Is feature averaging is not a valid method for multimodal feature fusion

Question

I have implemented a two-stream CNN architecture for human action recognition. The first stream uses a 1D-CNN to model accelerometer data, while the second stream is based on 2D