YOLOv8_TensorRT_CUDA_DeepSort

YOLOv8 is used for object detection, while CUDA is utilized for image preprocessing and postprocessing after the detection inference. All model inference is performed using TensorRT and implemented in a C++ environment.

Watch the video here

Docker

Build Docker container

Dockerfile.yolo includes

• ultralytics
• pytorch
• tensorflow
• onnxrt
• tensorrt
• opencv
• eigen

To build the Docker containers, use the following commands:

sudo docker build -t opencv:opencv --no-cache --rm --file Dockerfile.opencv .
sudo docker build -t detection:yolo --no-cache --rm --file Dockerfile.yolo .

Run Docker Container

To run the Docker container, navigate to the project directory and use the following command:

cd YOLOv8_CUDA_DeepSort
sudo docker run -it --rm --gpus all -p 5000:5000 \
--volume="$PWD:/workspace" \
--volume="/home/tingwei/Project/dataset:/workspace/dataset" \
detection:yolo

Models

Generate TensorRT engine model

Generate Yolov8 ONNX model

First, generate the YOLOv8 ONNX model by running the following commands: Change model name in export.py for other Yolov8 model.

cd yolov8
python export.py

Training extractor model and generate ONNX model

Follow the steps in the README.md of 'reid' folder to generate and train the extractor ONNX model.

Convert ONNX model to TensortRT engine.

Follow the steps in the README.md of the convert_onnx_to_engine_cpp folder to convert the ONNX model to a TensorRT engine.

Object Tracking

Usage

1. Build the project

   Create a build directory and compile the code using CMake and Make:

cd object_tracking
mkdir build
cd build
cmake ..
make

2. Run the Object Tracking Module

Use the following command to run the object tracking module, replacing [detector_engine_path] with the path to your detector engine model, [extractor_engine_path] with the path to your extractor engine model, [image_folder_path] with the path to your testing images, and [tracker_result_output_path] with the location for the tracking result text file.

./object_tracking [detector_engine_path] [extractor_engine_path] [image_folder_path] [tracker_result_output_path]

3. Visulize the result

   Generate video for tracking result.

usage: visualize.py [--tracker_result_path]
                    [--output_folder_path]

Speed

Host: intel i7
Device: GTX1080ti
Image Size: 1920x1080