janthmueller / fairmot_tensorrt Goto Github PK

This is C++ version of FairMOT using TensorRT in Windows.

The detection and feature extraction parts using TensorRT can refer to tensorRTIntegrate. If you would like to remove the dependency of DCN, please refer to another repository FairMOT_TensorRT_C.

1. export onnx model file in FairMOT by adding the following code in line 479 in "/src/lib/models/networks/pose_dla_dcn.py"

	z = {}
	for head in self.heads:
		z[head] = self.__getattr__(head)(y[-1])

	hm = z["hm"]
	wh = z["wh"]
	reg = z["reg"]
	hm = F.sigmoid(hm)
	hm_pool = F.max_pool2d(hm, kernel_size=3, stride=1, padding=1)


	id_feature = z['id']
	id_feature = F.normalize(id_feature, dim=1)
	id_feature = id_feature.permute(0, 2, 3, 1).contiguous() #switch id dim
	return [hm, wh, reg, hm_pool, id_feature]

2. Follow tensorRTIntegrate to build TensorRT engine which has been covered in this repository. You need to follow tensorRTIntegrate first to make sure DCN works and then add the MOT part in the folder "mot".

• Kalman Filter is borrowed from DeepSort, [deep_sort] https://github.com/apennisi/deep_sort
• [FairMOT]（https://github.com/ifzhang/FairMOT）
• [tensorRTIntegrate] (https://github.com/dlunion/tensorRTIntegrate)