apollo::perception::trafficlight::TrafficLightDetection类 参考

#include <detection.h>

类 apollo::perception::trafficlight::TrafficLightDetection 继承关系图:

apollo::perception::trafficlight::TrafficLightDetection 的协作图:

Public 成员函数
	TrafficLightDetection ()
	Construct a new traffic light detection object.
	~TrafficLightDetection ()=default
	Destroy the traffic light detection object.
bool	Init (const TrafficLightDetectorInitOptions &options) override
	Initialize traffic light detector parameters.
bool	Detect (camera::TrafficLightFrame *frame) override
	Detect traffic light from image.
bool	SelectOutputBoxes (const std::vector< base::RectI > &crop_box_list, const std::vector< float > &resize_scale_list_col, const std::vector< float > &resize_scale_list_row, std::vector< base::TrafficLightPtr > *lights)
	Dump output of inference results.
void	ApplyNMS (std::vector< base::TrafficLightPtr > *lights, double iou_thresh=0.6)
	Filter overlapping boxes using NMS.
bool	Inference (std::vector< base::TrafficLightPtr > *lights, camera::DataProvider *data_provider)
	Model inference process.
const std::vector< base::TrafficLightPtr > &	getDetectedBoxes ()
	Get the detected boxes object.
Public 成员函数 继承自 apollo::perception::trafficlight::BaseTrafficLightDetector
	BaseTrafficLightDetector ()=default
	Construct a new base traffic light detector object.
virtual	~BaseTrafficLightDetector ()=default
	Destroy the base traffic light detector object.
	DISALLOW_COPY_AND_ASSIGN (BaseTrafficLightDetector)

详细描述

在文件 detection.h 第 37 行定义.

构造及析构函数说明

TrafficLightDetection()

apollo::perception::trafficlight::TrafficLightDetection::TrafficLightDetection

(

)

Construct a new traffic light detection object.

在文件 detection.cc 第 35 行定义.

                                             :
        max_batch_size_(4), param_blob_length_(6), mean_{0, 0, 0}, gpu_id_(0) {}

~TrafficLightDetection()

apollo::perception::trafficlight::TrafficLightDetection::~TrafficLightDetection ( )

default

Destroy the traffic light detection object.

成员函数说明

ApplyNMS()

void apollo::perception::trafficlight::TrafficLightDetection::ApplyNMS	(	std::vector< base::TrafficLightPtr > *	lights,
		double	iou_thresh = 0.6
	)

Filter overlapping boxes using NMS.

参数

lights
iou_thresh

在文件 detection.cc 第 346 行定义.

                                                                                              {
    if (lights == nullptr) {
        AERROR << "lights are not available";
        return;
    }
 
    // (score, index) pairs sorted by detect score
    std::vector<std::pair<float, int>> score_index_vec(lights->size());
    for (size_t i = 0; i < lights->size(); ++i) {
        score_index_vec[i].first = lights->at(i)->region.detect_score;
        score_index_vec[i].second = static_cast<int>(i);
    }
    std::stable_sort(
            score_index_vec.begin(),
            score_index_vec.end(),
            [](const std::pair<float, int> &pr1, const std::pair<float, int> &pr2) { return pr1.first < pr2.first; });
 
    std::vector<int> kept_indices;
    while (!score_index_vec.empty()) {
        const int idx = score_index_vec.back().second;
        bool keep = true;
        for (size_t k = 0; k < kept_indices.size(); ++k) {
            const int kept_idx = kept_indices[k];
            const auto &rect1 = lights->at(idx)->region.detection_roi;
            const auto &rect2 = lights->at(kept_idx)->region.detection_roi;
            float overlap = static_cast<float>((rect1 & rect2).Area() / (rect1 | rect2).Area());
            // if current bbox has large overlap(>=iou_thresh) with any
            // kept bbox, drop it
            keep = std::fabs(overlap) < iou_thresh;
            if (!keep) {
                break;
            }
        }
        if (keep) {
            kept_indices.push_back(idx);
        }
        score_index_vec.pop_back();
    }
 
    int idx = 0;
    auto parted_itr = std::stable_partition(lights->begin(), lights->end(), [&](const base::TrafficLightPtr &light) {
        return std::find(kept_indices.begin(), kept_indices.end(), idx++) != kept_indices.end();
    });
    lights->erase(parted_itr, lights->end());
}

Detect()

bool apollo::perception::trafficlight::TrafficLightDetection::Detect ( camera::TrafficLightFrame *  frame )

overridevirtual

Detect traffic light from image.

参数

frame

返回

true

false

实现了 apollo::perception::trafficlight::BaseTrafficLightDetector.

在文件 detection.cc 第 207 行定义.

                                                                 {
    if (frame->traffic_lights.empty()) {
        AINFO << "no lights to detect";
        return true;
    }
 
    const auto &data_provider = frame->data_provider.get();
    auto input_blob = rt_net_->get_blob(net_inputs_[0]);
    int img_width = data_provider->src_width();
    int img_height = data_provider->src_height();
    std::vector<base::TrafficLightPtr> &lights_ref = frame->traffic_lights;
 
    AINFO << "detection input " << lights_ref.size() << " lights";
 
    selected_bboxes_.clear();
    detected_bboxes_.clear();
 
    for (auto &light : lights_ref) {
        base::RectI debug_rect(0, 0, 0, 0);
        light->region.detection_roi = light->region.projection_roi;
        light->region.debug_roi.clear();
        light->region.debug_roi_detect_scores.clear();
        light->region.debug_roi.push_back(debug_rect);
        light->region.debug_roi_detect_scores.push_back(0.0f);
    }
 
    for (auto &light : lights_ref) {
        if (light->region.outside_image || camera::OutOfValidRegion(light->region.projection_roi, img_width, img_height)
            || light->region.projection_roi.Area() <= 0) {
            light->region.projection_roi.x = 0;
            light->region.projection_roi.y = 0;
            light->region.projection_roi.width = 0;
            light->region.projection_roi.height = 0;
        }
    }
 
    PERF_BLOCK("traffic_light_detector_inference")
    Inference(&lights_ref, data_provider);
    PERF_BLOCK_END
 
    AINFO << "Dump output Done! Get box num:" << detected_bboxes_.size();
 
    for (size_t j = 0; j < detected_bboxes_.size(); ++j) {
        base::RectI &region = detected_bboxes_[j]->region.detection_roi;
        float score = detected_bboxes_[j]->region.detect_score;
        lights_ref[0]->region.debug_roi.push_back(region);
        lights_ref[0]->region.debug_roi_detect_scores.push_back(score);
    }
 
    AINFO << "start select";
    select_.SelectTrafficLights(detected_bboxes_, &lights_ref);
    AINFO << "select success";
 
    AINFO << "detection success";
    return true;
}

getDetectedBoxes()

const std::vector< base::TrafficLightPtr > & apollo::perception::trafficlight::TrafficLightDetection::getDetectedBoxes ( )

inline

Get the detected boxes object.

返回

const std::vector<base::TrafficLightPtr>&

在文件 detection.h 第 102 行定义.

                                                             {
        return detected_bboxes_;
    }

Inference()

bool apollo::perception::trafficlight::TrafficLightDetection::Inference	(	std::vector< base::TrafficLightPtr > *	lights,
		camera::DataProvider *	data_provider
	)

Model inference process.

参数

lights
data_provider

返回

true

false

在文件 detection.cc 第 134 行定义.

                                                                                                               {
    if (cudaSetDevice(gpu_id_) != cudaSuccess) {
        AERROR << "Failed to set device to " << gpu_id_;
        return false;
    }
    auto batch_num = lights->size();
    for (size_t i = 0; i < batch_num; ++i) {
        crop_box_list_.clear();
        resize_scale_list_.clear();
        int img_width = data_provider->src_width();
        int img_height = data_provider->src_height();
        int resize_index = 0;
        auto input_img_blob = rt_net_->get_blob(net_inputs_[0]);
        auto input_param = rt_net_->get_blob(net_inputs_[1]);
 
        input_img_blob->Reshape(
                1,
                static_cast<int>(detection_param_.min_crop_size()),
                static_cast<int>(detection_param_.min_crop_size()),
                3);
        param_blob_->Reshape(1, 6, 1, 1);
        float *param_data = param_blob_->mutable_cpu_data();
        param_data[0] = static_cast<float>(detection_param_.min_crop_size());
        param_data[1] = static_cast<float>(detection_param_.min_crop_size());
        param_data[2] = 1;
        param_data[3] = 1;
        param_data[4] = 0;
        param_data[5] = 0;
 
        AINFO << "reshape inputblob " << input_img_blob->shape_string();
 
        base::TrafficLightPtr light = lights->at(i);
        base::RectI cbox;
        crop_->getCropBox(img_width, img_height, light, &cbox);
        AINFO << "get crop box success " << cbox.x << " " << cbox.y << " " << cbox.width << " " << cbox.height;
 
        if (!camera::OutOfValidRegion(cbox, img_width, img_height) && cbox.Area() > 0) {
            crop_box_list_.push_back(cbox);
            light->region.debug_roi[0] = cbox;
            light->region.crop_roi = cbox;
 
            data_provider_image_option_.do_crop = true;
            data_provider_image_option_.crop_roi = cbox;
            data_provider_image_option_.target_color = base::Color::BGR;
            data_provider->GetImage(data_provider_image_option_, image_.get());
            AINFO << "get image data success ";
 
            float resize_scale = static_cast<float>(detection_param_.min_crop_size())
                    / static_cast<float>(std::min(cbox.width, cbox.height));
            resize_scale_list_.push_back(resize_scale);
 
            inference::ResizeGPU(
                    *image_, input_img_blob, img_width, resize_index, mean_[0], mean_[1], mean_[2], true, 1.0);
            resize_index++;
        }
 
        // inference
        PERF_BLOCK("traffic_light_detection_inference")
        cudaDeviceSynchronize();
        rt_net_->Infer();
        cudaDeviceSynchronize();
        PERF_BLOCK_END
        AINFO << "rt_net run success";
 
        // dump the output
        SelectOutputBoxes(crop_box_list_, resize_scale_list_, resize_scale_list_, &detected_bboxes_);
    }
 
    ApplyNMS(&detected_bboxes_);
 
    return true;
}

Init()

bool apollo::perception::trafficlight::TrafficLightDetection::Init ( const TrafficLightDetectorInitOptions &  options )

overridevirtual

Initialize traffic light detector parameters.

参数

options

返回

true

false

实现了 apollo::perception::trafficlight::BaseTrafficLightDetector.

在文件 detection.cc 第 38 行定义.

                                                                                           {
    std::string config_file = GetConfigFile(options.config_path, options.config_file);
 
    if (!cyber::common::GetProtoFromFile(config_file, &detection_param_)) {
        AERROR << "Failed to load traffic light detection parameters: " << config_file;
        return false;
    }
 
    AINFO << "Traffic light detection param: " << detection_param_.DebugString();
 
    const auto &model_info = detection_param_.info();
    std::string model_path = GetModelPath(model_info.name());
    AINFO << "Traffic light detection model_path: " << model_path;
 
    std::string proto_file = GetModelFile(model_path, model_info.proto_file().file());
    AINFO << "Traffic light detection proto_file: " << proto_file;
 
    std::string weight_file = GetModelFile(model_path, model_info.weight_file().file());
    AINFO << "Traffic light detection weight_file: " << weight_file;
 
    if (detection_param_.is_bgr()) {
        data_provider_image_option_.target_color = base::Color::BGR;
        mean_[0] = detection_param_.mean_b();
        mean_[1] = detection_param_.mean_g();
        mean_[2] = detection_param_.mean_r();
    } else {
        data_provider_image_option_.target_color = base::Color::RGB;
        mean_[0] = detection_param_.mean_r();
        mean_[1] = detection_param_.mean_g();
        mean_[2] = detection_param_.mean_b();
    }
 
    // Network input and output names
    net_inputs_ = inference::GetBlobNames(model_info.inputs());
    net_outputs_ = inference::GetBlobNames(model_info.outputs());
 
    const auto &model_type = model_info.framework();
    AINFO << "model_type: " << model_type;
 
    rt_net_.reset(inference::CreateInferenceByName(
            model_type, proto_file, weight_file, net_outputs_, net_inputs_, model_path));
 
    AINFO << "rt_net_ create succeed";
    rt_net_->set_gpu_id(options.gpu_id);
    AINFO << "set gpu id " << options.gpu_id;
    gpu_id_ = options.gpu_id;
 
    int resize_height = detection_param_.min_crop_size();
    int resize_width = detection_param_.min_crop_size();
    max_batch_size_ = detection_param_.max_batch_size();
    param_blob_length_ = 6;
 
    CHECK_GT(resize_height, 0);
    CHECK_GT(resize_width, 0);
    CHECK_GT(max_batch_size_, 0);
 
    std::map<std::string, std::vector<int>> input_reshape;
    inference::AddShape(&input_reshape, model_info.inputs());
 
    if (!rt_net_->Init(input_reshape)) {
        AINFO << "net init fail.";
        return false;
    }
    AINFO << "net init success.";
 
    mean_buffer_.reset(new base::Blob<float>(1, resize_height, resize_height, 3));
 
    param_blob_ = rt_net_->get_blob(net_inputs_[1]);
    float *param_data = param_blob_->mutable_cpu_data();
    for (int i = 0; i < max_batch_size_; ++i) {
        auto offset = i * param_blob_length_;
        param_data[offset + 0] = static_cast<float>(resize_width);
        param_data[offset + 1] = static_cast<float>(resize_height);
        param_data[offset + 2] = 1;
        param_data[offset + 3] = 1;
        param_data[offset + 4] = 0;
        param_data[offset + 5] = 0;
    }
 
    switch (detection_param_.crop_method()) {
    default:
    case 0:
        crop_.reset(new CropBox(detection_param_.crop_scale(), detection_param_.min_crop_size()));
        break;
    case 1:
        crop_.reset(new CropBoxWholeImage());
        break;
    }
 
    select_.Init(resize_width, resize_height);
    image_.reset(new base::Image8U(resize_height, resize_width, base::Color::BGR));
    return true;
}

SelectOutputBoxes()

bool apollo::perception::trafficlight::TrafficLightDetection::SelectOutputBoxes	(	const std::vector< base::RectI > &	crop_box_list,
		const std::vector< float > &	resize_scale_list_col,
		const std::vector< float > &	resize_scale_list_row,
		std::vector< base::TrafficLightPtr > *	lights
	)

Dump output of inference results.

参数

crop_box_list
resize_scale_list_col
resize_scale_list_row
lights

返回

true

false

在文件 detection.cc 第 264 行定义.

                                                {
    auto output_blob = rt_net_->get_blob(net_outputs_[0]);
    auto model_type = detection_param_.info().framework();
    int result_box_num, each_box_length;
    if (model_type == common::TensorRT) {
        result_box_num = output_blob->shape(1);
        each_box_length = output_blob->shape(2);
    } else {
        result_box_num = output_blob->shape(0);
        each_box_length = output_blob->shape(1);
    }
 
    AINFO << "output blob size " << output_blob->shape(0) << " " << output_blob->shape(1) << " "
          << output_blob->shape(2) << " " << output_blob->shape(3);
    AINFO << "result box number: " << result_box_num << " each box length: " << each_box_length;
 
    for (int candidate_id = 0; candidate_id < result_box_num; candidate_id++) {
        const float *result_data = output_blob->cpu_data() + candidate_id * each_box_length;
        int img_id = static_cast<int>(result_data[0]);
        if (img_id < 0) {
            continue;
        } else if (img_id >= static_cast<int>(crop_box_list.size())) {
            AINFO << "img id " << img_id << " > " << crop_box_list.size();
            continue;
        }
        base::TrafficLightPtr tmp(new base::TrafficLight);
        float inflate_col = 1 / resize_scale_list_col.at(img_id);
        float inflate_row = 1 / resize_scale_list_row.at(img_id);
        float x1 = result_data[1];
        float y1 = result_data[2];
        float x2 = result_data[3];
        float y2 = result_data[4];
        std::vector<float> score{result_data[5], result_data[6], result_data[7], result_data[8]};
        for (int i = 0; i < 9; ++i) {
            ADEBUG << "result_data " << result_data[i];
        }
 
        std::vector<float>::iterator biggest = std::max_element(std::begin(score), std::end(score));
        tmp->region.detect_class_id = base::TLDetectionClass(std::distance(std::begin(score), biggest) - 1);
 
        if (static_cast<int>(tmp->region.detect_class_id) >= 0) {
            tmp->region.detection_roi.x = static_cast<int>(x1 * inflate_col);
            tmp->region.detection_roi.y = static_cast<int>(y1 * inflate_row);
            tmp->region.detection_roi.width = static_cast<int>((x2 - x1 + 1) * inflate_col);
            tmp->region.detection_roi.height = static_cast<int>((y2 - y1 + 1) * inflate_row);
            tmp->region.detect_score = *biggest;
 
            if (camera::OutOfValidRegion(
                        tmp->region.detection_roi, crop_box_list.at(img_id).width, crop_box_list.at(img_id).height)
                || tmp->region.detection_roi.Area() <= 0) {
                AINFO << "Invalid width or height or x or y: " << tmp->region.detection_roi.width << " | "
                      << tmp->region.detection_roi.height << " | " << tmp->region.detection_roi.x << " | "
                      << tmp->region.detection_roi.y;
                AINFO << " max width " << crop_box_list.at(img_id).width << " max height "
                      << crop_box_list.at(img_id).height << " at img_id " << img_id;
                continue;
            }
 
            camera::RefineBox(
                    tmp->region.detection_roi,
                    crop_box_list.at(img_id).width,
                    crop_box_list.at(img_id).height,
                    &(tmp->region.detection_roi));
            tmp->region.detection_roi.x += crop_box_list.at(img_id).x;
            tmp->region.detection_roi.y += crop_box_list.at(img_id).y;
            tmp->region.is_detected = true;
            AINFO << "detect roi x " << tmp->region.detection_roi.x << " " << tmp->region.detection_roi.y << " "
                  << tmp->region.detection_roi.width << " " << tmp->region.detection_roi.height;
 
            lights->push_back(tmp);
        } else {
            AWARN << "Invalid class id: " << static_cast<int>(tmp->region.detect_class_id);
        }
    }
 
    return true;
}

---

该类的文档由以下文件生成:

• modules/perception/traffic_light_detection/detector/caffe_detection/detection.h
• modules/perception/traffic_light_detection/detector/caffe_detection/detection.cc