1 #include "darknet.h"
2 #include "yolo_v2_class.hpp"
3
4 #include "network.h"
5
6 extern "C" {
7 #include "detection_layer.h"
8 #include "region_layer.h"
9 #include "cost_layer.h"
10 #include "utils.h"
11 #include "parser.h"
12 #include "box.h"
13 #include "image.h"
14 #include "demo.h"
15 #include "option_list.h"
16 #include "stb_image.h"
17 }
18 //#include <sys/time.h>
19
20 #include <vector>
21 #include <iostream>
22 #include <algorithm>
23 #include <cmath>
24
25 #define NFRAMES 3
26
27 //static Detector* detector = NULL;
28 static std::unique_ptr<Detector> detector;
29
init(const char * configurationFilename,const char * weightsFilename,int gpu)30 int init(const char *configurationFilename, const char *weightsFilename, int gpu)
31 {
32 detector.reset(new Detector(configurationFilename, weightsFilename, gpu));
33 return 1;
34 }
35
detect_image(const char * filename,bbox_t_container & container)36 int detect_image(const char *filename, bbox_t_container &container)
37 {
38 std::vector<bbox_t> detection = detector->detect(filename);
39 for (size_t i = 0; i < detection.size() && i < C_SHARP_MAX_OBJECTS; ++i)
40 container.candidates[i] = detection[i];
41 return detection.size();
42 }
43
detect_mat(const uint8_t * data,const size_t data_length,bbox_t_container & container)44 int detect_mat(const uint8_t* data, const size_t data_length, bbox_t_container &container) {
45 #ifdef OPENCV
46 std::vector<char> vdata(data, data + data_length);
47 cv::Mat image = imdecode(cv::Mat(vdata), 1);
48
49 std::vector<bbox_t> detection = detector->detect(image);
50 for (size_t i = 0; i < detection.size() && i < C_SHARP_MAX_OBJECTS; ++i)
51 container.candidates[i] = detection[i];
52 return detection.size();
53 #else
54 return -1;
55 #endif // OPENCV
56 }
57
dispose()58 int dispose() {
59 //if (detector != NULL) delete detector;
60 //detector = NULL;
61 detector.reset();
62 return 1;
63 }
64
get_device_count()65 int get_device_count() {
66 #ifdef GPU
67 int count = 0;
68 cudaGetDeviceCount(&count);
69 return count;
70 #else
71 return -1;
72 #endif // GPU
73 }
74
built_with_cuda()75 bool built_with_cuda(){
76 #ifdef GPU
77 return true;
78 #else
79 return false;
80 #endif
81 }
82
built_with_cudnn()83 bool built_with_cudnn(){
84 #ifdef CUDNN
85 return true;
86 #else
87 return false;
88 #endif
89 }
90
built_with_opencv()91 bool built_with_opencv(){
92 #ifdef OPENCV
93 return true;
94 #else
95 return false;
96 #endif
97 }
98
99
get_device_name(int gpu,char * deviceName)100 int get_device_name(int gpu, char* deviceName) {
101 #ifdef GPU
102 cudaDeviceProp prop;
103 cudaGetDeviceProperties(&prop, gpu);
104 std::string result = prop.name;
105 std::copy(result.begin(), result.end(), deviceName);
106 return 1;
107 #else
108 return -1;
109 #endif // GPU
110 }
111
112 #ifdef GPU
check_cuda(cudaError_t status)113 void check_cuda(cudaError_t status) {
114 if (status != cudaSuccess) {
115 const char *s = cudaGetErrorString(status);
116 printf("CUDA Error Prev: %s\n", s);
117 }
118 }
119 #endif
120
121 struct detector_gpu_t {
122 network net;
123 image images[NFRAMES];
124 float *avg;
125 float* predictions[NFRAMES];
126 int demo_index;
127 unsigned int *track_id;
128 };
129
Detector(std::string cfg_filename,std::string weight_filename,int gpu_id)130 LIB_API Detector::Detector(std::string cfg_filename, std::string weight_filename, int gpu_id) : cur_gpu_id(gpu_id)
131 {
132 wait_stream = 0;
133 #ifdef GPU
134 int old_gpu_index;
135 check_cuda( cudaGetDevice(&old_gpu_index) );
136 #endif
137
138 detector_gpu_ptr = std::make_shared<detector_gpu_t>();
139 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
140
141 #ifdef GPU
142 //check_cuda( cudaSetDevice(cur_gpu_id) );
143 cuda_set_device(cur_gpu_id);
144 printf(" Used GPU %d \n", cur_gpu_id);
145 #endif
146 network &net = detector_gpu.net;
147 net.gpu_index = cur_gpu_id;
148 //gpu_index = i;
149
150 _cfg_filename = cfg_filename;
151 _weight_filename = weight_filename;
152
153 char *cfgfile = const_cast<char *>(_cfg_filename.c_str());
154 char *weightfile = const_cast<char *>(_weight_filename.c_str());
155
156 net = parse_network_cfg_custom(cfgfile, 1, 1);
157 if (weightfile) {
158 load_weights(&net, weightfile);
159 }
160 set_batch_network(&net, 1);
161 net.gpu_index = cur_gpu_id;
162 fuse_conv_batchnorm(net);
163
164 layer l = net.layers[net.n - 1];
165 int j;
166
167 detector_gpu.avg = (float *)calloc(l.outputs, sizeof(float));
168 for (j = 0; j < NFRAMES; ++j) detector_gpu.predictions[j] = (float*)calloc(l.outputs, sizeof(float));
169 for (j = 0; j < NFRAMES; ++j) detector_gpu.images[j] = make_image(1, 1, 3);
170
171 detector_gpu.track_id = (unsigned int *)calloc(l.classes, sizeof(unsigned int));
172 for (j = 0; j < l.classes; ++j) detector_gpu.track_id[j] = 1;
173
174 #ifdef GPU
175 check_cuda( cudaSetDevice(old_gpu_index) );
176 #endif
177 }
178
179
~Detector()180 LIB_API Detector::~Detector()
181 {
182 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
183 //layer l = detector_gpu.net.layers[detector_gpu.net.n - 1];
184
185 free(detector_gpu.track_id);
186
187 free(detector_gpu.avg);
188 for (int j = 0; j < NFRAMES; ++j) free(detector_gpu.predictions[j]);
189 for (int j = 0; j < NFRAMES; ++j) if (detector_gpu.images[j].data) free(detector_gpu.images[j].data);
190
191 #ifdef GPU
192 int old_gpu_index;
193 cudaGetDevice(&old_gpu_index);
194 cuda_set_device(detector_gpu.net.gpu_index);
195 #endif
196
197 free_network(detector_gpu.net);
198
199 #ifdef GPU
200 cudaSetDevice(old_gpu_index);
201 #endif
202 }
203
get_net_width() const204 LIB_API int Detector::get_net_width() const {
205 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
206 return detector_gpu.net.w;
207 }
get_net_height() const208 LIB_API int Detector::get_net_height() const {
209 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
210 return detector_gpu.net.h;
211 }
get_net_color_depth() const212 LIB_API int Detector::get_net_color_depth() const {
213 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
214 return detector_gpu.net.c;
215 }
216
217
detect(std::string image_filename,float thresh,bool use_mean)218 LIB_API std::vector<bbox_t> Detector::detect(std::string image_filename, float thresh, bool use_mean)
219 {
220 std::shared_ptr<image_t> image_ptr(new image_t, [](image_t *img) { if (img->data) free(img->data); delete img; });
221 *image_ptr = load_image(image_filename);
222 return detect(*image_ptr, thresh, use_mean);
223 }
224
load_image_stb(char * filename,int channels)225 static image load_image_stb(char *filename, int channels)
226 {
227 int w, h, c;
228 unsigned char *data = stbi_load(filename, &w, &h, &c, channels);
229 if (!data)
230 throw std::runtime_error("file not found");
231 if (channels) c = channels;
232 int i, j, k;
233 image im = make_image(w, h, c);
234 for (k = 0; k < c; ++k) {
235 for (j = 0; j < h; ++j) {
236 for (i = 0; i < w; ++i) {
237 int dst_index = i + w*j + w*h*k;
238 int src_index = k + c*i + c*w*j;
239 im.data[dst_index] = (float)data[src_index] / 255.;
240 }
241 }
242 }
243 free(data);
244 return im;
245 }
246
load_image(std::string image_filename)247 LIB_API image_t Detector::load_image(std::string image_filename)
248 {
249 char *input = const_cast<char *>(image_filename.c_str());
250 image im = load_image_stb(input, 3);
251
252 image_t img;
253 img.c = im.c;
254 img.data = im.data;
255 img.h = im.h;
256 img.w = im.w;
257
258 return img;
259 }
260
261
free_image(image_t m)262 LIB_API void Detector::free_image(image_t m)
263 {
264 if (m.data) {
265 free(m.data);
266 }
267 }
268
detect(image_t img,float thresh,bool use_mean)269 LIB_API std::vector<bbox_t> Detector::detect(image_t img, float thresh, bool use_mean)
270 {
271 detector_gpu_t &detector_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
272 network &net = detector_gpu.net;
273 #ifdef GPU
274 int old_gpu_index;
275 cudaGetDevice(&old_gpu_index);
276 if(cur_gpu_id != old_gpu_index)
277 cudaSetDevice(net.gpu_index);
278
279 net.wait_stream = wait_stream; // 1 - wait CUDA-stream, 0 - not to wait
280 #endif
281 //std::cout << "net.gpu_index = " << net.gpu_index << std::endl;
282
283 image im;
284 im.c = img.c;
285 im.data = img.data;
286 im.h = img.h;
287 im.w = img.w;
288
289 image sized;
290
291 if (net.w == im.w && net.h == im.h) {
292 sized = make_image(im.w, im.h, im.c);
293 memcpy(sized.data, im.data, im.w*im.h*im.c * sizeof(float));
294 }
295 else
296 sized = resize_image(im, net.w, net.h);
297
298 layer l = net.layers[net.n - 1];
299
300 float *X = sized.data;
301
302 float *prediction = network_predict(net, X);
303
304 if (use_mean) {
305 memcpy(detector_gpu.predictions[detector_gpu.demo_index], prediction, l.outputs * sizeof(float));
306 mean_arrays(detector_gpu.predictions, NFRAMES, l.outputs, detector_gpu.avg);
307 l.output = detector_gpu.avg;
308 detector_gpu.demo_index = (detector_gpu.demo_index + 1) % NFRAMES;
309 }
310 //get_region_boxes(l, 1, 1, thresh, detector_gpu.probs, detector_gpu.boxes, 0, 0);
311 //if (nms) do_nms_sort(detector_gpu.boxes, detector_gpu.probs, l.w*l.h*l.n, l.classes, nms);
312
313 int nboxes = 0;
314 int letterbox = 0;
315 float hier_thresh = 0.5;
316 detection *dets = get_network_boxes(&net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes, letterbox);
317 if (nms) do_nms_sort(dets, nboxes, l.classes, nms);
318
319 std::vector<bbox_t> bbox_vec;
320
321 for (int i = 0; i < nboxes; ++i) {
322 box b = dets[i].bbox;
323 int const obj_id = max_index(dets[i].prob, l.classes);
324 float const prob = dets[i].prob[obj_id];
325
326 if (prob > thresh)
327 {
328 bbox_t bbox;
329 bbox.x = std::max((double)0, (b.x - b.w / 2.)*im.w);
330 bbox.y = std::max((double)0, (b.y - b.h / 2.)*im.h);
331 bbox.w = b.w*im.w;
332 bbox.h = b.h*im.h;
333 bbox.obj_id = obj_id;
334 bbox.prob = prob;
335 bbox.track_id = 0;
336 bbox.frames_counter = 0;
337 bbox.x_3d = NAN;
338 bbox.y_3d = NAN;
339 bbox.z_3d = NAN;
340
341 bbox_vec.push_back(bbox);
342 }
343 }
344
345 free_detections(dets, nboxes);
346 if(sized.data)
347 free(sized.data);
348
349 #ifdef GPU
350 if (cur_gpu_id != old_gpu_index)
351 cudaSetDevice(old_gpu_index);
352 #endif
353
354 return bbox_vec;
355 }
356
tracking_id(std::vector<bbox_t> cur_bbox_vec,bool const change_history,int const frames_story,int const max_dist)357 LIB_API std::vector<bbox_t> Detector::tracking_id(std::vector<bbox_t> cur_bbox_vec, bool const change_history,
358 int const frames_story, int const max_dist)
359 {
360 detector_gpu_t &det_gpu = *static_cast<detector_gpu_t *>(detector_gpu_ptr.get());
361
362 bool prev_track_id_present = false;
363 for (auto &i : prev_bbox_vec_deque)
364 if (i.size() > 0) prev_track_id_present = true;
365
366 if (!prev_track_id_present) {
367 for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
368 cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
369 prev_bbox_vec_deque.push_front(cur_bbox_vec);
370 if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
371 return cur_bbox_vec;
372 }
373
374 std::vector<unsigned int> dist_vec(cur_bbox_vec.size(), std::numeric_limits<unsigned int>::max());
375
376 for (auto &prev_bbox_vec : prev_bbox_vec_deque) {
377 for (auto &i : prev_bbox_vec) {
378 int cur_index = -1;
379 for (size_t m = 0; m < cur_bbox_vec.size(); ++m) {
380 bbox_t const& k = cur_bbox_vec[m];
381 if (i.obj_id == k.obj_id) {
382 float center_x_diff = (float)(i.x + i.w/2) - (float)(k.x + k.w/2);
383 float center_y_diff = (float)(i.y + i.h/2) - (float)(k.y + k.h/2);
384 unsigned int cur_dist = sqrt(center_x_diff*center_x_diff + center_y_diff*center_y_diff);
385 if (cur_dist < max_dist && (k.track_id == 0 || dist_vec[m] > cur_dist)) {
386 dist_vec[m] = cur_dist;
387 cur_index = m;
388 }
389 }
390 }
391
392 bool track_id_absent = !std::any_of(cur_bbox_vec.begin(), cur_bbox_vec.end(),
393 [&i](bbox_t const& b) { return b.track_id == i.track_id && b.obj_id == i.obj_id; });
394
395 if (cur_index >= 0 && track_id_absent){
396 cur_bbox_vec[cur_index].track_id = i.track_id;
397 cur_bbox_vec[cur_index].w = (cur_bbox_vec[cur_index].w + i.w) / 2;
398 cur_bbox_vec[cur_index].h = (cur_bbox_vec[cur_index].h + i.h) / 2;
399 }
400 }
401 }
402
403 for (size_t i = 0; i < cur_bbox_vec.size(); ++i)
404 if (cur_bbox_vec[i].track_id == 0)
405 cur_bbox_vec[i].track_id = det_gpu.track_id[cur_bbox_vec[i].obj_id]++;
406
407 if (change_history) {
408 prev_bbox_vec_deque.push_front(cur_bbox_vec);
409 if (prev_bbox_vec_deque.size() > frames_story) prev_bbox_vec_deque.pop_back();
410 }
411
412 return cur_bbox_vec;
413 }
414
415
get_cuda_context()416 void *Detector::get_cuda_context()
417 {
418 #ifdef GPU
419 int old_gpu_index;
420 cudaGetDevice(&old_gpu_index);
421 if (cur_gpu_id != old_gpu_index)
422 cudaSetDevice(cur_gpu_id);
423
424 void *cuda_context = cuda_get_context();
425
426 if (cur_gpu_id != old_gpu_index)
427 cudaSetDevice(old_gpu_index);
428
429 return cuda_context;
430 #else // GPU
431 return NULL;
432 #endif // GPU
433 }