tracking
June 7, 2025 · View on GitHub
基于FairMot模型的目标追踪
下载模型
- track/fairmot.onnx: FairMot, Model Type: onnx, input size: Nx640x640x3, classes: 80, download
获取测试视频
wget https://bj.bcebos.com/paddlehub/fastdeploy/person.mp4
运行demo
注：请将PATH更换为自己对应的目录
- --name: 模型名称
- --inference_type: 推理后端类型
- --device_type: 推理后端的执行设备类型
- --model_type: 模型类型
- --is_path: 模型是否为路径
- --model_value: 模型路径或模型文件
- --codec_flag: 编解码类型
- --parallel_type: 并行类型
- --input_path: 输入视频路径
- --output_path: 输出视频路径
- --model_inputs: 模型输入
- --model_outputs: 模型输出
推理后端为onnxruntime，推理执行设备为CUDA
# 进入目录
cd /yourpath/nndeploy/build
# 链接
export LD_LIBRARY_PATH=$(pwd):$LD_LIBRARY_PATH
export LD_LIBRARY_PATH=/home/resource/third_party/onnxruntime-linux-aarch64-1.20.1/lib:$LD_LIBRARY_PATH
# 串行执行
./nndeploy_demo_track --name nndeploy::track::fairmot --inference_type kInferenceTypeOnnxRuntime --device_type kDeviceTypeCodeCuda:0 --model_type kModelTypeOnnx --is_path --codec_flag kCodecFlagVideo --parallel_type kParallelTypeSequential --input_path ./person.avi --output_path output.avi --model_value /home/for_all_users/model/track/fairmot/fairmot.onnx --model_inputs im_shape,image,scale_factor --model_outputs fetch_name_0,fetch_name_1
# 耗时
TimeProfiler: demo
------------------------------------------------------------------------------------------
name call_times sum cost_time(ms) avg cost_time(ms) gflops
------------------------------------------------------------------------------------------
graph->init() 1 265.533 265.533 0.000
graph->run 1 13108.209 13108.209 0.000
demo run() 200 13088.963 65.445 0.000
decode_node run() 200 749.108 3.746 0.000
nndeploy::track::fairmot run() 200 10419.823 52.099 0.000
preprocess run() 200 170.853 0.854 0.000
infer run() 200 8823.659 44.118 0.000
postprocess run() 200 1424.164 7.121 0.000
vismot_node run() 200 128.972 0.645 0.000
encode_node run() 200 1789.954 8.950 0.000
------------------------------------------------------------------------------------------
# 流水线执行
./nndeploy_demo_track --name nndeploy::track::fairmot --inference_type kInferenceTypeOnnxRuntime --device_type kDeviceTypeCodeCuda:0 --model_type kModelTypeOnnx --is_path --codec_flag kCodecFlagVideo --parallel_type kParallelTypeSequential --input_path ./person.avi --output_path output.avi --model_value /home/for_all_users/model/track/fairmot/fairmot.onnx --model_inputs im_shape,image,scale_factor --model_outputs fetch_name_0,fetch_name_1
# 耗时
TimeProfiler: demo
------------------------------------------------------------------------------------------
name call_times sum cost_time(ms) avg cost_time(ms) gflops
------------------------------------------------------------------------------------------
graph->init() 1 230.590 230.590 0.000
graph->run 1 9052.425 9052.425 0.000
decode_node run() 544 8781.201 16.142 0.000
demo run() 200 0.017 0.000 0.000
preprocess run() 200 216.381 1.082 0.000
nndeploy::track::fairmot run() 200 0.489 0.002 0.000
infer run() 200 9031.738 45.159 0.000
postprocess run() 200 1512.302 7.562 0.000
vismot_node run() 200 178.980 0.895 0.000
encode_node run() 200 1966.076 9.830 0.000
------------------------------------------------------------------------------------------
效果示例
输入视频
输出视频
