AI has seen you clearly, YOLO+ByteTrack+multi-label classification network

Today I will share with you a pedestrian attribute analysis system. Pedestrians can be identified from video or camera video streams and each person's attributes can be marked.

The recognized attributes include the following 10 categories

Some categories have multiple attributes. If the body orientation is: front , side and back, so there are 26 attributes in the final training.

Implementing such a system requires 3 steps:

• Use YOlOv5 to identify pedestrians
• Use ByteTrack to track and mark the same person
• Train multi-label images Classification network, identifying 26 attributes of pedestrians

1. Pedestrian recognition and tracking

Pedestrian recognition uses the YOLOv5 target detection model. You can train the model yourself, or you can directly use YOLOv5 pre-training Good model.

Pedestrian tracking uses Multi-Object Tracking Technology (MOT) technology. The video is composed of pictures. Although we humans can identify the same person in different pictures, if we do not track pedestrians, AI is unrecognizable. MOT technology is needed to track the same person and assign a unique ID to each pedestrian.

The training and use of the YOLOv5 model, as well as the principles and implementation plans of the multi-object tracking technology (MOT) technology, are detailed in the previous article. Interested friends can check out the article there. YOLOv5 ByteTrack counts traffic flow.

2. Training multi-label classification network

Most of the image classifications we first came into contact with were single-label classification, that is: a picture is classified into category 1, and the category can be two categories. It can also be multiple categories. Assuming there are three categories, the label corresponding to each picture may be in the following general format:

001.jpg010
002.jpg100
003.jpg100

label​Only one position is 1.

The multi-label classification network we are going to train today is a picture that contains multiple categories at the same time. The label format is as follows:

001.jpg011
002.jpg111
003.jpg100

label​ can have multiple positions of 1.

There are two options for training such a network. One is to treat each category as a single-label classification, calculate the loss separately, summarize the total, and calculate the gradient to update the network parameters.

The other one can be trained directly, but you need to pay attention to the network details, taking ResNet50 as an example

resnet50 = ResNet50(include_top=False, weights='imagenet')

# 迁移学习，不重新训练卷积层
for layer in resnet50.layers:
layer.trainable = False

# 新的全连接层
x = Flatten()(resnet50.output)
x = Dense(1024)(x)
x = Activation('relu')(x)
x = BatchNormalization()(x)
x = Dropout(0.5)(x)

# 输出 26 个属性的多分类标签
x = Dense(26, activatinotallow='sigmoid')(x)

model = Model(inputs = resnet50.input, outputs=x)

The activation function of the final output layer must be sigmoid, because each attribute needs to be calculated separately Probability. In the same way, the loss function during training also needs to use binary_crossentropy.

In fact, the principles of the above two methods are similar, but the development workload is different.

For convenience here, I use PaddleCls for training. Paddle's configuration is simple, but its disadvantage is that it is a bit of a black box. You can only follow its own rules, and it is more troublesome to customize it.

The model training uses the PA100K data set. It should be noted that although the original label defined by the PA100K data set has the same meaning as Paddle, the order is different.

For example: The 1st digit of the original label represents whether the label is female, while Paddle requires the 1st digit to represent whether the label is wearing a hat, and the 22nd digit represents whether the label is female.

#We can adjust the original label position according to Paddle’s requirements, so that it will be easier for us to reason later.

Download PaddleClas

git clone https://github.com/PaddlePaddle/PaddleClas

Unzip the downloaded dataset and place it in the dataset directory of PaddleClas.

Find the ppcls/configs/PULC/person_attribute/PPLCNet_x1_0.yaml configuration file and configure the image and label paths.

DataLoader:
Train:
dataset:
name: MultiLabelDataset
image_root: "dataset/pa100k/" #指定训练AI has seen you clearly, YOLO+ByteTrack+multi-label classification network所在根路径
cls_label_path: "dataset/pa100k/train_list.txt" #指定训练列表文件位置
label_ratio: True
transform_ops:

Eval:
dataset:
name: MultiLabelDataset
image_root: "dataset/pa100k/" #指定评估AI has seen you clearly, YOLO+ByteTrack+multi-label classification network所在根路径
cls_label_path: "dataset/pa100k/val_list.txt" #指定评估列表文件位置
label_ratio: True
transform_ops:

train_list.txt format is

00001.jpg0,0,1,0,....

After configuration, you can train directly

python3 tools/train.py
-c ./ppcls/configs/PULC/person_attribute/PPLCNet_x1_0.yaml

After training, export the model

python3 tools/export_model.py
-c ./ppcls/configs/PULC/person_attribute/PPLCNet_x1_0.yaml
-o Global.pretrained_model=output/PPLCNet_x1_0/best_model
-o Global.save_inference_dir=deploy/models/PPLCNet_x1_0_person_attribute_infer

will The exported results are placed in the ~/.paddleclas/inference_model/PULC/person_attribute/ directory

You can use the function provided by PaddleCls to directly call

import paddleclas

model = paddleclas.PaddleClas(model_name="person_attribute")

result = model.predict(input_data="./test_imgs/000001.jpg")

print(result)

for output The results are as follows:

[{'attributes': ['Female', 'Age18-60', 'Front', 'Glasses: False', 'Hat: False', 'HoldObjectsInFront: True', 'ShoulderBag', 'Upper: ShortSleeve', 'Lower:Trousers', 'No boots'], 'output': [0, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0, 1, 0, 1, 0, 1, 1, 0, 0], 'filename': './test_imgs/000001.jpg'}]

The model training process ends here, the data set and the source code of the entire project have been packaged.

The above is the detailed content of AI has seen you clearly, YOLO+ByteTrack+multi-label classification network. For more information, please follow other related articles on the PHP Chinese website!