Features

• Recognition and tracking of 2-digits target for automatic aeroplane for the Chinese Aeromodelling Design Competition (CADC)
• Detection model trained using Darknet-YOLO
• 2 digits recognition, automatic parameter searching using NNI , model built using Pytorch, trained on SVHN (The Street View House Numbers)
• Acceleration : TensorRT (tkDNN) (FP16), C++, multi-threading
• Real-time (20-25FPS) on TX2 (edge computing platform), ~ 200FPS on RTX 2070

Dependencies

• tkDNN
• darknet
• libtorch
• TensorRT
• OpenCV
• CUDA
• cuDNN

Build

• Build darknet by allowing OpenCV, GPU and cuDNN
• Add the following code to public method of class DetectionNN in tkDNN-master/include/tkDNN/DetectionNN.h and build tkDNN

void setthreshold(float threshold) confThreshold=threshold;

• Instead of using the libtorch downloaded, rebuild Torch by setting to allow CXX11_ABI in TorchConfig.cmake

if ("${CMAKE_CXX_COMPILER_ID}" STREQUAL "GNU")
  set(TORCH_CXX_FLAGS "-D_GLIBCXX_USE_CXX11_ABI=1")
endif()