Vehicle Detection in Infrared Imagery Using Neural Networks with Synthetic Training Data

Abstract

We present a new approach to the detection, localization, and recognition of vehicles in infrared imagery using a deep Convolutional Neural Network that completely avoids the need for manually-labelled training data by using synthetic imagery and a transfer learning strategy. Synthetic imagery is generated from CAD models using a rendering tool, allowing the network to be trained against a complete set of vehicle aspects and with automatically generated meta-data encoding the position of the vehicle in the image. The proposed approach is fast since a single network is used to compute class probabilities for individual pixels in the image. Results are presented illustrating the robust recognition and localization performance achievable with the novel approach for vehicle detection in real high-resolution infrared imagery.