你用SORT和deep SORT两种算法分别实现多目标行人的跟踪，并写出示例代码

好的，下面分别介绍使用 SORT 和 Deep SORT 两种算法实现多目标行人跟踪的示例代码。

首先介绍使用 SORT 算法实现多目标行人跟踪的示例代码：

import numpy as np
from filterpy.kalman import KalmanFilter
from scipy.optimize import linear_sum_assignment

class SORT:
    def __init__(self):
        self.dt = 1.0
        self.lamb = 100
        self.sigma_l = 10
        self.sigma_h = 1.0
        self.sigma_iou = 0.3
        self.min_hits = 3
        self.max_age = 20
        self.trackers_list = []
        self.frame_count = 0

    def __dist(self, bbox1, bbox2):
        x1, y1, w1, h1 = bbox1
        x2, y2, w2, h2 = bbox2
        c1 = np.array([x1 + w1 / 2.0, y1 + h1 / 2.0])
        c2 = np.array([x2 + w2 / 2.0, y2 + h2 / 2.0])
        return np.sqrt(np.sum(np.square(c1 - c2)))

    def __linear_assignment(self, cost_matrix):
        row_ind, col_ind = linear_sum_assignment(cost_matrix)
        return row_ind, col_ind

    def __get_cost_matrix(self, detections, trackers):
        cost_matrix = np.zeros((len(detections), len(trackers)), dtype=np.float32)
        for i, detection in enumerate(detections):
            for j, tracker in enumerate(trackers):
                cost_matrix[i, j] = self.__dist(detection, tracker.predict())
        return cost_matrix

    def __update_tracker(self, tracker, detection):
        tracker.kf.update(detection)
        tracker.hits += 1
        tracker.no_losses = 0

    def __create_tracker(self, detection):
        tracker = KalmanFilter(dim_x=7, dim_z=4)
        tracker.x[:4] = detection
        tracker.x[4:] = 0.0
        tracker.P[4:, 4:] *= 1000.0
        tracker.P *= 10.0
        tracker.R = np.diag([self.sigma_l ** 2, self.sigma_l ** 2,
                             self.sigma_h ** 2, self.sigma_h ** 2])
        tracker.Q[:4, :4] *= 0.01
        tracker.Q[4:, 4:] *= 0.01
        tracker.Q[4:, 4:] *= self.dt
        tracker.H = np.zeros((4, 7))
        tracker.H[:4, :4] = np.eye(4)
        tracker.hits = 1
        tracker.no_losses = 0
        return tracker

    def __remove_trackers(self):
        self.trackers_list = [tracker for tracker in self.trackers_list if tracker.no_losses <= self.max_age]

    def __remove_duplicate_trackers(self):
        for tracker1 in self.trackers_list:
            for tracker2 in self.trackers_list:
                if tracker1 == tracker2:
                    continue
                if self.__iou(tracker1.predict(), tracker2.predict()) > self.sigma_iou:
                    if tracker1.hits >= tracker2.hits:
                        tracker2.no_losses = self.max_age + 1
                    else:
                        tracker1.no_losses = self.max_age + 1

    def __ious(self, bboxes1, bboxes2):
        ious = np.zeros((len(bboxes1), len(bboxes2)), dtype=np.float32)
        for i, bbox1 in enumerate(bboxes1):
            for j, bbox2 in enumerate(bboxes2):
                ious[i, j] = self.__iou(bbox1, bbox2)
        return ious

    def __iou(self, bbox1, bbox2):
        x1, y1, w1, h1 = bbox1
        x2, y2, w2, h2 = bbox2
        xi1 = max(x1, x2)
        yi1 = max(y1, y2)
        xi2 = min(x1 + w1, x2 + w2)
        yi2 = min(y1 + h1, y2 + h2)
        inter_area = max(xi2 - xi1, 0) * max(yi2 - yi1, 0)
        union_area = w1 * h1 + w2 * h2 - inter_area
        return inter_area / union_area

    def update(self, detections):
        self.frame_count += 1

        for tracker in self.trackers_list:
            self.__update_tracker(tracker, np.array([0, 0, 0, 0]))

        cost_matrix = self.__get_cost_matrix(detections, self.trackers_list)

        row_ind, col_ind = self.__linear_assignment(cost_matrix)

        matched = set()
        for i, j in zip(row_ind, col_ind):
            if cost_matrix[i, j] > self.lamb:
                continue
            tracker = self.trackers_list[j]
            detection = detections[i]
            self.__update_tracker(tracker, detection)
            matched.add(j)

        unmatched_trackers = set(range(len(self.trackers_list))) - matched
        for j in unmatched_trackers:
            tracker = self.trackers_list[j]
            tracker.no_losses += 1

        unmatched_detections = set(range(len(detections))) - set(row_ind)
        for i in unmatched_detections:
            detection = detections[i]
            tracker = self.__create_tracker(detection)
            self.trackers_list.append(tracker)

        self.__remove_trackers()
        self.__remove_duplicate_trackers()

        return self.trackers_list

使用 SORT 算法跟踪多目标行人的示例代码：

import cv2
import numpy as np
from sort import SORT

def main():
    cap = cv2.VideoCapture("test.mp4")
    tracker = SORT()
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        detections = [[100, 100, 50, 100], [200, 200, 100, 200]]
        trackers = tracker.update(detections)
        for tracker in trackers:
            bbox = tracker.predict()
            x, y, w, h = bbox
            cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.imshow("frame", frame)
        if cv2.waitKey(1) == ord('q'):
            break
    cap.release()
    cv2.destroyAllWindows()

if __name__ == '__main__':
    main()

接着介绍使用 Deep SORT 算法实现多目标行人跟踪的示例代码：

import numpy as np
from filterpy.kalman import KalmanFilter
from scipy.optimize import linear_sum_assignment

class DeepSORT:
    def __init__(self):
        self.dt = 1.0
        self.lamb = 100
        self.sigma_l = 10
        self.sigma_h = 1.0
        self.sigma_iou = 0.3
        self.min_hits = 3
        self.max_age = 20
        self.trackers_list = []
        self.frame_count = 0

    def __dist(self, bbox1, bbox2):
        x1, y1, w1, h1 = bbox1
        x2, y2, w2, h2 = bbox2
        c1 = np.array([x1 + w1 / 2.0, y1 + h1 / 2.0])
        c2 = np.array([x2 + w2 / 2.0, y2 + h2 / 2.0])
        return np.sqrt(np.sum(np.square(c1 - c2)))

    def __linear_assignment(self, cost_matrix):
        row_ind, col_ind = linear_sum_assignment(cost_matrix)
        return row_ind, col_ind

    def __get_cost_matrix(self, detections, trackers):
        cost_matrix = np.zeros((len(detections), len(trackers)), dtype=np.float32)
        for i, detection in enumerate(detections):
            for j, tracker in enumerate(trackers):
                cost_matrix[i, j] = self.__dist(detection, tracker.predict())
        return cost_matrix

    def __update_tracker(self, tracker, detection):
        tracker.kf.update(detection)
        tracker.hits += 1
        tracker.no_losses = 0

    def __create_tracker(self, detection):
        tracker = KalmanFilter(dim_x=7, dim_z=4)
        tracker.x[:4] = detection
        tracker.x[4:] = 0.0
        tracker.P[4:, 4:] *= 1000.0
        tracker.P *= 10.0
        tracker.R = np.diag([self.sigma_l ** 2, self.sigma_l ** 2,
                             self.sigma_h ** 2, self.sigma_h ** 2])
        tracker.Q[:4, :4] *= 0.01
        tracker.Q[4:, 4:] *= 0.01
        tracker.Q[4:, 4:] *= self.dt
        tracker.H = np.zeros((4, 7))
        tracker.H[:4, :4] = np.eye(4)
        tracker.hits = 1
        tracker.no_losses = 0
        return tracker

    def __remove_trackers(self):
        self.trackers_list = [tracker for tracker in self.trackers_list if tracker.no_losses <= self.max_age]

    def __remove_duplicate_trackers(self):
        for tracker1 in self.trackers_list:
            for tracker2 in self.trackers_list:
                if tracker1 == tracker2:
                    continue
                if self.__iou(tracker1.predict(), tracker2.predict()) > self.sigma_iou:
                    if tracker1.hits >= tracker2.hits:
                        tracker2.no_losses = self.max_age + 1
                    else:
                        tracker1.no_losses = self.max_age + 1

    def __ious(self, bboxes1, bboxes2):
        ious = np.zeros((len(bboxes1), len(bboxes2)), dtype=np.float32)
        for i, bbox1 in enumerate(bboxes1):
            for j, bbox2 in enumerate(bboxes2):
                ious[i, j] = self.__iou(bbox1, bbox2)
        return ious

    def __iou(self, bbox1, bbox2):
        x1, y1, w1, h1 = bbox1
        x2, y2, w2, h2 = bbox2
        xi1 = max(x1, x2)
        yi1 = max(y1, y2)
        xi2 = min(x1 + w1, x2 + w2)
        yi2 = min(y1 + h1, y2 + h2)
        inter_area = max(xi2 - xi1, 0) * max(yi2 - yi1, 0)
        union_area = w1 * h1 + w2 * h2 - inter_area
        return inter_area / union_area

    def update(self, detections, features):
        self.frame_count += 1

        for tracker in self.trackers_list:
            self.__update_tracker(tracker, np.array([0, 0, 0, 0]))

        cost_matrix = self.__get_cost_matrix(detections, self.trackers_list)

        row_ind, col_ind = self.__linear_assignment(cost_matrix)

        matched = set()
        for i, j in zip(row_ind, col_ind):
            if cost_matrix[i, j] > self.lamb:
                continue
            tracker = self.trackers_list[j]
            detection = detections[i]
            feature = features[i]
            tracker.features.append(feature)
            self.__update_tracker(tracker, detection)
            matched.add(j)

        unmatched_trackers = set(range(len(self.trackers_list))) - matched
        for j in unmatched_trackers:
            tracker = self.trackers_list[j]
            tracker.no_losses += 1

        unmatched_detections = set(range(len(detections))) - set(row_ind)
        for i in unmatched_detections:
            detection = detections[i]
            feature = features[i]
            tracker = self.__create_tracker(detection)
            tracker.features.append(feature)
            self.trackers_list.append(tracker)

        self.__remove_trackers()
        self.__remove_duplicate_trackers()

        return self.trackers_list

使用 Deep SORT 算法跟踪多目标行人的示例代码：

import cv2
import numpy as np
from deep_sort import DeepSORT

def main():
    cap = cv2.VideoCapture("test.mp4")
    deepsort = DeepSORT()
    while True:
        ret, frame = cap.read()
        if not ret:
            break
        detections = [[100, 100, 50, 100], [200, 200, 100, 200]]
        features = np.random.rand(len(detections), 128)
        trackers = deepsort.update(detections, features)
        for tracker in trackers:
            bbox = tracker.predict()
            x, y, w, h = bbox
            cv2.rectangle(frame, (x, y), (x + w, y + h), (0, 255, 0), 2)
        cv2.imshow("frame", frame)
        if cv2.waitKey(1) == ord('q'):
            break
    cap.release()
    cv2.destroyAllWindows()

if __name__ == '__main__':
    main()

以上就是使用 SORT 和 Deep SORT 两种算法实现多目标行人跟踪的示例代码。