Constructive interference for Multi-view Time-of-Flight acquisition

An increasing number of computer vision applications, including medical systems, are using range imaging which require real-time and accurate 3D data acquisition. Time-of-Flight cameras allow us to acquire 3D data in real-time and at high frame-rate, however, the provided depth accuracy is in order of centimeters. This thesis describes a novel method for acquiring depth images using multi-view Time of Flight (ToF) cameras, which uses a constructive interference between the emitted signals to enhance the depth accuracy. This work proposes to combine the measurements of the multi-view cameras at the acquisition level, in opposite to approaches that filter, calibrate or do 3D reconstructions posterior to the image acquisition. This thesis presents an example using a pair of ToF cameras in stereo-set-up defining a three-stages procedure, in which the infrared lighting of the scene is actively modified: first, the two cameras emit an infrared signal one after the other (stages 1 and 2), and then, simultaneously (stage 3). The third stage is where the constructive interference between the two cameras is produced. These redundant measurements are optimized to obtain the enhanced depth information. Based on a simulation of the ToF cameras, a quantitative evaluation of the proposed method is provided. The performance of this novel 3D acquisition method is evaluated for different objects and configurations. Results on real images using different optimization framework are also presented. The acquisition is optimized first for each camera and then within the complete 3D reconstruction process. Both simulation and real images prove that the proposed Stereo ToF camera acquisition produces more accurate depth measurements. We then extend this concept and formulation from two ToF cameras to a multi-view ToF cameras set up. This thesis also presents two applications of Monocular Simultaneous Localization and Mapping (MonoSLAM). The first application is using MonoSLAM in endoscopic images obtaining a rough idea of the 3D scene observed by the endoscope. The proposed method applies a pre-processing step in order to enhance the texture information in the endoscopic images which normally have poor texture. The second application involves a RGB-D sensor as ToF cameras combined with a High-resolution cameras or Kinect. An extension of the traditional Monocular Simultaneous Localization and Mapping (MonoSLAM) for 3D reconstruction of the scene is also presented. The traditional Monocular SLAM (MonoSLAM) is extended in order to take advantage of the new depth data provided by the RGB-D sensors. The algorithm produces a sparse-map. The quantitative analysis shows that our proposed method improves the recovered trajectory of the camera in comparison to using only the traditional MonoSLAM (RGB). To simulate the RGB-D sensor, we use a ToF camera and an RGB camera in stereo set-up; and to define the ground-truth, a commercial optic tracking system is used. Note that this proposed method would function with any high-frame rate 3D acquisition modality including the Multi-view ToF system introduced in this thesis. Finally, a discussion about the feasibility of each proposed method and how they can be used in real-world applications is provided. Details of the implementation and results of the multi-view ToF acquisition process are then presented, demonstrating that acquisition and processing of the proposed constructive interferences result in considerable improvement of 3D acquisition and reconstruction accuracy.