/GPU-Accelerated-Monocular-Visual-Odometry

Monocular Visual odometry pipeline using C++, CUDA and OpenCV on the KITTI odometry dataset.

Primary LanguageC++

GPU accelerated Monocular Visual Odometry

  1. A Monocular Visual odometry pipeline using Modern C++, CUDA and OpenCV on the KITTI odometry dataset.

  2. ORB feature computation is implemented using CUDA kernels and compares performance improvements. CPU or GPU version can be chosen using input arguments.

    cudaVO-2024-03-03_21 53 14-ezgif com-crop (1)

Pipeline Overview:

  1. FAST Keypoint Detection: Detect keypoints in the images using the FAST algorithm.
  2. ORB Descriptor Computation: Compute ORB descriptors for the detected keypoints.
  3. Feature Matching: Match keypoints between images using a matching algorithm.
  4. Essential Matrix Estimation using RANSAC: Estimate the Essential Matrix (E) using RANSAC to handle outliers.
  5. Compute Pose from E: Extract camera pose information from the estimated Essential Matrix (E).
  6. Pose Tracking: Track the camera pose over time using the computed poses.

Dependencies

  • OpenCV 4.2.0
  • CUDA

Usage

mkdir build
cmake -DCMAKE_CUDA_ARCHITECTURES= $your_architecture ..
make

  • CPU_VO: Run CPU Visual Odometry.

    ./main IMAGE_PATH ORB_CPU

  • GPU_VO: Run CUDA-accelerated Visual Odometry.

    ./main IMAGE_PATH ORB_GPU

  • benchmark: Run benchmarking to evaluate the performance of the GPU-accelerated implementation.

    ./benchmark

Result

Computation Time Comparison

Method Compute Time for 1600 Images (ms)
ORB_CPU 20134 ms
ORB_CUDA 4474 ms

TODO:

Add documentation in code

Add tests