Ascend Deep-SORT

MOT using deepsort yolov5 with ascend

Requirement	Version
OS	Ubuntu 18.04
OpenCV	4.x compiled with GCC 7.5
OpenMP	default with GCC
CANN	3.3.0 or above

supported Ascend Hardware：Atlas300-3000/3010、Atlas800-3000/3010

Performance：

Hardware	e2e latency（10 tracks）	FPS
Ascend310	50ms	20

1.onnx model trans to ascend model

1.1 modify the focus modular to an Conv operator by scripts/modify_yolov5.py

yolov5m model can be download here:

链接：https://pan.baidu.com/s/1eZ9GBjEFgSIx-ayjzYzCBQ 提取码：1234

python3 modify_yolov5.py yolov5m.onnx yolov5m

1.2 trans yolov5.onnx to om with ATC

source /usr/local/Ascend/ascend-toolkit/set_env.sh
atc --model=yolov5m.onnx --framework=5 --output=yolov5m --soc_version=Ascend310 --insert_op_conf=aipp_rgb_yolo.cfg --input_format=NCHW --input_shape="images:1,3,640,640" --out_nodes="Conv_266:0;Conv_282:0;Conv_298:0" --output_type="Conv_266:0:FP32;Conv_282:0:FP32;Conv_298:0:FP32"

param	statement
--model	input onnx model
--output	output om file
--input_shape	model input shape: (batch, channel, width, height)
--out_nodes	select model output nodes, node's name in onnx graph
--output_type	define model output nodes' type, the node must corresponding to the items which out_nodes defined

which nodes you should select:

1.3 trans feature extractor model to om with ATC

feature extractor model can be download here

链接：https://pan.baidu.com/s/1nEfb3Yd0n32fHj9C9QdqQw 提取码：1234

source /usr/local/Ascend/ascend-toolkit/set_env.sh
atc --model=feature.onnx --framework=5 --output=feature --soc_version=Ascend310 --insert_op_conf=aipp_rgb_feature.cfg --input_format=NCHW --input_shape="input:1,3,128,64"

2.Compile and Run program

2.1 modify CMakeLists.txt

set(OpenCV_DIR /root/opencv4.4/lib/cmake/opencv4) #your own opencv path
...
set(ACL_PATH /usr/local/Ascend/ascend-toolkit/latest) #your own cann-toolkit path

2.2 Compile it

mkdir build
cd build
cmake ..
make -j8

2.3 Run it

./ACL_DEEPSORT model/yolov5m.om model/feature.om data/test.mp4

===============================
yolov5 infer time:20.759ms
Detections size:9
feature infer time:13.906ms
predict time:0.693ms
update time:11.268ms
fps:21.3538
===============================

Reference:

DeepSORT: https://github.com/shaoshengsong/DeepSORT

Yolov5_DeepSort_Pytorch: https://github.com/mikel-brostrom/Yolov5_DeepSort_Pytorch