Object Classification using Deep Learning on Extremely Low-Resolution Time-of-Flight Data

This paper proposes two novel deep learning models for 2D and 3D classification of objects in extremely low-resolution time-of-flight imagery. The models have been developed to suit contemporary range imaging hardware based on a recently fabricated Single Photon Avalanche Diode (SPAD) camera with 64 χ 64 pixel resolution. Being the first prototype of its kind, only a small data set has been collected so far which makes it challenging for training models. To bypass this hurdle, transfer learning is applied to the widely used VGG-16 convolutional neural network (CNN), with supplementary layers added specifically to handle SPAD data. This classifier and the renowned Faster-RCNN detector offer benchmark models for comparison to a newly created 3D CNN operating on time-of-flight data acquired by the SPAD sensor. Another contribution of this work is the proposed shot noise removal algorithm which is particularly useful to mitigate the camera sensitivity in situations of excessive lighting. Models have been tested in both low-light indoor settings and outdoor daytime conditions, on eight objects exhibiting small physical dimensions, low reflectivity, featureless structures and located at ranges from 25m to 700m. Despite antagonist factors, the proposed 2D model has achieved 95% average precision and recall, with higher accuracy for the 3D model.