4.17.8.1. deepsort_tracker_controller

针对DeepSORT算法，通过处理检测的结果和提取的特征，实现对目标的跟踪。

4.17.8.1.1. 构造函数

接口形式:

deepsort_tracker_controller(float max_cosine_distance,
                        int nn_budget,
                        int k_feature_dim,
                        float max_iou_distance = 0.7,
                        int max_age = 30,
                        int n_init = 3);

参数说明:

max_cosine_distance: float

输入参数。用于相似度计算的最大余弦距离阈值。

nn_budget: int

输入参数。用于最近邻搜索的最大数量限制。

k_feature_dim: int

输入参数。被检测的目标的特征维度。

max_iou_distance: float

输入参数。模用于跟踪器中的最大交并比（IoU）距离阈值。

max_age: int

输入参数。跟踪目标在跟踪器中存在的最大帧数。

n_init: int

输入参数。跟踪器中的初始化帧数阈值。

4.17.8.1.2. process

处理接口。

接口形式1:

int process(const vector<DeteObjRect>& detected_objects,
        vector<Tensor>& feature,
        vector<TrackObjRect>& tracked_objects);

参数说明1:

detected_objects: vector<DeteObjRect>

输入参数。检测出的物体框。

feature: vector<Tensor>

输入参数。检测出的物体的特征。

tracked_objects: vector<TrackObjRect>

输出参数。被跟踪的物体。

返回值说明1:

int

成功返回0，失败返回其他。

接口形式2:

int process(const vector<DeteObjRect>& detected_objects,
        vector<vector<float>>& feature,
        vector<TrackObjRect>& tracked_objects);

参数说明2:

detected_objects: vector<DeteObjRect>

输入参数。检测出的物体框。

feature: vector<float>

输入参数。检测出的物体的特征。

tracked_objects: vector<TrackObjRect>

输出参数。被跟踪的物体。

返回值说明2:

int

成功返回0，失败返回其他。

示例代码:

// The example code relies on sophon-demo/sample/YOLOv5/cpp/yolov5_bmcv/yolov5.h and sophon-demo/sample/DeepSORT/cpp/deepsort_bmcv/FeatureExtractor.h
#include <sail/cvwrapper.h>
#include "yolov5.h"
#include "FeatureExtractor.h"
#include <opencv2/opencv.hpp>
#include <vector>
#include <string>

using namespace std;

class YOLOv5Arg {
public:
    string bmodel;
    int dev_id;
    float conf_thresh;
    float nms_thresh;

    YOLOv5Arg(string bmodel, int dev_id, float conf_thresh, float nms_thresh) {
        this->bmodel = bmodel;
        this->dev_id = dev_id;
        this->conf_thresh = conf_thresh;
        this->nms_thresh = nms_thresh;
    }
};

int main() {
    string input = "data/test_car_person_1080P.mp4";
    string bmodel_detector = "models/BM1684X/yolov5s_v6.1_3output_int8_1b.bmodel";
    string bmodel_extractor = "models/BM1684X/extractor_int8_1b.bmodel";
    int dev_id = 0;
    float conf = 0.4;
    float nms = 0.7;

    YOLOv5Arg yolov5_args(bmodel_detector, dev_id, conf, nms);
    YOLOv5 yolov5(yolov5_args);
    Extractor extractor(bmodel_extractor, dev_id);

    cv::VideoCapture cap(input);
    vector<cv::Mat> img_batch;

    sail::deepsort_tracker_controller dstc(0.2, 100, extractor.output_shape[1], 0.7, 70, 3);

    vector<vector<float>> track_res_all_numpy;

    for (int i = 0; i < 15; i++) {
        cv::Mat img;
        cap.read(img);
        if (img.empty()) {
            break;
        }
        img_batch.push_back(img);
        vector<vector<float>> det = yolov5(img_batch);
        vector<cv::Rect> dets;
        for (auto& item : det) {
            int x1 = static_cast<int>(item[0]);
            int y1 = static_cast<int>(item[1]);
            int x2 = static_cast<int>(item[2]);
            int y2 = static_cast<int>(item[3]);
            cv::Rect roi(x1, y1, x2 - x1, y2 - y1);
            dets.push_back(roi);
        }
        vector<cv::Mat> im_crops;
        for (auto& roi : dets) {
            cv::Mat img_crop = img(roi);
            im_crops.push_back(img_crop);
        }
        vector<vector<float>> ext_results = extractor(im_crops);

        // The order of this API and the demo is inconsistent, and the class_id and score are reversed
        for (auto& row : det) {
            swap(row[4], row[5]);
        }
        img_batch.clear();

        vector<tuple<int, int, int, int, int, float, int>> det_tuple;
        for (auto& row : det) {
            det_tuple.push_back(make_tuple(static_cast<int>(row[0]), static_cast<int>(row[1]), static_cast<int>(row[2]), static_cast<int>(row[3]), static_cast<int>(row[4]), row[5], static_cast<int>(row[6])));
        }

    }
    return 0;
}