In autonomous vehicles (AVs), early warning systems rely on collision\nprediction to ensure occupant safety. However, state-of-the-art methods using\ndeep convolutional networks either fail at modeling collisions or are too\nexpensive/slow, making them less suitable for deployment on AV edge hardware.\nTo address these limitations, we propose sg2vec, a spatio-temporal scene-graph\nembedding methodology that uses Graph Neural Network (GNN) and Long Short-Term\nMemory (LSTM) layers to predict future collisions via visual scene perception.\nWe demonstrate that sg2vec predicts collisions 8.11% more accurately and 39.07%\nearlier than the state-of-the-art method on synthesized datasets, and 29.47%\nmore accurately on a challenging real-world collision dataset. We also show\nthat sg2vec is better than the state-of-the-art at transferring knowledge from\nsynthetic datasets to real-world driving datasets. Finally, we demonstrate that\nsg2vec performs inference 9.3x faster with an 88.0% smaller model, 32.4% less\npower, and 92.8% less energy than the state-of-the-art method on the\nindustry-standard Nvidia DRIVE PX 2 platform, making it more suitable for\nimplementation on the edge.\n