Page - 101 - in Proceedings - OAGM & ARW Joint Workshop 2016 on "Computer Vision and Robotics“

regression forest to figure out what this means in term of object movement. In our experiments, the use of these single set of points has proven to lead to a highly stable tracking performance. Pose Forecast. Tracking is performed on a frame by frame basis, by checking how depth values from sample points of the current positions vary when compared to the depth values of the current depth map. This means when an object is moving fast between two frames so that no or only few sample points of the previous position overlaps with the new position, the object cannot be tracked. In our algorithm we use the previously estimated movement prediction as a starting guess for the new movement prediction. This allows for objects to move further between two frames, and also provides a more accurate initial guess for the pose estimation. Initially when performing object localization, a movement of zero is assumed. 3.3. Combined Object Localization with Tracking The full object tracking framework as shown in Fig. 1 combines both global object localization and tracking. The goal of this framework is to produce a continuous, low latency stream of the current object position. Whenever possible the system uses the fast tracking described above, and if tracking was not successful it switches back to the slower but global object localization. To track multiple objects simultaneously this tracking framework is run in parallel for each object. From Object Localization to Tracking. Depending on the configuration of the object localization, it is possible to find multiple instances of the same object in a single frame. For our use case we assume that only a single instance is the correct one. To determine which of these instances the correct one is, we perform two checks. First, the instances are ranked by the number of inliers (r), and only if it passes a certain threshold it is considered as a potential correct pose. For each pose the multi forest tracker is evaluated and tracking is performed. When running the tracking algorithm on a correct pose, the estimated tracking transformation should be a minimal movement (m). When the tracking algorithm would estimate a large movement, this means that either this current position is wrong or cannot be tracked successfully. We use this tracking movement as an additional filtering criterion. In practice, we use both parameters to form a single sorting criterion quality (q) in (1): 𝑞 = 𝑟 𝑚 2      We rank all candidate poses by this criterion, which leads to more accurate results than the use of either one of the criterion separately. Figure 1. Object tracking framework. The framework first performs global object localization on the input depth data. After the detection results are filtered they are passed to the Object Tracking module. If tracking was successful, the framework will directly use the tracking module to track objects for the next sensor input. If not, it will revert back to global object localization for the next sensor input frame. The tracking results are published to a higher level system. Tracking Verification. After tracking has been performed, we calculate the movement of the object with respect to the camera position. To determine if tracking was successful, we calculate if the current movement is reasonably realistic. We do this by calculating the acceleration of the object within the last 3 frames. If the acceleration is above a threshold, we consider the tracking as 101