DRAD: A new model for dynamic real-time Avalanche detection from videos with residual depth-separable convolution and feature pyramid networks | Synapse