From f9a48f07b7708707d0cb79277e9f83874fe1ae52 Mon Sep 17 00:00:00 2001 From: GuanYuankai Date: Mon, 27 Oct 2025 07:23:35 +0000 Subject: [PATCH] =?UTF-8?q?=E9=99=8D=E4=BD=8E=E8=A7=86=E9=A2=91=E9=87=87?= =?UTF-8?q?=E9=9B=86=E7=9A=84CPU=E5=8D=A0=E6=9C=89=E7=8E=87?= MIME-Version: 1.0 Content-Type: text/plain; charset=UTF-8 Content-Transfer-Encoding: 8bit --- src/rknn/video_service.cc | 157 ++++++++++++++++++++++++++------------ 1 file changed, 110 insertions(+), 47 deletions(-) diff --git a/src/rknn/video_service.cc b/src/rknn/video_service.cc index 53056d0..baf9a2a 100644 --- a/src/rknn/video_service.cc +++ b/src/rknn/video_service.cc @@ -53,13 +53,24 @@ bool VideoService::start() { } printf("rknnPool init success.\n"); - setenv("OPENCV_FFMPEG_CAPTURE_OPTIONS", "rtsp_transport;tcp", 1); - printf("Set RTSP transport protocol to TCP\n"); + // setenv("OPENCV_FFMPEG_CAPTURE_OPTIONS", "rtsp_transport;tcp", 1); + // printf("Set RTSP transport protocol to TCP\n"); + std::string gst_input_pipeline = + "rtspsrc location=" + input_url_ + " latency=0 protocols=tcp ! " + "rtph265depay ! " + "h265parse ! " + "mppvideodec format=16 ! " + "video/x-raw,format=BGR ! " // <-- 关键：直接请求 mppvideodec 输出 BGR 格式 + "appsink"; - capture_.open(input_url_, cv::CAP_FFMPEG); + spdlog::info("Try to Open RTSP Stream"); + capture_.open(gst_input_pipeline, cv::CAP_GSTREAMER); + if (!capture_.isOpened()) { printf("Error: Could not open RTSP stream: %s\n", input_url_.c_str()); return false; + }else{ + spdlog::info("RTSP Stream Opened!"); } frame_width_ = static_cast(capture_.get(cv::CAP_PROP_FRAME_WIDTH)); @@ -70,13 +81,13 @@ bool VideoService::start() { printf("RTSP stream opened successfully! (%dx%d @ %.2f FPS)\n", frame_width_, frame_height_, frame_fps_); - std::string gst_pipeline = + std::string gst_pipeline = "appsrc ! " - "queue max-size-buffers=2 leaky=downstream ! " - "video/x-raw,format=BGR ! " - "videoconvert ! " - "video/x-raw,format=NV12 ! " - "mpph265enc gop=25 rc-mode=fixqp qp-init=26 ! " + "queue max-size-buffers=2 leaky=downstream ! " + "video/x-raw,format=BGR ! " // OpenCV VideoWriter 输入 BGR 数据 + "videoconvert ! " // <-- 使用 CPU 将 BGR 转换为 NV12 + "video/x-raw,format=NV12 ! " // 明确指定 videoconvert 输出 NV12 + "mpph265enc gop=25 rc-mode=fixqp qp-init=26 ! " // 硬件编码器接收 NV12 数据 "h265parse ! " "rtspclientsink location=" + output_rtsp_url_ + " latency=0 protocols=tcp"; @@ -131,18 +142,48 @@ void VideoService::stop() { void VideoService::update_tracker(detect_result_group_t &detect_result_group, const cv::Size& frame_size) { - if (intrusion_zone_.width == 0 || intrusion_zone_.height == 0) { + // 如果入侵区域无效，则设置为帧中心的 1/4 区域 (基于原始帧大小) + if (intrusion_zone_.width <= 0 || intrusion_zone_.height <= 0) { intrusion_zone_ = cv::Rect(frame_size.width / 4, frame_size.height / 4, frame_size.width / 2, frame_size.height / 2); } - std::vector current_detections; + // --- 缩放比例计算 --- + // !!! 重要: 请确保这里的 640.0f 与您 OpenCV resize 时的目标尺寸一致 !!! + const float model_input_width = 640.0f; + const float model_input_height = 640.0f; + float scale_x = (float)frame_size.width / model_input_width; + float scale_y = (float)frame_size.height / model_input_height; + // --- 结束缩放比例计算 --- + + std::vector current_detections; // 存储当前帧检测到的、已缩放到原始尺寸的框 for (int i = 0; i < detect_result_group.count; i++) { detect_result_t *det_result = &(detect_result_group.results[i]); + // 只处理 "person" 类别 if (strcmp(det_result->name, "person") == 0) { - current_detections.push_back(cv::Rect( - det_result->box.left, det_result->box.top, - det_result->box.right - det_result->box.left, - det_result->box.bottom - det_result->box.top)); + // --- 将模型输出坐标按比例缩放回原始帧坐标 --- + int original_left = static_cast(det_result->box.left * scale_x); + int original_top = static_cast(det_result->box.top * scale_y); + int original_right = static_cast(det_result->box.right * scale_x); + int original_bottom = static_cast(det_result->box.bottom * scale_y); + + // --- 边界检查与修正 --- + // 确保坐标不会超出原始图像边界 + original_left = std::max(0, std::min(original_left, frame_size.width - 1)); + original_top = std::max(0, std::min(original_top, frame_size.height - 1)); + // 确保 right >= left, bottom >= top + original_right = std::max(original_left, std::min(original_right, frame_size.width)); + original_bottom = std::max(original_top, std::min(original_bottom, frame_size.height)); + // --- 结束边界检查 --- + + // 只有当框有效时（宽度和高度大于0）才添加到检测列表 + if (original_right > original_left && original_bottom > original_top) { + current_detections.push_back(cv::Rect( + original_left, original_top, + original_right - original_left, // width + original_bottom - original_top // height + )); + } + // --- 结束坐标缩放 --- } } @@ -150,65 +191,86 @@ void VideoService::update_tracker(detect_result_group_t &detect_result_group, co it->second.frames_unseen++; } - for (const auto& det_box : current_detections) { - bool is_matched = false; - int best_match_id = -1; - double max_iou = 0.3; // IoU阈值 + std::vector matched_track_ids; // 记录本帧已匹配到的跟踪ID，防止一个跟踪ID匹配多个检测框 - for (auto const& [id, person] : this->tracked_persons_) { + // 尝试将当前检测框与已跟踪目标进行匹配 + for (const auto& det_box : current_detections) { + int best_match_id = -1; + double max_iou_threshold = 0.3; // IoU 匹配阈值 + double best_iou = 0.0; // 用于寻找最佳匹配 + + for (auto const& [id, person] : tracked_persons_) { + // 检查该跟踪ID是否已在本帧匹配过 + bool already_matched = false; + for(int matched_id : matched_track_ids) { + if (id == matched_id) { + already_matched = true; + break; + } + } + if (already_matched) { + continue; // 跳过已匹配的跟踪目标 + } + + // 计算 IoU (现在 det_box 和 person.box 都是原始坐标系) double iou = (double)(det_box & person.box).area() / (double)(det_box | person.box).area(); - if (iou > max_iou) { - max_iou = iou; + if (iou > best_iou && iou >= max_iou_threshold) { // 必须大于等于阈值才能成为候选 + best_iou = iou; best_match_id = id; } } - + if (best_match_id != -1) { + // 找到匹配，更新跟踪信息 tracked_persons_[best_match_id].box = det_box; tracked_persons_[best_match_id].frames_unseen = 0; - is_matched = true; + matched_track_ids.push_back(best_match_id); // 记录已匹配 } else { + // 没有找到匹配，创建新的跟踪目标 TrackedPerson new_person; - new_person.id = this->next_track_id_++; // 使用成员变量 + new_person.id = next_track_id_++; // 分配新ID new_person.box = det_box; new_person.entry_time = 0; new_person.is_in_zone = false; new_person.alarm_triggered = false; new_person.frames_unseen = 0; - tracked_persons_[new_person.id] = new_person; // 使用成员变量 + tracked_persons_[new_person.id] = new_person; } } + // 更新每个跟踪目标的区域状态和报警状态 double current_time = get_current_time_seconds(); for (auto it = tracked_persons_.begin(); it != tracked_persons_.end(); ++it) { TrackedPerson& person = it->second; - // 使用成员变量 - bool currently_in_zone = (this->intrusion_zone_ & person.box).area() > 0; + // 检查人与入侵区域的交集 (现在都是原始坐标系) + bool currently_in_zone = (intrusion_zone_ & person.box).area() > 0; if (currently_in_zone) { - if (!person.is_in_zone) { + if (!person.is_in_zone) { // 刚进入区域 person.is_in_zone = true; person.entry_time = current_time; - } else { - // 使用成员变量 - if (!person.alarm_triggered && (current_time - person.entry_time) > this->intrusion_time_threshold_) { + person.alarm_triggered = false; // 重置报警状态 + } else { // 持续在区域内 + // 检查是否达到报警时间阈值且尚未报警 + if (!person.alarm_triggered && (current_time - person.entry_time) >= intrusion_time_threshold_) { person.alarm_triggered = true; - trigger_alarm(person.id, person.box); // 调用成员函数 + trigger_alarm(person.id, person.box); // 触发报警 } } - } else { - person.is_in_zone = false; - person.entry_time = 0; - person.alarm_triggered = false; + } else { // 当前不在区域内 + if (person.is_in_zone) { // 刚离开区域 + person.is_in_zone = false; + person.entry_time = 0; // 重置进入时间 + person.alarm_triggered = false; // 重置报警状态 + } } } - for (auto it = tracked_persons_.begin(); it != tracked_persons_.end(); ) { - // (建议) 增加到50帧 (约2秒) 提高鲁棒性，减少ID切换 - if (it->second.frames_unseen > 50) { - it = tracked_persons_.erase(it); + for (auto it = tracked_persons_.begin(); it != tracked_persons_.end(); /* 无自增 */) { + if (it->second.frames_unseen > 50) { // 超过 50 帧未见则移除 + it = tracked_persons_.erase(it); // erase 返回下一个有效迭代器 } else { - ++it; + ++it; // 手动移动到下一个 } } } @@ -265,8 +327,6 @@ void VideoService::processing_loop() { detect_result_group_t detection_results; while (running_) { - // auto t_start = std::chrono::high_resolution_clock::now(); - { std::unique_lock lock(frame_mutex_); @@ -283,14 +343,17 @@ void VideoService::processing_loop() { new_frame_available_ = false; } - // auto t_read = std::chrono::high_resolution_clock::now(); if (frame.empty()) { continue; } + cv::Mat model_input_image; + cv::resize(frame, model_input_image, cv::Size(640, 640)); + if (!model_input_image.isContinuous()) { + model_input_image = model_input_image.clone(); + } - // [保留] 后续的 AI 推理、跟踪、绘图、推流逻辑完全不变 - if (rknn_pool_->put(frame) != 0) { + if (rknn_pool_->put(model_input_image) != 0) { spdlog::error("VideoService: Failed to put frame into rknnPool. Stopping."); running_ = false; break;