The L-CAS dataset was collected by a Velodyne VLP-16 3D LiDAR, mounted at a height of 0.8 m from the floor on the top of a Pioneer 3-AT robot, in one of the main buildings (a large indoor public space, including a canteen, a coffee shop and the resting area) of Lincoln University, UK. This dataset captures new research challenges for indoor service robots including human groups, children, people with trolleys, etc. Similar to the STRANDS dataset, the data were recorded in the sensor reference frame, and all human detections and tracks were then transformed to the world frame. We conducted our experiments on the first 19 minutes of data, in which 935 pedestrian trajectories were extracted.
Trajectory examples in the L-CAS dataset including: extracted pedestrian trajectories (left), detected point clusters (middle), and trajectories heatmap (right). In the heatmap, warmer colors indicate higher frequencies of pedestrian occupancy. The map is normalized between 0 and 1.
Following the previous works, the Average Displacement Error (ADE) is used to measure the error of predicted positions, which is computed using mean square error over all predicted position and ground truth positions as:
As our approach predicts the 3DOF pose rather than 2D position, the Average Eulerian angle Difference error (AEDE) is used to measure the rotation loss. We first convert the prediction and the ground truth quaternion to Eulerian angles, and calculate the absolute error of the yaw angle as:
The 3DOF trajectory dataset can be downloaded [here]
The raw Velodyne data can be found [here]
If you are interested in our research or have questions, pls contact Li Sun (Kevin) via: lisunsirATgmail.com