基于mmpose的动作分类

基于mmpose的动作分类

说明：本项目使用了检测+关键点+跟踪+分类

检测模型

本项目使用yolov7的模型进行人体的检测

克隆项目

git clone https://github.com/WongKinYiu/yolov7.git

下载模型：

https://github.com/WongKinYiu/yolov7/releases/download/v0.1/yolov7.pt

在文件路径输入cmd进入终端然后创建环境，安装依赖

conda create -n 环境名 python=X.X 

pip install -r requirements.txt -i https://pypi.tuna.tsinghua.edu.cn/simple
该命令可以下载快一点，使用的国内源

测试环境，如果有检测结果，说明环境安装完成

python detect.py --weights weights/yolov7.pt --source inference/images

封装

创建新的项目PoseAndSort

将v7的utils和models文件夹复制到新的项目,然后新建一个pose.py，对检测模型进行封装

class MyDetect:
    def __init__(self, weights, device_name, img_size=640):
        # print('detectapi __init__')
        self.opt = simulation_opt(weights=weights, img_size=img_size)
        weights, imgsz = self.opt.weights, self.opt.img_size

        # Initialize
        set_logging()
        if device_name != "cuda":
            self.device = select_device("0")
        else:
            self.device = select_device(device_name)
        self.half = self.device.type != 'cpu'  # half precision only supported on CUDA

        # model = torch.load(weights,
        #                    map_location=self.device)  # load FP32 model
        # self.model = model['model']
        self.model = attempt_load(weights, map_location=self.device)  # load FP32 model
        self.stride = int(self.model.stride.max())  # model stride
        self.imgsz = check_img_size(imgsz, s=self.stride)  # check img_size

        if self.half:
            self.model.half()  # to FP16

        self.classify = False

        # read names and colors
        self.names = self.model.module.names if hasattr(self.model, 'module') else self.model.names
        self.colors = [[random.randint(0, 255) for _ in range(3)] for _ in self.names]

        def run(self, source):  # 使用时，调用这个函数
            检测代码省略......
            返回nms后的scale_coords检测结果

mmpose-关键点

说明：因为特殊要求，需要脚部的关键点，所以没有使用v7的关键点模型

模型使用的是 人体 2d 关键点 (17 Keypoints)，RTMPose-l

安装

创建一个 conda 虚拟环境并激活它。

conda create --name openmmlab python=3.8 -y
conda activate openmmlab

使用 MIM 安装 MMEngine 和 MMCV

pip install -U openmim
mim install mmengine
mim install "mmcv>=2.0.1"

模型推理

# RTMPose
python tools/deploy.py \
    configs/mmpose/pose-detection_simcc_onnxruntime_dynamic.py \
    {RTMPOSE_PROJECT}/rtmpose/body_2d_keypoint/rtmpose-m_8xb256-420e_coco-256x192.py \
    ../rtmpose_m/rtmpose_m.pth \
    demo/resources/human-pose.jpg \
    --work-dir mmdeploy_models/mmpose/sdk \
    --device cpu \
    --show \
    --dump-info  # 导出 sdk info

导出onnx

# 前往 mmdeploy 目录
cd ${PATH_TO_MMDEPLOY}
# 转换 RTMPose
python tools/deploy.py \
    configs/mmpose/pose-detection_simcc_onnxruntime_dynamic.py \
    {RTMPOSE_PROJECT}/rtmpose/body_2d_keypoint/rtmpose-m_8xb256-420e_coco-256x192.py \
    ../rtmpose_m/rtmpose_m.pth \
    demo/resources/human-pose.jpg \
    --work-dir mmdeploy_models/mmpose/ort \
    --device cpu \
    --show

open-mmlab 提供了在线的模型转换和通用的onnxm模型

硬件模型库 (openmmlab.com)

封装

class MyRtmpose():
    def __init__(self, pose_model_path, device_name):
        self.model = PoseDetector(
            model_path=pose_model_path, device_name=device_name)

    def run(self, images, bboxes):
        result = self.model(images, bboxes)
        return result

说明需要安装mmdeploy_runtime

跟踪模型

本项目使用了传统的sort和Deepsort

sort

class Sort(object):
    def __init__(self, max_age=1, min_hits=3, iou_threshold=0.3):
        """
        Parameters for SORT
        """
        self.max_age = max_age
        self.min_hits = min_hits
        self.iou_threshold = iou_threshold
        self.trackers = []
        self.frame_count = 0
        self.color_list = []

    def getTrackers(self, ):
        return self.trackers

    def update(self, dets=np.empty((0, 6)), unique_color=False):
        """
        Parameters:
        'dets' - a numpy array of detection in the format [[x1, y1, x2, y2, score], [x1,y1,x2,y2,score],...]

        Ensure to call this method even frame has no detections. (pass np.empty((0,5)))

        Returns a similar array, where the last column is object ID (replacing confidence score)

        NOTE: The number of objects returned may differ from the number of objects provided.
        """
        self.frame_count += 1

        # Get predicted locations from existing trackers
        trks = np.zeros((len(self.trackers), 6))
        to_del = []
        ret = []
        for t, trk in enumerate(trks):

            pos = self.trackers[t].predict()[0]
            trk[:] = [pos[0], pos[1], pos[2], pos[3], 0, 0]
            if np.any(np.isnan(pos)):
                to_del.append(t)
        trks = np.ma.compress_rows(np.ma.masked_invalid(trks))
        for t in reversed(to_del):
            self.trackers.pop(t)
            if unique_color:
                self.color_list.pop(t)
        matched, unmatched_dets, unmatched_trks = associate_detections_to_trackers(dets, trks, self.iou_threshold)

        # Update matched trackers with assigned detections
        for m in matched:
            self.trackers[m[1]].update(dets[m[0], :])

        # Create and initialize new trackers for unmatched detections
        for i in unmatched_dets:
            trk = KalmanBoxTracker(np.hstack((dets[i, :], np.array([0]))))
            self.trackers.append(trk)
            if unique_color:
                self.color_list.append(get_color())

        i = len(self.trackers)
        for trk in reversed(self.trackers):
            d = trk.get_state()[0]
            if (trk.time_since_update < 1) and (trk.hit_streak >= self.min_hits or self.frame_count <= self.min_hits):
                ret.append(np.concatenate((d, [trk.id + 1])).reshape(1,
                                                                     -1))  # +1'd because MOT benchmark requires positive value
            i -= 1
            # remove dead tracklet
            if (trk.time_since_update > self.max_age):
                self.trackers.pop(i)
                if unique_color:
                    self.color_list.pop(i)

        if (len(ret) > 0):
            return np.concatenate(ret)
        return np.empty((0, 6))

deepsort

class DeepSort(object):
    def __init__(self, model_path, max_dist=0.2, min_confidence=0.3, nms_max_overlap=1.0, max_iou_distance=0.7, max_age=70, n_init=3, nn_budget=100, use_cuda=True):
        self.min_confidence = min_confidence
        self.nms_max_overlap = nms_max_overlap

        self.extractor = Extractor(model_path, use_cuda=use_cuda)

        max_cosine_distance = max_dist
        nn_budget = 100
        metric = NearestNeighborDistanceMetric("cosine", max_cosine_distance, nn_budget)
        self.tracker = Tracker(metric, max_iou_distance=max_iou_distance, max_age=max_age, n_init=n_init)

    def update(self, bbox_xywh, confidences, ori_img):
        self.height, self.width = ori_img.shape[:2]
        # generate detections
        features = self._get_features(bbox_xywh, ori_img)
        bbox_tlwh = self._xywh_to_tlwh(bbox_xywh)
        detections = [Detection(bbox_tlwh[i], conf, features[i]) for i,conf in enumerate(confidences) if conf>self.min_confidence]

        # run on non-maximum supression
        boxes = np.array([d.tlwh for d in detections])
        scores = np.array([d.confidence for d in detections])
        indices = non_max_suppression(boxes, self.nms_max_overlap, scores)
        detections = [detections[i] for i in indices]

        # update tracker
        self.tracker.predict()
        self.tracker.update(detections)

        # output bbox identities
        outputs = []
        for track in self.tracker.tracks:
            if not track.is_confirmed() or track.time_since_update > 1:
                continue
            box = track.to_tlwh()
            x1,y1,x2,y2 = self._tlwh_to_xyxy(box)
            track_id = track.track_id
            outputs.append(np.array([x1,y1,x2,y2,track_id], dtype=np.int))
        if len(outputs) > 0:
            outputs = np.stack(outputs,axis=0)
        return outputs


    """
    TODO:
        Convert bbox from xc_yc_w_h to xtl_ytl_w_h
    Thanks JieChen91@github.com for reporting this bug!
    """
    @staticmethod
    def _xywh_to_tlwh(bbox_xywh):
        if isinstance(bbox_xywh, np.ndarray):
            bbox_tlwh = bbox_xywh.copy()
        elif isinstance(bbox_xywh, torch.Tensor):
            bbox_tlwh = bbox_xywh.clone()
        bbox_tlwh[:,0] = bbox_xywh[:,0] - bbox_xywh[:,2]/2.
        bbox_tlwh[:,1] = bbox_xywh[:,1] - bbox_xywh[:,3]/2.
        return bbox_tlwh


    def _xywh_to_xyxy(self, bbox_xywh):
        x,y,w,h = bbox_xywh
        x1 = max(int(x-w/2),0)
        x2 = min(int(x+w/2),self.width-1)
        y1 = max(int(y-h/2),0)
        y2 = min(int(y+h/2),self.height-1)
        return x1,y1,x2,y2

    def _tlwh_to_xyxy(self, bbox_tlwh):
        """
        TODO:
            Convert bbox from xtl_ytl_w_h to xc_yc_w_h
        Thanks JieChen91@github.com for reporting this bug!
        """
        x,y,w,h = bbox_tlwh
        x1 = max(int(x),0)
        x2 = min(int(x+w),self.width-1)
        y1 = max(int(y),0)
        y2 = min(int(y+h),self.height-1)
        return x1,y1,x2,y2

    def _xyxy_to_tlwh(self, bbox_xyxy):
        x1,y1,x2,y2 = bbox_xyxy

        t = x1
        l = y1
        w = int(x2-x1)
        h = int(y2-y1)
        return t,l,w,h

    def _get_features(self, bbox_xywh, ori_img):
        im_crops = []
        for box in bbox_xywh:
            x1,y1,x2,y2 = self._xywh_to_xyxy(box)
            im = ori_img[y1:y2,x1:x2]
            im_crops.append(im)
        if im_crops:
            features = self.extractor(im_crops)
        else:
            features = np.array([])
        return features

分类

因数据原因使用机器学习模型进行动作的分类

根据获取的26个关键点数据进行归一化

随机森林模型

X_train, X_test, y_train, y_test = train_test_split(all_embedding, all_target, test_size=0.25, shuffle=True,
                                                    stratify=all_target)
rfc = RandomForestClassifier(random_state=0, n_estimators=59, max_depth=13,min_samples_leaf=1,min_samples_split=24)

rfc = rfc.fit(X_train, y_train)
score_r = rfc.score(X_test, y_test)

pre = rfc.predict(X_test)  # 预测 类别
print("测试集准确率", accuracy_score(y_test, pre))
print("测试集召回率", recall_score(y_test, pre, average="weighted"))
print("测试集精确率", precision_score(y_test, pre, average="weighted"))
print("测试集f1", f1_score(y_test, pre, average="weighted"))

决策树

# 决策树分类器
clf = DecisionTreeClassifier(random_state=0)
clf = clf.fit(X_train, y_train)
score_c = clf.score(X_test, y_test)
pre = clf.predict(X_test)  # 预测 类别
print("测试集准确率", accuracy_score(y_test, pre))
print("测试集召回率", recall_score(y_test, pre, average="weighted"))
print("测试集精确率", precision_score(y_test, pre, average="weighted"))
print("测试集f1", f1_score(y_test, pre, average="weighted"))

SVC

# 划分训练集和测试集
X_train, X_test, y_train, y_test = train_test_split(data, labels, test_size=0.2, random_state=0)

# 训练模型
model = SVC()
model.fit(X_train, y_train)

# 测试模型
accuracy = model.score(X_test, y_test)
print("准确率:", accuracy)

KNN分类器

X_train, X_test, y_train, y_test = train_test_split(all_embedding, all_target, test_size=0.25, shuffle=True,
                                                    stratify=all_target)
# 创建KNN分类器
knn = KNeighborsClassifier(n_neighbors=18)

# 训练分类器
knn.fit(X_train, y_train)

# 对测试集进行预测
y_pred = knn.predict(X_test)

# 计算准确率
accuracy = accuracy_score(y_test, y_pred)
print("准确率:", accuracy)

使用神经网络

# 定义神经网络模型
class ComplexNet(nn.Module):
    def __init__(self):
        super(ComplexNet, self).__init__()
        self.fc1 = nn.Linear(78, 936)
        self.fc2 = nn.Linear(936, 780)
        self.fc3 = nn.Linear(780, 624)
        self.fc4 = nn.Linear(624, 468)
        # self.fc5 = nn.Linear(468, 312)
        self.fc5 = nn.Linear(468, 5)
        self.relu = nn.Sigmoid()
        # self.relu = nn.ReLU()
        self.dropout = nn.Dropout(0.2)

    #
    def forward(self, x):
        x = self.fc1(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc2(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc3(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc4(x)
        x = self.relu(x)
        x = self.dropout(x)
        x = self.fc5(x)
        # x = self.relu(x)
        # x = self.dropout(x)
        # x = self.fc6(x)
        return x