This paper focuses on oxygen consumption (VO2) estimation using 6-axis motion data (3-axis acceleration and 3-axis angular velocity) that are obtained from small motion sensors attached to people playing sports with different intensities. In order to achieve high estimation accuracy over a wide range of intensities of exercises, we apply neural network that is trained by experimental data consisting of the measured VO2 and motion sensing data of people with a wide range of intensities of exercises. We first investigate the gain brought by applying neural network by comparing its accuracy with an approach based on the linear regression model. Then, we analyze how much improvement the information on angular velocity can bring as compared with the estimation with the acceleration data alone. Our numerical results show that the employed framework exploiting neural network can improve the estimation accuracy in comparison to the linear regression model and the exploitation of information on the angular velocity plays an important role to improve the accuracy over higher intensities of exercises.