Hu Chunming / village_ascend_arm

Commit 2ae5809357fc6f8b8aed74894bbac743c4de25ef

Authored by Hu Chunming
1 parent f171c20a

添加road_seg算法

Inline Side-by-side
Showing 10 changed files with 806 additions and 2 deletions
bin/models/road_seg/tzroad_seg240108_310p.om 0 → 100755
View file @2ae5809
No preview for this file type
build/src/Makefile
  View file @2ae5809
@@ -20,6 +20,7 @@ INCLUDES = -I$(PROJ_ALL_PATH)/src/common \ @@ -20,6 +20,7 @@ INCLUDES = -I$(PROJ_ALL_PATH)/src/common \
20 -I$(PROJ_ALL_PATH)/src/common/dvpp \ 20 -I$(PROJ_ALL_PATH)/src/common/dvpp \
21 -I$(PROJ_ALL_PATH)/src/common/cnn \ 21 -I$(PROJ_ALL_PATH)/src/common/cnn \
22 -I$(PROJ_ALL_PATH)/src/common/cnn_cls \ 22 -I$(PROJ_ALL_PATH)/src/common/cnn_cls \
  23 + -I$(PROJ_ALL_PATH)/src/common/road_cnn \
23 -I$(PROJ_ALL_PATH)/src/common/dvppx \ 24 -I$(PROJ_ALL_PATH)/src/common/dvppx \
24 -I$(PROJ_ALL_PATH)/src/common/model_process \ 25 -I$(PROJ_ALL_PATH)/src/common/model_process \
25 -I$(CUR_PROJ_PATH)/../va \ 26 -I$(CUR_PROJ_PATH)/../va \
@@ -65,6 +66,7 @@ SRCS := $(wildcard $(CUR_PROJ_PATH)/*.cpp) \ @@ -65,6 +66,7 @@ SRCS := $(wildcard $(CUR_PROJ_PATH)/*.cpp) \
65 $(wildcard $(CUR_PROJ_PATH)/common/*.cpp) \ 66 $(wildcard $(CUR_PROJ_PATH)/common/*.cpp) \
66 $(wildcard $(CUR_PROJ_PATH)/common/dvpp/*.cpp) \ 67 $(wildcard $(CUR_PROJ_PATH)/common/dvpp/*.cpp) \
67 $(wildcard $(CUR_PROJ_PATH)/common/cnn_cls/*.cpp) \ 68 $(wildcard $(CUR_PROJ_PATH)/common/cnn_cls/*.cpp) \
  69 + $(wildcard $(CUR_PROJ_PATH)/common/road_cnn/*.cpp) \
68 $(wildcard $(CUR_PROJ_PATH)/common/cnn/*.cpp) \ 70 $(wildcard $(CUR_PROJ_PATH)/common/cnn/*.cpp) \
69 $(wildcard $(CUR_PROJ_PATH)/common/dvppx/*.cpp) \ 71 $(wildcard $(CUR_PROJ_PATH)/common/dvppx/*.cpp) \
70 $(wildcard $(CUR_PROJ_PATH)/common/model_process/*.cpp) \ 72 $(wildcard $(CUR_PROJ_PATH)/common/model_process/*.cpp) \
@@ -95,6 +97,9 @@ $(TARGET):$(OBJS) @@ -95,6 +97,9 @@ $(TARGET):$(OBJS)
95 %.o:$(CUR_PROJ_PATH)/common/cnn_cls/%.cpp 97 %.o:$(CUR_PROJ_PATH)/common/cnn_cls/%.cpp
96 $(XX) $(CXXFLAGS) -c $< 98 $(XX) $(CXXFLAGS) -c $<
97 99
  100 +%.o:$(CUR_PROJ_PATH)/common/road_cnn/%.cpp
  101 + $(XX) $(CXXFLAGS) -c $<
  102 +
98 %.o:$(CUR_PROJ_PATH)/common/dvppx/%.cpp 103 %.o:$(CUR_PROJ_PATH)/common/dvppx/%.cpp
99 $(XX) $(CXXFLAGS) -c $< 104 $(XX) $(CXXFLAGS) -c $<
100 105
src/PicAnalysis.cpp
  View file @2ae5809
@@ -50,7 +50,12 @@ int PicAnalysis::init(int dev_id) { @@ -50,7 +50,12 @@ int PicAnalysis::init(int dev_id) {
50 // return -1; 50 // return -1;
51 // } 51 // }
52 52
53 - ret = m_motor_phone_algorithm.init(dev_id); 53 + // ret = m_motor_phone_algorithm.init(dev_id);
  54 + // if(0 != ret){
  55 + // return -1;
  56 + // }
  57 +
  58 + ret = m_road_seg_algorithm.init(dev_id);
54 if(0 != ret){ 59 if(0 != ret){
55 return -1; 60 return -1;
56 } 61 }
@@ -117,7 +122,9 @@ int PicAnalysis::analysis_sync(vector&lt;string&gt; vec_file_path){ @@ -117,7 +122,9 @@ int PicAnalysis::analysis_sync(vector&lt;string&gt; vec_file_path){
117 122
118 // m_motor_rainshed_algorithm.detect(vec_img); 123 // m_motor_rainshed_algorithm.detect(vec_img);
119 124
120 - m_motor_phone_algorithm.detect(vec_img); 125 + // m_motor_phone_algorithm.detect(vec_img);
  126 +
  127 + m_road_seg_algorithm.detect(vec_img);
121 128
122 LOG_INFO("analysis_sync finished!"); 129 LOG_INFO("analysis_sync finished!");
123 130
src/PicAnalysis.h
  View file @2ae5809
@@ -6,6 +6,7 @@ @@ -6,6 +6,7 @@
6 #include "./ai_engine_module/HumanCarAnalysis.h" 6 #include "./ai_engine_module/HumanCarAnalysis.h"
7 #include "./ai_engine_module/MotorRainshedAnalysis.h" 7 #include "./ai_engine_module/MotorRainshedAnalysis.h"
8 #include "./ai_engine_module/MotorPhoneAnalysis.h" 8 #include "./ai_engine_module/MotorPhoneAnalysis.h"
  9 +#include "./ai_engine_module/RoadSegAnalysis.h"
9 10
10 using namespace std; 11 using namespace std;
11 12
@@ -35,6 +36,7 @@ private: @@ -35,6 +36,7 @@ private:
35 HumanCarAnalysis m_human_car_algorithm; 36 HumanCarAnalysis m_human_car_algorithm;
36 MotorRainshedAnalysis m_motor_rainshed_algorithm; 37 MotorRainshedAnalysis m_motor_rainshed_algorithm;
37 MotorPhoneAnalysis m_motor_phone_algorithm; 38 MotorPhoneAnalysis m_motor_phone_algorithm;
  39 + RoadSegAnalysis m_road_seg_algorithm;
38 }; 40 };
39 41
40 42
src/ai_engine_module/RoadSegAnalysis.cpp 0 → 100644
  View file @2ae5809
  1 +#include "RoadSegAnalysis.h"
  2 +#include "road_seg.h"
  3 +
  4 +RoadSegAnalysis::RoadSegAnalysis(/* args */)
  5 +{
  6 +}
  7 +
  8 +RoadSegAnalysis::~RoadSegAnalysis()
  9 +{
  10 + release();
  11 +}
  12 +
  13 +int RoadSegAnalysis::init(int devId){
  14 + ACL_CALL(aclrtCreateContext(&ctx, devId), SY_SUCCESS, SY_FAILED);
  15 +
  16 + rs_param param;
  17 + param.modelNames = "./models/road_seg/tzroad_seg240108_310p.om";
  18 + param.thresld = 0.25;
  19 + param.devId = devId;
  20 +
  21 + cout << "rs_init start " << endl;
  22 + int ret = rs_init(&m_handle, param);
  23 + if (ret != 0) {
  24 + return -1;
  25 + }
  26 +
  27 + cout << "rs_init success " << endl;
  28 +
  29 + return SY_SUCCESS;
  30 +}
  31 +
  32 +int RoadSegAnalysis::detect(vector<sy_img> vec_img){
  33 +
  34 + ACL_CALL(aclrtSetCurrentContext(ctx), SY_SUCCESS, SY_FAILED);
  35 +
  36 + const int batchsize = vec_img.size();
  37 + rs_result results[batchsize];
  38 + const int fea_size = 360*640;
  39 + for (int b = 0; b < batchsize; b++) {
  40 + results[b].seg_array = new unsigned char[fea_size];
  41 + results[b].direct_seg = new unsigned char[fea_size];
  42 + }
  43 +
  44 + int ret = SY_FAILED;
  45 +
  46 + do
  47 + {
  48 + ret = rs_batch(m_handle, vec_img.data(), batchsize, results);
  49 + if (SY_SUCCESS != ret) {
  50 + printf("mrc_batch failed!");
  51 + break;
  52 + }
  53 +
  54 + // todo 结果处理
  55 + } while (0);
  56 +
  57 + for (int b = 0; b < batchsize; b++) {
  58 + delete[] results[b].seg_array; results[b].seg_array = NULL;
  59 + delete[] results[b].direct_seg; results[b].direct_seg = NULL;
  60 + }
  61 +
  62 + return ret;
  63 +}
  64 +
  65 +int RoadSegAnalysis::release() {
  66 +
  67 + ACL_CALL(aclrtSetCurrentContext(ctx), SY_SUCCESS, SY_FAILED);
  68 +
  69 + if (m_handle) {
  70 + rs_release(&m_handle);
  71 + }
  72 +
  73 + if(ctx){
  74 + aclrtDestroyContext(ctx);
  75 + ctx = nullptr;
  76 + }
  77 +
  78 + return SY_SUCCESS;
  79 +}
0 \ No newline at end of file 80 \ No newline at end of file
src/ai_engine_module/RoadSegAnalysis.h 0 → 100644
  View file @2ae5809
  1 +#include "include.h"
  2 +
  3 +class RoadSegAnalysis
  4 +{
  5 +public:
  6 + RoadSegAnalysis(/* args */);
  7 + ~RoadSegAnalysis();
  8 +
  9 + int init(int devId);
  10 +
  11 + int detect(vector<sy_img> vec_img);
  12 +
  13 +private:
  14 + int release();
  15 +
  16 +private:
  17 + void* m_handle{nullptr};
  18 + aclrtContext ctx{nullptr};
  19 +};
  20 +
src/ai_engine_module/road_seg.cpp 0 → 100755
  View file @2ae5809
  1 +#include "road_seg.h"
  2 +#include "sy_errorinfo.h"
  3 +#include "road_extractor.h"
  4 +#include "dvpp_processx.h"
  5 +#include <time.h>
  6 +#include <sys/time.h>
  7 +#include "stream_data.h"
  8 +#include <algorithm>
  9 +#include <map>
  10 +#include <cstring>
  11 +#include <cmath>
  12 +
  13 +
  14 +using namespace atlas_utils;
  15 +using namespace std;
  16 +
  17 +struct Resource {
  18 + aclrtContext ctx;
  19 + aclrtStream stream;
  20 +};
  21 +
  22 +typedef struct Tools {
  23 + Resource src;
  24 + ROADExtract* roadsegExtract;
  25 + DvppProcessx* dvpp;
  26 +
  27 +}Tools;
  28 +
  29 +
  30 +int rs_init(void **handle, rs_param param){
  31 +
  32 + int ret = SY_SUCCESS;
  33 + Tools* tools = new Tools;
  34 + // init resource
  35 + ACL_CALL(aclrtCreateStream(&tools->src.stream), ACL_SUCCESS, SY_FAILED);
  36 +
  37 + // head_shoulder detection init
  38 + tools->roadsegExtract = new ROADExtract();
  39 + tools->roadsegExtract->config.confThr = param.thresld;
  40 +
  41 + ret = tools->roadsegExtract->Init(param.modelNames);
  42 + if (ret != SY_SUCCESS) {
  43 + delete tools->roadsegExtract;
  44 + tools->roadsegExtract = nullptr;
  45 + return SY_FAILED;
  46 + }
  47 +
  48 + tools->dvpp = new DvppProcessx();
  49 + tools->dvpp->InitResource(tools->src.stream);
  50 +
  51 + *handle = tools;
  52 +
  53 + return SY_SUCCESS;
  54 +}
  55 +
  56 +namespace road_seg_7cls
  57 +{
  58 + int lane_cls_classes = 6; //5 + 1
  59 + int road_seg_classes = 9; //8类别+背景
  60 + int direct_seg_classes = 5; //4类别+背景
  61 + std::vector<float> lane_softmax(const std::vector<float>& x) {
  62 + std::vector<float> result(x.size());
  63 + // 实现softmax逻辑
  64 + float max_val = *std::max_element(x.begin(), x.end());
  65 + float sum = 0.0;
  66 + for (size_t i = 0; i < x.size(); i++) {
  67 + result[i] = std::exp(x[i] - max_val);
  68 + sum += result[i];
  69 + }
  70 + for (size_t i = 0; i < x.size(); i++) {
  71 + result[i] /= sum;
  72 + }
  73 +
  74 + return result;
  75 + }
  76 +
  77 + std::vector<float> lane_linspace(float start, float end, int n) {
  78 + std::vector<float> result;
  79 + float step = (end - start) / (n - 1);
  80 + for (int i = 0; i < n; i++) {
  81 + result.push_back(start + step * i);
  82 + }
  83 + return result;
  84 + }
  85 +
  86 + bool Lane_IOU(const std::vector<float>& parent_box, const std::vector<float>& compared_box, int threshold) {
  87 + // 实现Lane_IOU逻辑
  88 + int y = lane_cls_classes, length = y + 2, offset_start = y + 3;
  89 +
  90 + int n_offsets = 72;
  91 + int n_strips = n_offsets - 1;
  92 +
  93 + int start_a = static_cast<int>((parent_box[y] * n_strips) + 0.5);
  94 + int start_b = static_cast<int>((compared_box[y] * n_strips) + 0.5);
  95 + int start = std::max(start_a, start_b);
  96 + int end_a = start_a + static_cast<int>(parent_box[length]) - 1 + 0.5 - ((parent_box[length] - 1 < 0) ? 1 : 0);
  97 + int end_b = start_b + static_cast<int>(compared_box[length]) - 1 + 0.5 - ((compared_box[length] - 1 < 0) ? 1 : 0);
  98 + int end = std::min(std::min(end_a, end_b), n_strips);
  99 + if (end - start < 0) {
  100 + return false;
  101 + }
  102 + int dist = 0;
  103 + for (int i = offset_start + start; i < offset_start + end; i++) {
  104 + if (i > (offset_start + end)) {
  105 + break;
  106 + }
  107 + if (parent_box[i] < compared_box[i]) {
  108 + dist += (compared_box[i] - parent_box[i]);
  109 + } else {
  110 + dist += (parent_box[i] - compared_box[i]);
  111 + }
  112 + }
  113 + return dist < (threshold * (end - start + 1));
  114 + }
  115 +
  116 + void Lane_nms(const std::vector<std::vector<float>>& proposals, std::vector<size_t> &keep_index, size_t &num_to_keep, const std::vector<float>& scores, int overlap = 50, int top_k = 4) {
  117 + // 实现Lane_nms逻辑
  118 + /*std::vector<float> sorted_score = scores;
  119 + std::sort(sorted_score.begin(), sorted_score.end(), std::greater<float>());*/
  120 + std::vector<int> indices(scores.size());
  121 + for (size_t i = 0; i < indices.size(); i++) {
  122 + indices[i] = i;
  123 + }
  124 + std::sort(indices.begin(), indices.end(), [&](int a, int b) { return scores[a] > scores[b]; });
  125 +
  126 + std::vector<int> r_filters(scores.size(), 0);
  127 +
  128 + for (size_t i = 0; i < indices.size(); i++) {
  129 + int indice = indices[i];
  130 + if (r_filters[i] == 1) { // continue if this proposal is filtered by nms before
  131 + continue;
  132 + }
  133 + keep_index.push_back(indice);
  134 + if (keep_index.size() > static_cast<size_t>(top_k)) { // break if more than top_k
  135 + break;
  136 + }
  137 + if (i == indices.size() - 1) { // break if indice is the last one
  138 + break;
  139 + }
  140 + std::vector<int> sub_indices(indices.begin() + i + 1, indices.end());
  141 + for (size_t sub_i = 0; sub_i < sub_indices.size(); sub_i++) {
  142 + int sub_indice = sub_indices[sub_i];
  143 + bool r_filter = Lane_IOU(proposals[indice], proposals[sub_indice], overlap);
  144 + if (r_filter) {
  145 + r_filters[i + 1 + sub_i] = 1;
  146 + }
  147 + }
  148 + }
  149 +
  150 + num_to_keep = static_cast<int>(keep_index.size());
  151 + }
  152 +
  153 +
  154 + int predictions_to_pred(const std::vector<std::vector<float>>& predictions, rs_lane* reg_array, int n_strips, int n_offsets) {
  155 + std::vector<float> prior_ys = lane_linspace(1.0, 0.0, n_offsets);
  156 + int lane_count = 0;
  157 + int start_x = lane_cls_classes + 1, start_y = lane_cls_classes, offset_start = lane_cls_classes + 4, length_ = lane_cls_classes + 3;
  158 + for (const auto& lane : predictions) {
  159 + std::vector<float> lane_xs(lane.begin() + offset_start, lane.end()); // normalized value
  160 + int start = std::min(std::max(0, static_cast<int>(std::round(lane[start_y] * n_strips))), n_strips);
  161 + int length = static_cast<int>(std::round(lane[length_]));
  162 + int end = start + length - 1;
  163 + end = std::min(end, static_cast<int>(prior_ys.size()) - 1);
  164 + /*
  165 + std::vector<bool> mask(prior_ys.size(), false);
  166 + for (int i = 0; i < start; i++) {
  167 + if (lane_xs[i] >= 0.0 && lane_xs[i] <= 1.0) {
  168 + mask[i] = true;
  169 + }
  170 + }
  171 + for (int i = end + 1; i < static_cast<int>(prior_ys.size()); i++) {
  172 + if (lane_xs[i] >= 0.0 && lane_xs[i] <= 1.0) {
  173 + mask[i] = true;
  174 + }
  175 + }*/
  176 +
  177 + std::fill(lane_xs.begin() + end + 1, lane_xs.end(), -2);
  178 + std::fill(lane_xs.begin(), lane_xs.begin() + start, -2);
  179 +
  180 + std::vector<float> lane_ys;
  181 + for (size_t i = 0; i < prior_ys.size(); i++) {
  182 + // if (!mask[i]) {
  183 + if (lane_xs[i] >= 0) {
  184 + lane_ys.push_back(prior_ys[i]);
  185 + }
  186 + }
  187 +
  188 + std::vector<float> lane_xs_new;
  189 + for (size_t i = 0; i < prior_ys.size(); i++) {
  190 + // if (!mask[i]) {
  191 + if (lane_xs[i] >= 0) {
  192 + lane_xs_new.push_back(lane_xs[i]);
  193 + }
  194 + }
  195 +
  196 +
  197 + if (lane_xs_new.empty()) {
  198 + continue;
  199 + }
  200 +
  201 + int lane_xs_new_size = lane_xs_new.size();
  202 + // if (lane_xs_new_size <= 1) {
  203 + // continue;
  204 + // }
  205 +
  206 + if (lane_xs_new_size > 180) {
  207 + lane_xs_new_size = 180;
  208 + }
  209 +
  210 + int count = 0;
  211 + #if 1
  212 + for (size_t i = 0; i < lane_xs_new_size; i++) {
  213 + // float lane_x = lane_xs_new[i], lane_y = lane_ys[i];
  214 + float lane_x = lane_xs_new[i] * 640;
  215 + float lane_y = lane_ys[i] * 360;
  216 + reg_array[lane_count].points[count].x_ = lane_x;
  217 + reg_array[lane_count].points[count].y_ = lane_y;
  218 + // printf("lane_x:%f lane_y:%f %d %d\n", lane_x, lane_y, reg_array[lane_count].points[count].x_, reg_array[lane_count].points[count].y_);
  219 + count ++;
  220 + }
  221 + #endif
  222 +
  223 + reg_array[lane_count].start_x = lane[start_x]* 640;
  224 + reg_array[lane_count].start_y = lane[start_y]* 360;
  225 +
  226 +
  227 + std::vector<float> lane_info(lane.begin()+1, lane.begin()+lane_cls_classes); // 去除背景类别
  228 + std::vector<float> softmax_res = lane_softmax(lane_info);
  229 + reg_array[lane_count].cls = std::distance(softmax_res.begin(),std::max_element(softmax_res.begin(),softmax_res.end())) + 1;
  230 + reg_array[lane_count].conf = *(std::max_element(softmax_res.begin(),softmax_res.end()));
  231 + reg_array[lane_count].num_points = count;
  232 + lane_count ++;
  233 + }
  234 + return lane_count;
  235 + }
  236 +
  237 + // 统一求nms
  238 + int get_lanes(const vector<float> output, rs_lane* reg_array, float conf_threshold, int nms_thres, int max_lanes) {
  239 +
  240 + int stride1 = 192, stride2 = 82, num_points = 72, input_width = 640, input_height = 360;
  241 + int theta = lane_cls_classes + 2, length = lane_cls_classes + 3;
  242 + int n_strips = num_points - 1, n_offsets = num_points;
  243 + std::vector<std::vector<float>> predictions, nms_predictions;
  244 + std::vector<float> scores;
  245 + for (int i = 0; i < stride1; i++) {
  246 + std::vector<float> pred_score(output.begin() + i*stride2, output.begin() + i*stride2+lane_cls_classes); // 存储类别得分
  247 + std::vector<float> softmax_res = lane_softmax(pred_score);
  248 + std::vector<float> selected_scores(softmax_res.begin()+1, softmax_res.begin()+lane_cls_classes); // 选择目标类别的概率值，去除背景类别
  249 + float max_score = *std::max_element(selected_scores.begin(), selected_scores.end());
  250 + if (max_score >= conf_threshold) {
  251 + std::vector<float> pred(output.begin() + i*stride2, output.begin() + (i+1)*stride2);
  252 + nms_predictions.push_back(pred);
  253 + predictions.push_back(pred);
  254 + scores.push_back(max_score);
  255 + }
  256 + }
  257 +
  258 + for (auto& row : nms_predictions) {
  259 + row[theta] *= n_strips;
  260 + for (size_t i = length; i < row.size(); i++) {
  261 + row[i] *= (input_width - 1);
  262 + }
  263 + }
  264 +
  265 + std::vector<size_t> keep;
  266 + size_t num_to_keep;
  267 + Lane_nms(nms_predictions, keep, num_to_keep, scores, nms_thres, max_lanes);
  268 + keep.resize(num_to_keep);
  269 +
  270 + std::vector<std::vector<float>> final_predictions;
  271 + for (auto& index : keep) {
  272 + predictions[index][length] = round(predictions[index][length] * n_strips);
  273 + final_predictions.push_back(predictions[index]);
  274 + }
  275 +
  276 + int lane_count = predictions_to_pred(final_predictions, reg_array, n_strips, n_offsets);
  277 + return lane_count;
  278 + }
  279 +
  280 + // 对每个类别分别求nms
  281 + int get_lanesV2(const vector<float> output, rs_lane* reg_array, float conf_threshold, int nms_thres, int max_lanes) {
  282 +
  283 + int stride1 = 192, stride2 = 82, num_points = 72, input_width = 640, input_height = 360;
  284 + int theta = lane_cls_classes + 2, length = lane_cls_classes + 3;
  285 + int n_strips = num_points - 1, n_offsets = num_points;
  286 +
  287 + std::vector<std::vector<float>> final_predictions;
  288 + for (int cls = 0; cls < lane_cls_classes; cls++) {
  289 + std::vector<std::vector<float>> predictions, nms_predictions;
  290 + std::vector<float> scores;
  291 + for (int i = 0; i < stride1; i++) {
  292 + std::vector<float> pred_score(output.begin() + i*stride2, output.begin() + i*stride2+lane_cls_classes); // 存储类别得分
  293 + std::vector<float> softmax_res = lane_softmax(pred_score);
  294 + std::vector<float> selected_scores(softmax_res.begin()+1, softmax_res.begin()+lane_cls_classes); // 选择目标类别的概率值，去除背景类别
  295 +
  296 + auto iter = std::max_element(selected_scores.begin(), selected_scores.end());
  297 + int indice = std::distance(selected_scores.begin(), iter);
  298 + float max_score = *iter;
  299 + if (indice == cls && max_score >= conf_threshold) {
  300 + std::vector<float> pred(output.begin() + i*stride2, output.begin() + (i+1)*stride2);
  301 + nms_predictions.push_back(pred);
  302 + predictions.push_back(pred);
  303 + scores.push_back(max_score);
  304 + }
  305 + }
  306 +
  307 + for (auto& row : nms_predictions) {
  308 + row[theta] *= n_strips;
  309 + for (size_t i = length; i < row.size(); i++) {
  310 + row[i] *= (input_width - 1);
  311 + }
  312 + }
  313 +
  314 + std::vector<size_t> keep;
  315 + size_t num_to_keep;
  316 + Lane_nms(nms_predictions, keep, num_to_keep, scores, nms_thres, max_lanes);
  317 + keep.resize(num_to_keep);
  318 +
  319 +
  320 + for (auto& index : keep) {
  321 + predictions[index][length] = round(predictions[index][length] * n_strips);
  322 + final_predictions.push_back(predictions[index]);
  323 + }
  324 + }
  325 +
  326 + int lane_count = predictions_to_pred(final_predictions, reg_array, n_strips, n_offsets);
  327 + return lane_count;
  328 + }
  329 +
  330 +
  331 + // 获取道路分割结果中最大值的索引（类别）
  332 + // int road_seg_show(const float* inference_features, uint8_t *road_seg) {
  333 + int road_seg_show(const vector<float> inference_features, uint8_t *road_seg, int seg_num_cls) {
  334 + int h = 360, w = 640;
  335 + int step_size = h*w;
  336 + for (int i = 0; i < h; i++) {
  337 + for (int j = 0; j < w; j++) {
  338 + float max_score = 0; int max_cls = 0;
  339 + for (int k = 0; k < seg_num_cls; k++) {
  340 + int cur_steps = i*w+j+k*step_size;
  341 + if (inference_features[cur_steps] > max_score) {
  342 + max_score = inference_features[cur_steps];
  343 + max_cls = k;
  344 + }
  345 + }
  346 + // if (max_cls == 5 || max_cls == 6) max_cls = 0; // 将分割的车道线变成背景，防止混淆================
  347 + road_seg[i*w+j] = max_cls;
  348 + }
  349 + }
  350 +
  351 + return 0;
  352 + }
  353 +}
  354 +
  355 +int rs_process(void *handle, sy_img img_data, rs_result result)
  356 +{
  357 + return rs_batch(handle, &img_data, 1, &result);
  358 +}
  359 +
  360 +int rs_batch(void * handle, sy_img *image_data_array, int batchsize, rs_result *result){
  361 + Tools* tools = (Tools*) handle;
  362 +
  363 + int inputW = tools->roadsegExtract->GetInputWidth();
  364 + int inputH = tools->roadsegExtract->GetInputHeight();
  365 +
  366 + //printf("debug inputw:%d,inputh:%d\n",inputW,inputH);
  367 +
  368 + for (int b = 0; b < batchsize; b++) {
  369 + if (image_data_array[b].data_ == NULL || image_data_array[b].w_ == 0 || image_data_array[b].h_ == 0) {
  370 + ERROR_LOG(" RoadSegExtract get null input ptr!");
  371 + return SY_FAILED;
  372 + }
  373 +
  374 + ImageData resizeImg, src;
  375 + // Utils::CopysyImageDataToDvpp(src, image_data_array[b]);
  376 + ACL_CALL(Utils::CopysyImageDataToDvppV2(src, image_data_array[b]), SY_SUCCESS, SY_FAILED);
  377 + ACL_CALL(tools->dvpp->CropAndPadding(resizeImg, src, inputW, inputH), SY_SUCCESS, SY_FAILED);
  378 + // forward
  379 + // double t1, t2;
  380 + // t1 = msecond();
  381 + int ret = tools->roadsegExtract->Inference(resizeImg);
  382 + if (ret != SY_SUCCESS) {
  383 + return SY_MODEL_FORWARD_ERROR;
  384 + }
  385 + // t2 = msecond();
  386 + // printf("debug infer time: %.2f\n", t2 - t1);
  387 +
  388 + vector<float> detRes;
  389 + ret = tools->roadsegExtract->PostProcess(detRes);
  390 + if (ret != SY_SUCCESS) {
  391 + return SY_MODEL_GETRESULT_ERROR;
  392 + }
  393 +
  394 + int lane_length = 192 * 82;
  395 + int roadseg_length = 360 * 640 * road_seg_7cls::road_seg_classes;
  396 + std::vector<float> clane_detection(detRes.begin(), detRes.begin() + lane_length);
  397 + std::vector<float> croad_segmentation(detRes.begin() + lane_length, detRes.begin() + lane_length + roadseg_length);
  398 + std::vector<float> cdirect_segmentation(detRes.begin() + lane_length + roadseg_length, detRes.end());
  399 +
  400 + int res = road_seg_7cls::road_seg_show(croad_segmentation, result[b].seg_array, road_seg_7cls::road_seg_classes);
  401 + res = road_seg_7cls::road_seg_show(cdirect_segmentation, result[b].direct_seg, road_seg_7cls::direct_seg_classes);
  402 + float conf_threshold = 0.4;
  403 + int nms_thres = 50, max_lanes = 21;
  404 + int lane_count = road_seg_7cls::get_lanesV2(clane_detection, result[b].reg_array, conf_threshold, nms_thres, max_lanes);
  405 + result[b].lane_count = lane_count;
  406 +
  407 + vector<float>().swap(detRes);
  408 + vector<float>().swap(clane_detection);
  409 + vector<float>().swap(croad_segmentation);
  410 + vector<float>().swap(cdirect_segmentation);
  411 + }
  412 +
  413 + return SY_SUCCESS;
  414 +}
  415 +
  416 +void rs_release(void **handle) {
  417 + if (*handle) {
  418 + Tools* tools = (Tools*) *handle;
  419 + if (tools->roadsegExtract) {
  420 + delete tools->roadsegExtract;
  421 + tools->roadsegExtract = nullptr;
  422 + }
  423 + if (tools->dvpp) {
  424 + delete tools->dvpp;
  425 + tools->dvpp = nullptr;
  426 + }
  427 + // aclFinalize();
  428 + delete tools;
  429 + tools = NULL;
  430 + }
  431 +}
  432 +
  433 +const char * rs_get_version() {
  434 + return "road_seg_vdec_arm_v310p_0.0.2.20240110_without_timelimit";
  435 +}
0 \ No newline at end of file 436 \ No newline at end of file
src/ai_engine_module/road_seg.h 0 → 100755
  View file @2ae5809
  1 +/*************************************************************************
  2 +* Version: road_seg_v0.0.0.20230313
  3 +* CopyRight : 中国科学院自动化所模式识别实验室图像视频组
  4 +* UpdateDate：
  5 +* Content : 道路分割
  6 +*************************************************************************/
  7 +#ifndef ROADSEG_H_
  8 +#define ROADSEG_H_
  9 +
  10 +#if _MSC_VER
  11 +#ifdef ROADSEG_EXPORTS
  12 +#define ROADSEG_API __declspec(dllexport)
  13 +#else
  14 +#define ROADSEG_API __declspec(dllimport)
  15 +#endif
  16 +#else
  17 +#define ROADSEG_API __attribute__ ((visibility ("default")))
  18 +#endif
  19 +
  20 +#include "sy_common.h"
  21 +
  22 +#ifdef __cplusplus
  23 +extern "C"
  24 +{
  25 +#endif
  26 +
  27 +#ifndef __RS_LANE__
  28 +#define __RS_LANE__
  29 +typedef struct rs_lane {
  30 + sy_point points[180]; // 车道线上的点坐标数组
  31 + int num_points; // 车道线上点的数量
  32 + float start_x; // 起始点x坐标
  33 + float start_y; // 起始点y坐标
  34 + float cls; // 车道线类别 1-实线 2-虚线
  35 + float conf; // 车道线置信度
  36 +}rs_lane;
  37 +#endif
  38 +
  39 +#ifndef __rs_RESULT__
  40 +#define __rs_RESULT__
  41 + typedef struct rs_result
  42 + {
  43 + rs_lane reg_array[20];
  44 + int lane_count;
  45 + unsigned char *seg_array;
  46 + unsigned char *direct_seg;
  47 + }rs_result;
  48 +#endif
  49 +
  50 +#ifndef __rs_PARAM__
  51 +#define __rs_PARAM__
  52 + typedef struct rs_param
  53 + {
  54 + int devId; //运行卡号
  55 + char* modelNames;
  56 + float thresld; //检测阈值 默认为0.3
  57 + }rs_param;
  58 +#endif
  59 +
  60 + /*************************************************************************
  61 + * FUNCTION: rs_init
  62 + * PURPOSE: 载入模型
  63 + * PARAM:
  64 + [in] handle - 句柄
  65 + [in] params - 参数
  66 + * RETURN: 成功（0）或者错误代码
  67 + * NOTES:
  68 + *************************************************************************/
  69 + ROADSEG_API int rs_init(void ** handle, rs_param param);
  70 +
  71 + /*************************************************************************
  72 + * FUNCTION: rs_process
  73 + * PURPOSE: 车颜色识别
  74 + * PARAM:
  75 + [in] handle - 检测句柄
  76 + [in] img_data - 图像数据
  77 + [in] result - 结果 内存在外部申请
  78 + * RETURN: 成功(0) 或 错误代码（< 0）
  79 + * NOTES:
  80 + *************************************************************************/
  81 + ROADSEG_API int rs_process(void *handle, sy_img img_data, rs_result result);
  82 +
  83 + /*************************************************************************
  84 + * FUNCTION: rs_batch
  85 + * PURPOSE: 车颜色识别 batch
  86 + * PARAM:
  87 + [in] handle - 检测句柄
  88 + [in] img_data_array - 图像数据
  89 + [in] batch_size - 图像数目
  90 + [in] result - 结果 内存在外部申请
  91 + * RETURN: 成功(0) 或 错误代码（< 0）
  92 + * NOTES:
  93 + *************************************************************************/
  94 + ROADSEG_API int rs_batch(void *handle, sy_img* img_data_array, int batch_size, rs_result* result);
  95 +
  96 +
  97 + /*************************************************************************
  98 + * FUNCTION: rs_release
  99 + * PURPOSE: 释放
  100 + * PARAM:
  101 + [in] handle - handle
  102 + * RETURN: NULL
  103 + * NOTES:
  104 + *************************************************************************/
  105 + ROADSEG_API void rs_release(void ** handle);
  106 +
  107 +
  108 + /*************************************************************************
  109 + * FUNCTION: rs_get_version
  110 + * PURPOSE:
  111 + * PARAM: NULL
  112 + * RETURN: 版本号
  113 + * NOTES:
  114 + *************************************************************************/
  115 + ROADSEG_API const char * rs_get_version();
  116 +
  117 +#ifdef __cplusplus
  118 +};
  119 +#endif
  120 +
  121 +#endif
src/common/road_cnn/road_extractor.cpp 0 → 100755
  View file @2ae5809
  1 +#include "road_extractor.h"
  2 +#include <iostream>
  3 +#include "acl/acl.h"
  4 +#include "model_process.h"
  5 +#include "sy_errorinfo.h"
  6 +#include <time.h>
  7 +#include <sys/time.h>
  8 +#include <algorithm>
  9 +
  10 +using namespace std;
  11 +
  12 +namespace atlas_utils {
  13 +
  14 +int ROADExtract::Init(const char* modelPath) {
  15 + ACL_CALL(aclrtGetRunMode(&runMode_), SY_SUCCESS, SY_FAILED);//获取当前昇腾AI软件栈的运行模式，根据不同的运行模式，后续的接口调用方式不同
  16 + ACL_CALL(model_.LoadModelFromFileWithMem(modelPath), SY_SUCCESS, SY_FAILED);//从文件加载离线模型数据，由用户自行管理模型运行的内存
  17 + ACL_CALL(model_.CreateDesc(), SY_SUCCESS, SY_FAILED);//获取模型的描述信息
  18 + ACL_CALL(model_.CreateOutput(outDims_), SY_SUCCESS, SY_FAILED);
  19 + ACL_CALL(model_.GetInputDims(inDims_), SY_SUCCESS, SY_FAILED);
  20 + modelHeight_ = inDims_[0][1];
  21 + modelWidth_ = inDims_[0][2];
  22 +
  23 + return SY_SUCCESS;
  24 +}
  25 +
  26 +
  27 +int ROADExtract::Inference(ImageData& input) {
  28 + model_.CreateInput(input.data.get(), input.size);
  29 + ACL_CALL(model_.Execute(), SY_SUCCESS, SY_FAILED);
  30 + model_.DestroyInput(); //需调用CreateInput的销毁类接口DestroyInput！！！
  31 + return SY_SUCCESS;
  32 +}
  33 +
  34 +int ROADExtract::GetInputWidth() {
  35 + return modelWidth_;
  36 +}
  37 +
  38 +int ROADExtract::GetInputHeight() {
  39 + return modelHeight_;
  40 +}
  41 +
  42 +int ROADExtract::PostProcess(vector<float>& results) {
  43 + aclmdlDataset* modelOutput = model_.GetModelOutputData();
  44 + int outDatasetNum = aclmdlGetDatasetNumBuffers(modelOutput);
  45 + for (int i = 0; i < outDatasetNum; i++) {
  46 + aclDataBuffer* dataBuffer = aclmdlGetDatasetBuffer(modelOutput, i);
  47 + if (dataBuffer == nullptr) {
  48 + return SY_FAILED;
  49 + }
  50 +
  51 + uint32_t dataBufferSize = aclGetDataBufferSize(dataBuffer);
  52 + void* data = aclGetDataBufferAddr(dataBuffer);
  53 + if (data == nullptr) {
  54 + return SY_FAILED;
  55 + }
  56 +
  57 + int length = dataBufferSize/sizeof(float);
  58 + // float outInfo[length]; //栈上分配，受栈大小限制，数组过大可能导致栈溢出
  59 + float* outInfo = new float[length]; //堆上分配
  60 + if (runMode_ == ACL_HOST) {
  61 + ACL_CALL(aclrtMemcpy(outInfo, dataBufferSize, data, dataBufferSize, ACL_MEMCPY_DEVICE_TO_HOST),ACL_SUCCESS, SY_FAILED);
  62 + }
  63 + else {
  64 + ACL_CALL(aclrtMemcpy(outInfo, dataBufferSize, data, dataBufferSize, ACL_MEMCPY_DEVICE_TO_DEVICE),ACL_SUCCESS, SY_FAILED);
  65 + }
  66 +
  67 + for(uint32_t b = 0; b < length; b++) {
  68 + results.emplace_back(outInfo[b]);
  69 + }
  70 +
  71 + delete[] outInfo; outInfo = nullptr;
  72 + }
  73 +
  74 + return SY_SUCCESS;
  75 +}
  76 +
  77 +void ROADExtract::Release() {
  78 + model_.Unload();
  79 + model_.DestroyDesc();
  80 + model_.DestroyOutput();
  81 +}
  82 +
  83 +}
src/common/road_cnn/road_extractor.h 0 → 100755
  View file @2ae5809
  1 +#ifndef _ROAD_EXTRACT_H_
  2 +#define _ROAD_EXTRACT_H_
  3 +
  4 +#include <memory>
  5 +#include "utils.h"
  6 +#include "acl/acl.h"
  7 +#include "model_process.h"
  8 +
  9 +using namespace std;
  10 +
  11 +namespace atlas_utils {
  12 +
  13 +
  14 +#ifndef DATA_TYPE_SIZE
  15 +#define DATA_TYPE_SIZE 4
  16 +#endif
  17 +
  18 +class ROADExtract {
  19 +public:
  20 + struct ConfigParams {
  21 + float confThr = 0.0;
  22 + };
  23 +
  24 + ROADExtract() {}
  25 + ~ROADExtract() {
  26 + Release();
  27 + }
  28 +
  29 + int Init(const char* modelPath);
  30 + int Inference(ImageData& input);
  31 + int PostProcess(vector<float>& result);
  32 +
  33 + int GetInputWidth();
  34 + int GetInputHeight();
  35 +
  36 + ConfigParams config;
  37 +
  38 +private:
  39 + void Release();
  40 +
  41 + vector<vector<int>> outDims_;
  42 + vector<vector<int>> inDims_;
  43 + ModelProcess model_;
  44 + uint32_t modelWidth_;
  45 + uint32_t modelHeight_;
  46 + aclrtRunMode runMode_;
  47 +
  48 +};
  49 +
  50 +}
  51 +
  52 +#endif