Wearable sensor-based human activity recognition (HAR) plays a significant role in the current smart applications with the development of the theory of artificial intelligence and popularity of the wearable sensors. Salient and discriminative features improve the performance of HAR. To capture the local dependence over time and space on the same axis from multi-location sensor data on convolutional neural networks (CNN), which is ignored by existing methods with 1D kernel and 2D kernel, this study proposes two methods, T-2D and M-2D. They construct three activity images from each axis of multi-location 3D accelerometers and one activity image from the other sensors. Then it implements the CNN networks named T-2DCNN and M-2DCNN based on T-2D and M-2D respectively, which fuse the four activity image features for the classifier. To reduce the number of the CNN weight, the weight-shared CNN, TS-2DCNN and MS-2DCNN, are proposed. In the default experiment settings on public datasets, the proposed methods outperform the existing methods with the F₁-value increased by 6.68% and 1.09% at most in OPPORTUNITY and SKODA respectively. It concludes that both naïve and weight-shared model have better performance in most F₁-values with different number of sensors and F₁-value difference of each class.