perception_stack.py

# Semester Project: Perception Stack for Autonomous Driving
# Maksim Pikovskiy <mp8671@rit.edu>

import os
import time
import numpy as np
import open3d as o3d

# Print num of clusters per frame and vehicle info per frame
DEBUG_MODE = False

# Save latencies for each frame and average latency
SAVE_LATENCIES = False

# Path to where the Point Cloud frames are stored
path_frames = "dataset/PointClouds"

# Assuming the max height of a normal vehicle is 4.75 meters (with some overhead)
# Normal car: ~ 2 meters
# Normal truck: ~ 4 meters
car_height = 4.75
# Assuming the ground level is at 0 meters
ground_level = 0.005

# Threshold for determining if two vehicles are the same
distance_threshold = np.asarray([2, 2, 2])


def read_and_preprocess(filename):
    """
    Read a PCD file and preprocess it for vehicle detection.

    Parameters:
    - filename: Path to the PCD file.

    Returns:
    - pcd: Processed point cloud.
    """
    # Read point cloud
    pcd = o3d.io.read_point_cloud(filename)

    # Down sample to increase speed of processing
    # pcd = pcd.voxel_down_sample(voxel_size=0.075) # results in more clusters and more false positives

    # Alternative for downsampling
    # Remove points out possible range for cars and trucks
    pcd = pcd.select_by_index(np.where(np.asarray(pcd.points)[:, 2] < car_height)[0])
    pcd = pcd.select_by_index(np.where(np.asarray(pcd.points)[:, 2] > ground_level)[0])

    # pcd, _ = pcd.remove_statistical_outlier(nb_neighbors=3, std_ratio=2.0)

    return pcd


def detect_moving_objects(curr_frame, prev_frame):
    """
    Detect moving objects in the current frame using background subtraction.

    Parameters:
    - curr_frame: Current point cloud frame.
    - prev_frame: Previous point cloud frame.

    Returns:
    - frame_diff: Point cloud containing only the points that have changed.
    """

    # Compute distances between points that near each other
    dists = curr_frame.compute_point_cloud_distance(prev_frame)
    dists = np.asarray(dists)
    ind = np.where(dists > 0.0020)[0]  # get only distances that are significant

    # Get the points that have changed
    frame_diff = curr_frame.select_by_index(ind)

    frame_diff.estimate_normals()

    # Delete points with no 5 neighbors in 2 meter radius
    # frame_diff, _ = frame_diff.remove_radius_outlier(nb_points=5, radius=2)

    return frame_diff


def get_vehicles(frame):
    """
    Get vehicles from a point cloud frame using DBSCAN.

    Parameters:
    - frame: Point cloud frame.

    Returns:
    - vehicles: List of clusters representing vehicles.
    """
    # Check if there are any points in the frame
    if len(frame.points) == 0:
        return None

    # Cluster the points using DBSCAN with eps of 1.75 and min_points of 3
    labels = np.array(frame.cluster_dbscan(eps=1.75, min_points=3))

    # Check if there are any clusters generated by DBSCAN
    if len(labels) == 0:
        return None

    vehicles = []
    max_label = labels.max()  # Get the max label

    if DEBUG_MODE:
        print(f"\t Point Cloud has {max_label + 1} clusters")

    # Iterate through all the clusters and add them to the list of vehicles
    for i in range(max_label + 1):
        cluster = frame.select_by_index(np.where(labels == i)[0])
        vehicles.append(cluster)

    return vehicles


def compute_bounding_box(cluster):
    """
    Compute the bounding box for a cluster.

    Parameters:
    - cluster: Cluster representing a vehicle.

    Returns:
    - bounds: Bounding box coordinates [x_min, y_min, z_min, x_max, y_max, z_max].
    """
    # Get the min and max values for each axis of the cluster
    bounds = [
        np.min(np.asarray(cluster.points)[:, 0]),
        np.min(np.asarray(cluster.points)[:, 1]),
        np.min(np.asarray(cluster.points)[:, 2]),
        np.max(np.asarray(cluster.points)[:, 0]),
        np.max(np.asarray(cluster.points)[:, 1]),
        np.max(np.asarray(cluster.points)[:, 2]),
    ]

    return bounds


def create_bounding_box(bbox):
    """
    Create a line set representing a bounding box.

    Parameters:
    - bbox: Bounding box coordinates [x_min, y_min, z_min, x_max, y_max, z_max].

    Returns:
    - line_set: Open3D LineSet representing the bounding box.
    """
    # Create a line set representing the bounding box
    lines = [
        [0, 1],
        [1, 2],
        [2, 3],
        [3, 0],
        [4, 5],
        [5, 6],
        [6, 7],
        [7, 4],
        [0, 4],
        [1, 5],
        [2, 6],
        [3, 7],
    ]

    # Set the color of the lines to red
    colors = [[1, 0, 0] for _ in range(len(lines))]
    # Create the line set
    line_set = o3d.geometry.LineSet()
    # Set the points and lines of the line set
    line_set.points = o3d.utility.Vector3dVector(
        [
            [bbox[0], bbox[1], bbox[2]],
            [bbox[3], bbox[1], bbox[2]],
            [bbox[3], bbox[4], bbox[2]],
            [bbox[0], bbox[4], bbox[2]],
            [bbox[0], bbox[1], bbox[5]],
            [bbox[3], bbox[1], bbox[5]],
            [bbox[3], bbox[4], bbox[5]],
            [bbox[0], bbox[4], bbox[5]],
        ]
    )
    # Set the lines and colors of the line set
    line_set.lines = o3d.utility.Vector2iVector(lines)
    line_set.colors = o3d.utility.Vector3dVector(colors)

    return line_set


def get_position(bounds):
    """
    Get the position of a vehicle. Calculate the center of the bounding box.

    Parameters:
    - bounds: Bounding box coordinates [x_min, y_min, z_min, x_max, y_max, z_max].

    Returns:
    - position: Position of the vehicle [x, y, z] (center of bounding box).
    """
    # Calculate the center of the bounding box
    position = [
        (bounds[0] + bounds[3]) / 2,
        (bounds[1] + bounds[4]) / 2,
        (bounds[2] + bounds[5]) / 2,
    ]

    return position


def calculate_volume(bounds):
    """
    Calculate the volume of a bounding box for a vehicle.

    Parameters:
    - bounds: Bounding box coordinates [x_min, y_min, z_min, x_max, y_max, z_max].

    Returns:
    - volume: Volume of the bounding box for a vehicle.
    """
    # Calculate the length, width, and height of the bounding box
    length = abs(bounds[3] - bounds[0])
    width = abs(bounds[4] - bounds[1])
    height = abs(bounds[5] - bounds[2])

    return length * width * height


def save_results(detected_vehicles, frame_number):
    """
    Save detected vehicle information to a CSV file.

    Parameters:
    - detected_vehicles: Information about vehicles detected in the current frame.
    - frame_number: Current frame number.
    """
    # Create a directory for saving the results
    result_dir = "perception_results"
    os.makedirs(result_dir, exist_ok=True)

    # Create a filename for the CSV file
    csv_filename = "frame_{i}.csv".format(i=frame_number)
    csv_path = os.path.join(result_dir, csv_filename)

    # Extract relevant information for saving
    ids = ["vehicle_id"] + [vehicle["id"] for vehicle in detected_vehicles]
    position_x = ["position_x"] + [
        vehicle["position"][0] for vehicle in detected_vehicles
    ]
    position_y = ["position_y"] + [
        vehicle["position"][1] for vehicle in detected_vehicles
    ]
    position_z = ["position_z"] + [
        vehicle["position"][2] for vehicle in detected_vehicles
    ]
    mvec_x = ["mvec_x"] + [vehicle["motion_vector"][0] for vehicle in detected_vehicles]
    mvec_y = ["mvec_y"] + [vehicle["motion_vector"][1] for vehicle in detected_vehicles]
    mvec_z = ["mvec_z"] + [vehicle["motion_vector"][2] for vehicle in detected_vehicles]
    bbox_x_min = ["bbox_x_min"] + [
        vehicle["bounding_box"][0] for vehicle in detected_vehicles
    ]
    bbox_x_max = ["bbox_x_max"] + [
        vehicle["bounding_box"][3] for vehicle in detected_vehicles
    ]
    bbox_y_min = ["bbox_y_min"] + [
        vehicle["bounding_box"][1] for vehicle in detected_vehicles
    ]
    bbox_y_max = ["bbox_y_max"] + [
        vehicle["bounding_box"][4] for vehicle in detected_vehicles
    ]
    bbox_z_min = ["bbox_z_min"] + [
        vehicle["bounding_box"][2] for vehicle in detected_vehicles
    ]
    bbox_z_max = ["bbox_z_max"] + [
        vehicle["bounding_box"][5] for vehicle in detected_vehicles
    ]

    # Save the results to a CSV file
    np.savetxt(
        csv_path,
        [
            p
            for p in zip(
                ids,
                position_x,
                position_y,
                position_z,
                mvec_x,
                mvec_y,
                mvec_z,
                bbox_x_min,
                bbox_x_max,
                bbox_y_min,
                bbox_y_max,
                bbox_z_min,
                bbox_z_max,
            )
        ],
        delimiter=",",
        fmt="%s",
    )


def track_vehicles(vehicles, tracked_vehicles):
    """
    Track vehicles across frames. Matches the vehicles detected in the current
    frame with the vehicles detected in the previous frame. If a match is found,
    the vehicle is assigned the same ID as the vehicle in the previous frame. If
    no match is found, a new ID is assigned to the vehicle. If there are more than
    6 vehicles detected in the current frame, the vehicles with the smallest volume
    are removed.

    Parameters:
    - vehicles: Information about vehicles detected in the current frame.
    - tracked_vehicles: Information about vehicles detected in previous frames.

    Returns:
    - new_vehicles: Information about vehicles detected in the current frame.
    """
    # Check if there are any vehicles detected in the current frame
    new_vehicles = []
    if tracked_vehicles is None or len(tracked_vehicles) == 0:
        # Assign IDs to the vehicles of initial frame
        for idx, vehicle in enumerate(vehicles):
            if len(vehicles) == 6:
                break
            vehicle["id"] = idx
            vehicle["color"] = np.random.rand(3)
            new_vehicles.append(vehicle)

    else:
        # Match vehicles in current frame with vehicles in previous frame
        detected_vehicles = []
        for vehicle in vehicles:
            has_match = False
            for veh_track in tracked_vehicles:
                # Calculate the distance between the vehicles
                distance = abs(
                    np.asarray(vehicle["position"]) - np.asarray(veh_track["position"])
                )
                # Check if the distance is less than the threshold
                if (
                    np.all(distance < distance_threshold)
                    # and veh_track["id"] not in detected_vehicles
                ):
                    vehicle["id"] = veh_track["id"]
                    vehicle["color"] = veh_track["color"]
                    detected_vehicles.append(veh_track["id"])
                    new_vehicles.append(vehicle)
                    has_match = True
                    break
            # If no match is found, assign a new ID to the vehicle
            # and add it to the list of new vehicles
            if not has_match and len(new_vehicles) <= 6:
                vehicle["id"] = len(new_vehicles)
                vehicle["color"] = np.random.rand(3)
                new_vehicles.append(vehicle)

        # Get the vehicles that were not detected in the current frame
        undetected_vehicles = [
            vehicle
            for vehicle in tracked_vehicles
            if vehicle["id"] not in detected_vehicles
        ]

        # Update the position and bounding box of the undetected vehicles
        # Based on the motion vector (Unreliable due to loss of points in blind spot)
        # for undetected_vehicle in undetected_vehicles:
        #     undetected_vehicle["position"] = np.array(
        #         undetected_vehicle["position"]
        #     ) + np.array(undetected_vehicle["motion_vector"])

        #     undetected_vehicle["bounding_box"][0] += undetected_vehicle[
        #         "motion_vector"
        #     ][0]
        #     undetected_vehicle["bounding_box"][3] += undetected_vehicle[
        #         "motion_vector"
        #     ][0]
        #     undetected_vehicle["bounding_box"][1] += undetected_vehicle[
        #         "motion_vector"
        #     ][1]
        #     undetected_vehicle["bounding_box"][4] += undetected_vehicle[
        #         "motion_vector"
        #     ][1]
        #     undetected_vehicle["bounding_box"][2] += undetected_vehicle[
        #         "motion_vector"
        #     ][2]
        #     undetected_vehicle["bounding_box"][5] += undetected_vehicle[
        #         "motion_vector"
        #     ][2]

        #     new_vehicles.append(undetected_vehicle)

        # Add the undetected vehicles to the list of new vehicles
        new_vehicles.extend(undetected_vehicles)

        # Remove vehicles with small volume if there are more than 6 vehicles
        if len(new_vehicles) > 6:
            for vehicle in new_vehicles:
                if calculate_volume(vehicle["bounding_box"]) < 1.0:
                    new_vehicles.remove(vehicle)

    return new_vehicles


def get_similar_vehicle(vehicles, vehicle_position):
    """
    Get a vehicle from a list of vehicles that is similar to the vehicle
    with the given position.

    Parameters:
    - vehicles: List of vehicles.
    - vehicle_position: Position of the vehicle to find a match for.

    Returns:
    - veh: Vehicle that is similar to the vehicle with the given position.
    """
    # Match vehicle in current frame with similar vehicle in previous frame
    # using the distance between the vehicles' positions
    for veh in vehicles:
        distance = abs(np.array(vehicle_position) - np.array(veh["position"]))
        if np.all(distance < distance_threshold):
            return veh
    return None


def get_vehicle_infos(vehicles, prev_vehicles_info):
    """
    Get information about vehicles detected in the current frame.
    Information is id, position, bounding box, and motion vector.

    Parameters:
    - vehicles: List of clusters representing vehicles.
    - prev_vehicles_info: Information about vehicles detected in previous frame.

    Returns:
    - vehicles_info: Information about vehicles detected in the current frame.
    """
    vehicles_info = []
    for idx, vehicle in enumerate(vehicles):
        # Compute the bounding box and position for the vehicle
        bounds = compute_bounding_box(vehicle)
        position = get_position(bounds)

        # Compute the motion vector for the vehicle
        motion_vector = [0, 0, 0]
        if prev_vehicles_info is not None:
            prev_vehicle = get_similar_vehicle(prev_vehicles_info, position)
            if prev_vehicle is not None:
                motion_vector = np.array(prev_vehicle["position"]) - np.array(position)

        # Check if the volume of the bounding box is greater than 0.0005
        if calculate_volume(bounds) > 0.0005:
            if DEBUG_MODE:
                print(
                    f"\t Vehicle {idx + 1} - Bounding Box: {bounds} - Position: {position} - Motion Vector: {motion_vector}"
                )

            # Add the vehicle information to the list of vehicles information
            vehicles_info.append(
                {
                    "id": idx,
                    "position": position,
                    "bounding_box": bounds,
                    "motion_vector": motion_vector,
                }
            )

    return vehicles_info


def main():
    """
    Retrieves point cloud frames from the dataset folder and processes them
    to detect vehicles. Vehicles are tracked across frames and bounding boxes
    are created for each vehicle. Furthermore, the vehicles and their bounding
    boxes are visualized in an animated fashion. The results are saved to a CSV file.

    Also calculates the latency for each frame and the average latency.

    """
    # Get the list of PCD files and sort them numerically
    pcd_files = os.listdir(path_frames)
    pcd_files.sort(key=lambda x: int(os.path.splitext(x)[0]))

    # Create a point cloud for visualization and a visualizer
    geometry = o3d.geometry.PointCloud()
    vis = o3d.visualization.Visualizer()
    vis.create_window()
    vis.add_geometry(geometry)

    # Does not work for Open3d 0.17.0, Python 3.10.11
    # Settings are not adjusted to fit the scene
    # ctr = vis.get_view_control()
    # ctr.set_zoom(0.01)
    # ctr.set_lookat([0, 0, 0])
    # ctr.set_front([0, 0, -1])
    # ctr.set_up([0, 1, 0])

    # List of latencies for each frame
    latencies = []

    # Previous frame
    prev_frame = None

    # Previous moving objects and tracked vehicles
    prev_moving_objects = None
    tracked_vehicles = None
    # Iterate through all the PCD files
    for i in range(len(pcd_files)):
        start_time = time.time()  # Start timer for latency

        if DEBUG_MODE:
            print(pcd_files[i])

        # If it is the first frame, read the next frame instead of previous frame
        if i == 0:
            prev_frame = read_and_preprocess(
                os.path.join(path_frames, pcd_files[i + 1])
            )
        # Read the current frame and preprocess it
        curr_frame = read_and_preprocess(os.path.join(path_frames, pcd_files[i]))

        # Detect moving objects in the current frame compared to the previous frame
        moving_objects = detect_moving_objects(curr_frame, prev_frame)

        # Get the vehicles from the moving objects
        vehicles = get_vehicles(moving_objects)

        vehicles_info = []
        # Check if there are any vehicles detected in the current frame
        if vehicles is None or len(vehicles) == 0:
            # If no vehicles are detected, use the previous moving objects and tracked vehicles
            moving_objects = prev_moving_objects
            vehicles_info = tracked_vehicles
        else:
            # Get information about the vehicles detected in the current frame
            vehicles_info = get_vehicle_infos(vehicles, tracked_vehicles)

        geometries = []
        # Update the tracked vehicles
        tracked_vehicles = track_vehicles(vehicles_info, tracked_vehicles)

        # Calculate the latency after getting all the information and before visualization
        # end_time = time.time()

        # Create a cluster for each vehicle and a bounding box for each vehicle
        for vehicle in tracked_vehicles:
            # Get the points within the bounding box
            points = np.asarray(curr_frame.points)
            cluster_indices = np.where(
                (points[:, 0] >= vehicle["bounding_box"][0])
                & (points[:, 0] <= vehicle["bounding_box"][3])
                & (points[:, 1] >= vehicle["bounding_box"][1])
                & (points[:, 1] <= vehicle["bounding_box"][4])
                & (points[:, 2] >= vehicle["bounding_box"][2])
                & (points[:, 2] <= vehicle["bounding_box"][5])
            )[0]

            # Create a point cloud for the points within the bounding box
            veh_cluster = o3d.geometry.PointCloud()
            veh_cluster.points = o3d.utility.Vector3dVector(points[cluster_indices])

            # Color each vehicle's points differently
            veh_cluster.paint_uniform_color(vehicle["color"])

            # Add the cluster of vehicle and bounding box to the list of geometries
            geometries.append(veh_cluster)
            geometries.append(create_bounding_box(vehicle["bounding_box"]))

        # Visualize moving objects and bounding boxes
        vis.clear_geometries()
        for geo in geometries:
            vis.add_geometry(geo)

        # Does not work for Open3d 0.17.0, Python 3.10.11
        # Settings are not adjusted to fit the scene
        # ctr.set_zoom(0.01)
        # ctr.set_lookat([0, 0, 0])
        # ctr.set_front([0, 0, -1])
        # ctr.set_up([0, 1, 0])

        vis.poll_events()
        vis.update_renderer()

        # Update the previous frame and moving objects
        prev_frame = curr_frame
        prev_moving_objects = moving_objects

        # Save the results to a CSV file
        save_results(tracked_vehicles, i)

        # Calculate the latency for the current frame
        end_time = time.time()  # End timer for latency including visualization
        latency = round(end_time - start_time, 5)
        print(f"Frame {i} - Latency: {latency} seconds")

        # Add the latency to the list of latencies
        latencies.append(latency)

    # Print the average latency
    print(f"\nAverage latency: {round(np.mean(latencies), 5)} seconds")

    # Save the latencies to a text file
    if SAVE_LATENCIES:
        formatted_latencies = [
            "Frame " + str(idx) + " - Latency: " + str(latency) + " seconds"
            for idx, latency in enumerate(latencies)
        ]

        formatted_latencies.append(
            f"\nAverage latency: {round(np.mean(latencies), 5)} seconds"
        )

        np.savetxt(
            "latencies.txt",
            formatted_latencies,
            delimiter=",",
            fmt="%s",
        )


# Run the main function
if __name__ == "__main__":
    main()