Visual-Fusion

• LiDAR Fusion with Vision
• Data taken from KITTI Dataset
• Download Yolov4 model weights from here

Low-Level Fusion

Yolo Detections

Visualizing 3D LiDAR points in Open3D

3D Lidar Points Projected on the image plane

LiDAR points Fused with YOLO detections

• LiDAR points are projected on the image using camera instrinsic and extrinsic matrix
• The points that lie within the detected 2D Bounding Box by YOLO are stored and rest are ignored
• There are some outliers inside bboxes that do not belong to that category, to reject these outliers there are several ways.
• One way is to shrink the bounding box size so that the points that absolutely belong to the desired objects are only considered.
• Another way is to use the Sigma Rule, i.e include the points that are within 1 sigma or 2 sigma away from gaussian mean, based on the distance of points

Mid-Level Fusion

Yolo Detections

LiDAR Points projected on Image

3D Bounding Boxes From LiDAR

3D BBox converted to 2D BBox

LiDAR 2D BBox Fused with YOLO 2D BBox using Intersection Over Union

• 2D Bboxes from LiDAR are associated with YOLO 2D Bboxes using Hungarian Algorithm
• Green Bounding Boxes are detected by YOlO whereas Blue Bounding Boxes are calculated using LiDAR points
• YOLO missed 1 vehicle, whereas 2 vehicles are missed by LiDAR, one of which is half out of frame, at the bottom right side

File Structure

.
├── Code
|  ├── main.py
|  ├── Fusion.py
|  ├── Lidar2Camera.py
|  ├── YoloDetector.py
|  ├── Utils.py
|  ├── FusionUtils.py
|  ├── LidarUtils.py
|  ├── YoloUtils.py
├── Data
   ├── calibs
   |  ├── 000031.txt
   |  ├── 000035.txt
   |  ├── ...
   ├── images
   |  ├── 000031.png
   |  ├── 000035.png 
   |  ├── ...
   ├── labels
   |  ├── 000031.txt
   |  ├── 000035.txt 
   |  ├── ...
   ├── models
      ├── yolov4
      |  ├── yolov4.cfg
      |  ├── coco.names
   ├── output
      ├── images
      ├── videos
   ├── points
   |  ├── 000031.pcd
   |  ├── 000035.pcd 
   |  ├── ...

TODO

• Add Run Instructions
• Add Dependencies
• Add References
• High-Level Fusion