GK SDK 源码库: XMIPCLinuxV100R005C00SPC030 (kernel/tools/open_source excluded)

source/gmp/usr/svp/include/post_process/yolov5.h

@@ -0,0 +1,137 @@
+#ifndef __YOLOV5__
+#define __YOLOV5__
+#include <stdio.h>
+#include <string.h>
+
+#include "xmedia_svp_std.h"
+#include "xmedia_svp.h"
+#include "svp_nms.h"
+#include "xmedia_svp_quantize.h"
+#include "svp_core.h"
+#include "svp_movement.h"
+#include "svp_bytetracker.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+// sigmoid 查表法
+#define TABLE_SIZE 100000
+#define TABLE_MIN   (-100.0f)
+#define TABLE_MAX   (100.0f)
+#define TABLE_200   (200.0f)
+#define UNSIGNED_8_BIT_NUM 256
+#define TABLE_500 (500.0f)
+
+// 每个特征中的x、y、w、h、socre长度
+#define XYWHF_LEN 5
+#define XYWH_LEN 4
+#define LAYER_MAX_NUM  10
+#define LAYER_MIN_NUM  3
+#define LAYER_NUM_FOUR 4
+
+// 每层feature_map对应的anchor
+#define ANCHOR_NUM 3
+
+/** 模型输入默认分辨率  */
+#define INPUT_DEFAULT_WIDE 640
+#define INPUT_DEFAULT_HIGH 360
+
+/** 默认检测阈值 */
+#define DEFAULT_DETECT_THRESHOLD 0.45f
+/** 默认IOU阈值 */
+#define DEFAULT_IOU_THRESHOLD 0.5f
+/** 分类器默认阈值*/
+#define DEFAULT_CLASSIFIER_THRESHOLD 0.01f
+#define DEFAULT_BBOX_NUM 512
+
+/** 模型检测类型数量  */
+#define DETECTION_MODEL_SINGLE 1
+#define DETECTION_MODEL_DOUBLE 2
+#define DETECTION_MODEL_TRIPLE 3
+#define DETECTION_MODEL_MULTI  4
+#define DETECTION_MODEL_SIX 6
+#define DETECTION_MODEL_PCNMV_FIRESMOKE 11
+#define DETECTION_MODEL_MAX  80
+
+// 稳定框算法系数
+#define SVP_IOU_STABLE_THRESH_UPPER   0.85f
+#define SVP_IOU_STABLE_THRESH_MIDDLE  0.7f
+#define SVP_STABLE_HORIZONTAL_FACTOR 6
+#define SVP_STABLE_VERTICAL_FACTOR 5
+#define SVP_UNIFORM_FACTOR_A 2
+#define SVP_UNIFORM_FACTOR_B 3
+#define SVP_STABLE_HISTORY_RATIO 0.8f
+#define SVP_STABLE_CURRENT_RATIO 0.2f
+
+// 目标框面积大小限制
+#define MAX_AREA_ATTR 4
+#define MINI_AREA 100
+
+typedef struct {
+    xmedia_svp_rect rect;
+    xmedia_svp_class_type class_type;
+} svp_stable_rect;
+
+typedef struct {
+    svp_stable_rect stable_targets[XMEDIA_SVP_MAX_TARGET_NUM];
+    xmedia_u32 num;
+} svp_stable_box;
+
+typedef struct {
+    xmedia_svp_point anchors[LAYER_MAX_NUM][ANCHOR_NUM];   // 锚点值
+    xmedia_u8 layer_num;                                 // 锚点层数
+} xmedia_svp_detect_anchors;
+
+typedef struct {
+    xmedia_u32 w;                            // 模型支持的图像宽度
+    xmedia_u32 h;                            // 模型支持的图像高度
+    xmedia_u32 num;                          // 模型的检测类别数量
+    xmedia_u32 feature[LAYER_MAX_NUM];       // 每个feature_map的大小
+    xmedia_u32 output_num;                   // npu输出所有的特征数量
+    xmedia_float thres_desig;                // 检测阈值的反sigmoid值
+    quanlize_param quanlize[LAYER_MAX_NUM];  // 反量化需要的参数
+    xmedia_npu_model model;                  // npu参数
+    xmedia_svp_alg_type type;                // 算法类型
+    xmedia_float detect_threshold;           // 置信度阈值，建议值0.55f
+    xmedia_float classifier_threshold;       // 分类器阈值，建议值0.01f
+    xmedia_float iou_threshold;              // iou相交比阈值，建议值0.5f
+    xmedia_u32 max_target_num;               // 最大目标数，最大值10
+    xmedia_bool smart_venc_enable;
+    xmedia_bool smart_ae_enable;
+    xmedia_bool smart_venc_array[XMEDIA_SVP_MAX_VENC_CHN_NUM];
+    xmedia_bool smart_ae_array[XMEDIA_SVP_MAX_VI_PIPE_NUM];
+    svp_stable_box record_result;            // 记录历史框,用于稳定算法
+    xmedia_svp_detect_anchors anchors;       // 锚点值
+    xmedia_u64 cost_time;                    // process接口耗时
+    xmedia_bool aov_flag;                    // 是否为aov模型
+    xmedia_bool aov_only_target;             // 是否仅判断有无目标
+    svp_movement movement;                   // 静止过滤相关参数
+    svp_tracklet tracklet;
+    xmedia_u8 track_id_arry[SVP_ALG_MAX_TARGET_NUM * SVP_MAX_LOST_COUNT]; // 追踪id数组
+    xmedia_s32  track_id_grow;                                               // 追踪id增加
+    xmedia_void *private_data;               // 实例私有数据
+    xmedia_float *sigmoid_table;             // sigmoid查表
+    xmedia_float *sigmoid_dequantize_table;
+    xmedia_float *sigmoid_dequantize_table_x2;
+    xmedia_float *sigmoid_dequantize_table_x6;
+    xmedia_float *sigmoid_dequantize_table_x2_sq;
+    svp_base_result *bbox;
+} yolov5_detect_param;
+
+xmedia_void set_default_yolov5_detect_param(yolov5_detect_param *param);
+
+xmedia_s32 detect_process(yolov5_detect_param *param, const xmedia_video_frame_info *input_image,
+                                xmedia_svp_yolov5_output *result, const xmedia_svp_detect_anchors anchors_yaml);
+xmedia_s32 detect_init(yolov5_detect_param *param, const xmedia_svp_alg_type type,
+                          const xmedia_svp_modules *model, xmedia_s32 layer_num);
+xmedia_s32 detect_deinit(yolov5_detect_param *param);
+
+xmedia_s32 svp_get_detect_result_yolov5_bz(yolov5_detect_param *param, xmedia_svp_yolov5_output *result,
+                                            const xmedia_svp_detect_anchors anchors_yaml);
+
+#ifdef __cplusplus
+}
+#endif
+
+#endif

source/gmp/usr/svp/include/post_process/yolov8.h

@@ -0,0 +1,65 @@
+#ifndef __YOLOV8__
+#define __YOLOV8__
+
+#include <stdio.h>
+#include <string.h>
+
+#include "xmedia_svp.h"
+#include "xmedia_svp_std.h"
+#include "xmedia_svp_quantize.h"
+
+#include "svp_nms.h"
+#include "svp_core.h"
+#include "yolov5.h"
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define DEFAULT_KEYPOINT_LEN 0
+#define PEOPLE_KEYPOINT_LEN 17
+#define SVP_PLATE_KPS 4
+
+typedef struct {
+    xmedia_u32 w;                            // 模型支持的图像宽度
+    xmedia_u32 h;                            // 模型支持的图像高度
+    xmedia_u32 num;                          // 模型的检测类别数量
+    xmedia_u32 feature[LAYER_MAX_NUM];       // 每个feature_map的大小
+    xmedia_u32 output_num;                   // npu输出所有的特征数量
+    xmedia_float thres_desig;                // 检测阈值的反sigmoid值
+    quanlize_param quanlize[LAYER_MAX_NUM];  // 反量化需要的参数
+    xmedia_npu_model model;                  // npu参数
+    xmedia_svp_alg_type type;                // 算法类型
+    xmedia_float detect_threshold;           // 置信度阈值，建议值0.55f
+    xmedia_float iou_threshold;              // iou相交比阈值，建议值0.5f
+    xmedia_float classifier_threshold;       // 分类器阈值，建议值0.01f
+    xmedia_u32 max_target_num;               // 最大目标数，最大值10
+    svp_stable_box record_result;            // 记录历史框,用于稳定算法
+    svp_movement movement;                   // 静止过滤相关参数
+    svp_tracklet tracklet;
+    xmedia_u8 track_id_arry[SVP_ALG_MAX_TARGET_NUM * SVP_MAX_LOST_COUNT]; // 追踪id数组
+    xmedia_s32  track_id_grow;                                                 // 追踪id增加
+    xmedia_u8 keypoint_len;                  // 关键点数量
+    xmedia_void *private_data;               // 实例私有数据
+    xmedia_float *deqnt_u8_to_f32_table;
+    xmedia_float *sigmoid_deqnt_table;
+    svp_base_yolov8_result *bbox;
+} yolov8_detect_param;
+
+xmedia_void set_default_yolov8_detect_param(yolov8_detect_param *param);
+
+xmedia_s32 yolov8_init(yolov8_detect_param *param, const xmedia_svp_modules *model);
+
+xmedia_s32 yolov8_uninit(yolov8_detect_param *param);
+
+xmedia_s32 detect_yolov8_process(yolov8_detect_param *param, const xmedia_video_frame_info *frame_info,
+    xmedia_svp_yolov8_output *result);
+
+xmedia_float iouv8(svp_base_yolov8_result *box1, svp_base_yolov8_result *box2);
+xmedia_s32 svp_get_detect_result_bz(yolov5_detect_param *param, xmedia_svp_yolov8_output *result,
+    const xmedia_svp_detect_anchors anchors_yaml);
+#ifdef __cplusplus
+}
+#endif
+
+#endif