This paper evaluates the effectiveness of Convolutional Neural Networks (CNNs) in recognizing fitness activities using Inertial Measurement Unit (IMU) data. The study introduces a new dataset of running exercises recorded by 20 participants wearing IMUs on their ankles and wrists. Four CNN architectures—Deep-CNN, FCN, ResNet, and Scaling-FCN—are adapted to the fitness activity recognition task (FAR). Traditional machine learning methods, including Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN), are also evaluated for comparison. The performance of these models is assessed under different conditions, such as varying input data sizes and sensor counts. The results show that CNNs achieve at least 94% test accuracy, with the Scaling-FCN achieving the highest accuracy of 99.86% on the test set when only data from a single foot sensor is used. The study concludes that CNNs are well-suited for fitness activity recognition, and selective sensor selection can significantly improve performance, although traditional ML methods can still compete or outperform CNNs when favorable input data are utilized.This paper evaluates the effectiveness of Convolutional Neural Networks (CNNs) in recognizing fitness activities using Inertial Measurement Unit (IMU) data. The study introduces a new dataset of running exercises recorded by 20 participants wearing IMUs on their ankles and wrists. Four CNN architectures—Deep-CNN, FCN, ResNet, and Scaling-FCN—are adapted to the fitness activity recognition task (FAR). Traditional machine learning methods, including Random Forest (RF), Support Vector Machine (SVM), and K-Nearest Neighbors (K-NN), are also evaluated for comparison. The performance of these models is assessed under different conditions, such as varying input data sizes and sensor counts. The results show that CNNs achieve at least 94% test accuracy, with the Scaling-FCN achieving the highest accuracy of 99.86% on the test set when only data from a single foot sensor is used. The study concludes that CNNs are well-suited for fitness activity recognition, and selective sensor selection can significantly improve performance, although traditional ML methods can still compete or outperform CNNs when favorable input data are utilized.