Velocity estimation from monocular video for automotive applications using convolutional neural networks

We aim to determine the speed of ego-vehicle motion from a video stream. Previous work by Konda et al. [1] has shown that motion can be detected and quantified with the help of a synchrony autoencoder, which has multiplicative gating interactions introduced between its hidden units, and hence, across video frames. In this work we modify their synchrony autoencoder method to achieve a "real time" performance in a wide variety of driving environments. Our modifications led to a model which is 1.5 times faster and uses only half of the total memory by comparison with the original. We also benchmark the speed estimation performance against a model based on CaffeNet. CaffeNet is known for visual classification and localization but we employ its architecture with a little tweak for speed determination using sequential video frames and blur patterns. We evaluate our models on self-collected data, KITTI, and other standard sets.