Object-Specific Time-to-Collision Estimates from Monocular Vision for Drones

This paper presents three Time-to-Collision (TTC) estimation algorithms combining non-object specific low-level vi-sual estimates known from visual geometry with high-level visual object detection and scene depth inference using lightweight deep neural networks. All methods are two-stage models that first detect a specific object, then segment a region of interest on the object to extract visual cues that are relevant for TTC estimation. More precisely, sparse optical flow, bounding box sizes, and scaled monocular dense depth maps are used as visual cues for TTC calculation. The algorithms are evaluated on real data giving insights which visual cues are more reliable applying a statistical error analysis. It can be seen that the accuracy of the TTC calculation of the individual methods depends on the calculated TTC itself and that this dependency is distributed in a complementary manner across the methods. These results allow an assessment of how the individual methods can be combined to obtain more robust and accurate TTC estimates. This also contributes to how purely reactive visual navigation strategies can be improved.