Temporal 3D Fully Connected Network for Water-Hazard Detection

Abstract

Water on road is a potential hazard for road vehicles. In complex traffic, it is difficult for drivers to safely share the road with other vehicles and pedestrians while avoiding water puddles and pot holes. Running into these could potentially damage electronic components and corrode metal parts of the car, or even lead to loss of control. Similarly, driver-less cars also need to detect water puddles and plan safe path around them or slow down if it deems to be safe to do so. Such detection needs to be both accurate and fast enough for realtime assistance or planning. We present a new temporal 3D fully convolutional network called T3D-FCN for water detection that exploit temporal information to achieve accuracy comparable to the-state-of-the-art accuracy while requires less computation resources. We also show that by adding temporal information the detection performance significantly improves with a small additional computation as compared to single image detection using a similar network architecture.