the Scaling-FCN reaching up to 99.86% test accuracy. While traditional ML methods, particularly SVMs, outperform CNNs in some scenarios, CNNs generally perform well across different input configurations. The study also investigates the impact of sensor data on model performance, finding that reducing the number of sensors can improve CNN performance but may negatively affect traditional ML methods. Additionally, the study highlights the importance of data preprocessing, including synchronization, standardization, and segmentation, in achieving accurate results. The study concludes that CNNs are well-suited for IMU-based fitness activity recognition and that performance can be further improved by selectively dropping sensors. The results suggest that CNNs can be effectively used in mobile fitness applications to provide real-time feedback on user activity. The dataset used in this study is publicly available for further research and evaluation.This paper presents a study on the application of convolutional neural networks (CNNs) for recognizing fitness activities using inertial measurement units (IMUs). The study evaluates the performance of CNNs and traditional machine learning (ML) methods on a newly collected dataset of running exercises performed by 20 participants. The dataset includes IMU data from four sensors (two ankles and two wrists) and consists of seven different running exercises. The study compares the performance of four CNN architectures—Deep-CNN, FCN, ResNet, and Scaling-FCN—with three traditional ML methods (Random Forest, SVM, and K-NN) under various input conditions, including different sensor configurations and data sizes. The results show that CNNs achieve high accuracy in recognizing fitness activities, with the Scaling-FCN reaching up to 99.86% test accuracy. While traditional ML methods, particularly SVMs, outperform CNNs in some scenarios, CNNs generally perform well across different input configurations. The study also investigates the impact of sensor data on model performance, finding that reducing the number of sensors can improve CNN performance but may negatively affect traditional ML methods. Additionally, the study highlights the importance of data preprocessing, including synchronization, standardization, and segmentation, in achieving accurate results. The study concludes that CNNs are well-suited for IMU-based fitness activity recognition and that performance can be further improved by selectively dropping sensors. The results suggest that CNNs can be effectively used in mobile fitness applications to provide real-time feedback on user activity. The dataset used in this study is publicly available for further research and evaluation.