Nvidia jetson-nano yolov5 russian license plate recognition

Nvidia jetson nano + RTSP IP camera + yolov5 + russian license plate recognition

The project if for building on HOST machine (not on target jetson-nano board)

This demo project is distibuted as is without any support

requires:

• Linux machine with any distribution as a build host
• CUDA Toolkit 10.2 for x86 installed
• CUDA == 10.2, TensorRT >= 8.0, OpenCV+CUDA >= 4.4.x libraries prebuilt for ARM64 + headers
• live555 library prebuilt for ARM64 + headers
• Linux for Tegra rootfs deployed on the build host
• quemu-user-static for deploying Linux for Tegra rootfs
• GCC compiler for ARM64 (aarch64-linux-gnu-7.5.0 has been used)

prepare:

• download and unpack Linux for Tegra roots
• with root privilegies run ./apply_binaries.sh from the Linux_for_Tegra directory
• install CUDA Toolkit
• download or install GCC compiler
• prepare libraries form Nvidia (TensorRT, OpenCV+CUDA, CUDA) for ARM64
• change paths in Makefile and Rules.mk

build

run make

TensorRT engines (yolov5m_lpr_pre_v2.engine, yolov5s_lpr_plate.engine) for jetson-nano are in engines directory

run on jetson-nano board

run from the RTSP ip camera: ./lpr -u rtsp://user:psssword@192.168.100.64:554 -w 1920 -h 1080 -f H264 run from the H264 video file: (result output to the H264 file lpr.h264 ) ./lpr -i ./city_drive.h264 -o ./lpr.h264 -w 1280 -h 720 -f H264 run from the folder with images: (result images to another folder img_out) ./lpr -F img_in -T img_out