risc-nyuad/drone_detection

Drone Detection Based on Yolo and optimized with TKDNN and TensorRT

Drone_Detection

Drone Detection Based on Yolo and optimized with TKDNN and TensorRT

This Drone Detector is based on Darknet(YOLO) and under process. This detector works as a part in a long-term drone tracking system. It works both on visual domain and thermal domain

Demo

Thermal Domain

Visual Domain

Speed comparison

Experiments on RTX 2060

Resolution	Config	Speed,fps	TkDNN TensorRT speedup(fp16),fps
416*416	yolov4	82	162
512*512	yolov4	69	134
608*608	yolov4	53	103
416*416	yolov4tiny	300	790
512*512	yolov4tiny	140	-
608*608	yolov4tiny	89	-
416*416(cpu)	yolov4tiny	4	42

The library is compiled as .so file which can be used in c++(./src/main) or python environment(./python) The code is been optimized.

Instruction

Now It support CPU or GPU mode with sperate file "libdarknet.so" and "libdarknetcpu.so" To compile the demo:

1. Download the weights file from googledrivelink and put them under weights folder

2. Configuration

# set the path for configuration, weights, videos in your main.cpp file

string cfgfile = "../../cfg/yolov4-tiny-3l-drone.cfg";
string weightfile = "../../weights/yolov4-tiny-3l-drone.weights";
VideoCapture capture("../../demo/cut_drone.mp4");
ifstream classNamesFile("../../cfg/drone.names");

3. Compile

cd src
mkdir build && cd build
cmake ..
make

4. The excuate file is under build folder ./YoloDroneDetection

Test with CPU + opencv or OPENCL + opencv

The speed on cpu with opencv can achieve 35 fps!!!

1. Install OpenCV3.4.10 (or Higher) with CUDA support(if CUDA needed) Under opencv folder, change the configuration in main.cpp line 89-96, default is using cpu for inference.

//net.setPreferableBackend(cv::dnn::DNN_BACKEND_CUDA);  //If use cuda for backend
//net.setPreferableTarget(cv::dnn::DNN_TARGET_CUDA);    //If use cuda for optimization

net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);  //If use cpu for backend
net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);       //If use cpu for optimization

//net.setPreferableBackend(cv::dnn::DNN_BACKEND_OPENCV);  //If use cpu for detection
//net.setPreferableTarget(cv::dnn::DNN_TARGET_OPENCL);   //If use opencl for optimization
//net.setPreferableTarget(cv::dnn::DNN_TARGET_OPENCL_FP16); // or use opencl fp16 for  optimization

//net.setPreferableBackend(cv::dnn::DNN_BACKEND_INFERENCE_ENGINE);
//net.setPreferableTarget(cv::dnn::DNN_TARGET_CPU);

2. compile main.cpp

g++ -o main  main.cpp  `pkg-config opencv4 --cflags --libs`

3. run

./main ../cfg/yolov4-tiny-3l-drone.cfg ../weights/yolov4-tiny-3l-drone.weights ../demo/cut_drone.mp4 ../cfg/drone.names

TODO

• C++ multithread and memory safe
• Intergrate tracking under C++ framework
• Provide TKDNN optimized model and test script