Track directly from rtsp

README.md

@@ -11,10 +11,11 @@
 
 These are roughly the steps to go from datagathering to training
 
-1. Make sure to have some recordings with a fixed camera.
+1. Make sure to have some recordings with a fixed camera. [UPDATE: not needed anymore, except for calibration & homography footage]
     * Recording can be done with `ffmpeg  -rtsp_transport udp -i rtsp://USER:PASS@IP:554/Streaming/Channels/1.mp4   hof2-cam-$(date "+%Y%m%d-%H%M").mp4`
 2. Follow the steps in the auxilary  [traptools](https://git.rubenvandeven.com/security_vision/traptools) repository to obtain (1) camera matrix, lens distortion, image dimensions, and (2+3) homography
 3. Run the tracker, e.g. `poetry run tracker --detector ultralytics  --homography ../DATASETS/NAME/homography.json   --video-src ../DATASETS/NAME/*.mp4 --calibration ../DATASETS/NAME/calibration.json  --save-for-training EXPERIMENTS/raw/NAME/`
+    * Note: You can run this right of the camera stream: `poetry run tracker --eval_device cuda:0 --detector ultralytics   --video-src  rtsp://USER:PW@ADDRESS/STREAM --homography ../DATASETS/NAME/homography.json   --calibration ../DATASETS/NAME/calibration.json  --save-for-training EXPERIMENTS/raw/NAME/`, each recording adding a new file to the `raw` folder.
 4. Parse tracker data to Trajectron format: `poetry run process_data --src-dir EXPERIMENTS/raw/NAME --dst-dir EXPERIMENTS/trajectron-data/ --name NAME`
 5. Train Trajectron model `poetry run trajectron_train --eval_every 10 --vis_every 1 --train_data_dict NAME_train.pkl --eval_data_dict NAME_val.pkl --offline_scene_graph no --preprocess_workers 8 --log_dir EXPERIMENTS/models --log_tag _NAME --train_epochs 100 --conf EXPERIMENTS/config.json --batch_size 256 --data_dir EXPERIMENTS/trajectron-data `
 6. The run!

trap/animation_renderer.py

@@ -109,10 +109,10 @@ class AnimationRenderer:
         self.batch_anim = pyglet.graphics.Batch()
 
         self.debug_lines = [
-            pyglet.shapes.Line(1380, self.config.camera.h, 1380, 670, 2, (255,255,255,255), batch=self.batch_overlay),
+            pyglet.shapes.Line(1370, self.config.camera.h-360, 1380, 670-360, 2, (255,255,255,255), batch=self.batch_overlay),
-            pyglet.shapes.Line(0, 660, 1380, 670, 2, (255,255,255,255), batch=self.batch_overlay),
+            pyglet.shapes.Line(0, 660-360, 1380, 670-360, 2, (255,255,255,255), batch=self.batch_overlay),
-            pyglet.shapes.Line(1140, 760, 1140, 675, 2, (255,255,255,255), batch=self.batch_overlay),
+            pyglet.shapes.Line(1140, 760-360, 1140, 675-360, 2, (255,255,255,255), batch=self.batch_overlay),
-            pyglet.shapes.Line(0, 750, 1380, 760, 2, (255,255,255,255), batch=self.batch_overlay),
+            pyglet.shapes.Line(0, 770-360, 1380, 770-360, 2, (255,255,255,255), batch=self.batch_overlay),
 
         ]
 

trap/prediction_server.py

@@ -113,6 +113,20 @@ def get_maps_for_input(input_dict, scene, hyperparams):
     return maps_dict
 
 
+# If homography is in cm, predictions can be terrible. Correct that here
+# TODO)) This should actually not be here, but we should use alternative homography
+# and then scale up in rendering
+def history_cm_to_m(history):
+    return [(h[0]/100, h[1]/100) for h in history]
+
+def prediction_m_to_cm(source):
+    # histories_dict[t][node] 
+    for t in source:
+        for node in source[t]:
+            source[t][node] *= 100
+    # print(t,node, source[t][node])
+    return source
+
 class PredictionServer:
     def __init__(self, config: Namespace, is_running: Event):
         self.config = config
@@ -270,6 +284,7 @@ class PredictionServer:
                     # TODO: modify this into a mapping function between JS data an the expected Node format
                     # node = FakeNode(online_env.NodeType.PEDESTRIAN)
                     history = [[h['x'], h['y']] for h in track.get_projected_history_as_dict(frame.H, self.config.camera)]
+                    history = history_cm_to_m(history)
                     history = np.array(history)
                     x = history[:, 0]
                     y = history[:, 1]
@@ -360,6 +375,10 @@ class PredictionServer:
                                                                                         hyperparams['prediction_horizon']
                                                                                         )
             
+            
+            prediction_dict, histories_dict, futures_dict = prediction_m_to_cm(prediction_dict), prediction_m_to_cm(histories_dict), prediction_m_to_cm(futures_dict)
+            
+            
             assert(len(prediction_dict.keys()) <= 1)
             if len(prediction_dict.keys()) == 0:
                 return

trap/tracker.py

@@ -27,6 +27,7 @@ from trap.frame_emitter import DetectionState, Frame, Detection, Track
 
 from tsmoothie.smoother import KalmanSmoother, ConvolutionSmoother
 import tsmoothie.smoother
+from datetime import datetime
 
 # Detection = [int, int, int, int, float, int]
 # Detections = [Detection]
@@ -56,6 +57,24 @@ def _yolov8_track(frame: Frame, model: YOLO, **kwargs) -> List[Detection]:
         
         return [Detection(track_id, bbox[0]-.5*bbox[2], bbox[1]-.5*bbox[3], bbox[2], bbox[3], 1, DetectionState.Confirmed, frame.index) for bbox, track_id in zip(results[0].boxes.xywh.cpu(), results[0].boxes.id.int().cpu().tolist())]
 
+class Multifile():
+    def __init__(self, srcs: List[Path]):
+        self.srcs = srcs
+        self.g = self.__iter__()
+
+    @property
+    def name(self):
+        return ", ".join([s.name for s in self.srcs])
+    
+    def __iter__(self):
+        for path in self.srcs:
+            with path.open('r') as fp:
+                for l in fp:
+                    yield l
+    
+    def readline(self):
+        return self.g.__next__()
+
 
 class TrainingDataWriter:
     def __init__(self, training_path = Optional[Path]):
@@ -77,7 +96,8 @@ class TrainingDataWriter:
 
     def __enter__(self):
         if self.path:
-            self.training_fp = open(self.path / 'all.txt', 'w')
+            d = datetime.now().isoformat(timespec="minutes")
+            self.training_fp = open(self.path / f'all-{d}.txt', 'w')
             # following https://github.com/StanfordASL/Trajectron-plus-plus/blob/master/experiments/pedestrians/process_data.py
             self.csv = csv.DictWriter(self.training_fp, fieldnames=['frame_id', 'track_id', 'l', 't', 'w', 'h', 'x', 'y', 'state'], delimiter='\t', quoting=csv.QUOTE_NONE)
             self.count = 0
@@ -110,13 +130,23 @@ class TrainingDataWriter:
             return
         
         self.training_fp.close()
+
+        source_files = list(self.path.glob("*.txt")) # we loop twice, so need a list instead of generator
+        total = 0
+        sources = Multifile(source_files)
+        for line in sources:
+            if len(line) > 3: # make sure not to count empty lines
+                total += 1 
+
+
         lines = {
-            'train': int(self.count * .8),
+            'train': int(total * .8),
-            'val': int(self.count * .12),
+            'val': int(total * .12),
-            'test': int(self.count * .08),
+            'test': int(total * .08),
         }
-        logger.info(f"Splitting gathered data from {self.training_fp.name}")
+
-        with open(self.training_fp.name, 'r') as source_fp:
+        logger.info(f"Splitting gathered data from {sources.name}")
+        # for source_file in source_files:
         for name, line_nrs in lines.items():
             dir_path = self.path / name
             dir_path.mkdir(exist_ok=True)
@@ -124,8 +154,7 @@ class TrainingDataWriter:
             logger.debug(f"- Write {line_nrs} lines to {file}")
             with file.open('w') as target_fp:
                 for i in range(line_nrs):
-                        target_fp.write(source_fp.readline())
+                    target_fp.write(sources.readline())
-