src.dataset.sdd_dataset module

src.dataset.sdd_dataset module#

class src.dataset.sdd_dataset.SDDDataset(*args: Any, **kwargs: Any)[source]#

Bases: BaseDataset

Stanford Drone Dataset (SDD) loader.

This dataset contains top-down drone footage with multiple categories such as pedestrians, bicycles, skaters, and vehicles.

raw_fps = 30#
classmethod load_data(args: Namespace, data_path: str) → SDDDataset[source]#

Load data for a single SDD video scene.

Read annotations.txt, convert coordinates from pixels to meters, apply nontrivial data-cleaning logic such as abnormal-speed removal and trajectory smoothing, and load map.png if it exists; otherwise create a blank map.

Parameters:

  • args (Namespace) – Global arguments.

  • data_path (str) – Path to annotations.txt.

Returns:

Initialized dataset instance.

Return type:

SDDDataset

classmethod load_data_batch(args: Namespace, data_path: str, show_tqdm=True) → List[SDDDataset][source]#

Batch-load SDD datasets.

static get_homography_mat(data_path)[source]#

Get the homography matrix from the dataset directory structure, effectively capturing the pixel-to-meter ratio.

Different SDD scenes such as bookstore and deathCircle use different scale factors.

Parameters:

data_path (Path) – Data file path, used to infer the scene name.

Returns:

3x3 scaling matrix, effectively diagonal in this case.

Return type:

np.ndarray

static clean_trajectory(traj, median_k=5, do_savgol=True, sg_win=9, sg_poly=2, do_loess=False, loess_frac=0.08, speed_window=11, speed_k=5.0, angle_window=11, angle_k=5.0, theta_thresh=numpy.deg2rad, split_angle_thresh=numpy.deg2rad, split_speed_delta=0.5, split_time_thresh=1.0)[source]#

Static utility method for trajectory cleaning and smoothing.

Denoise and smooth the raw trajectory, using Savitzky-Golay or LOESS, detect speed and angle anomalies, and split the trajectory into reasonable segments.

Parameters:

  • traj (np.ndarray) – Raw trajectory (N, 3) with columns [t, x, y].

  • median_k (int) – Median-filter window size.

  • do_savgol (bool) – Whether to use Savitzky-Golay filtering.

  • ... (other smoothing and splitting threshold parameters)

Returns:

List of cleaned trajectory segments.

Return type:

List[np.ndarray]

Contents