Merge pull request #1 from blakeblackshear/regions

Regions

Dockerfile

@@ -61,17 +61,17 @@ RUN cd /usr/local/src/ \
 RUN jupyter nbextension enable --py --sys-prefix widgetsnbextension
 
 # Download & build OpenCV
-RUN wget -q -P /usr/local/src/ --no-check-certificate https://github.com/opencv/opencv/archive/3.4.1.zip
+RUN wget -q -P /usr/local/src/ --no-check-certificate https://github.com/opencv/opencv/archive/4.0.1.zip
 RUN cd /usr/local/src/ \
- && unzip 3.4.1.zip \
+ && unzip 4.0.1.zip \
- && rm 3.4.1.zip \
+ && rm 4.0.1.zip \
- && cd /usr/local/src/opencv-3.4.1/ \
+ && cd /usr/local/src/opencv-4.0.1/ \
  && mkdir build \
- && cd /usr/local/src/opencv-3.4.1/build \ 
+ && cd /usr/local/src/opencv-4.0.1/build \ 
  && cmake -D CMAKE_INSTALL_TYPE=Release -D CMAKE_INSTALL_PREFIX=/usr/local/ .. \
  && make -j4 \
  && make install \
- && rm -rf /usr/local/src/opencv-3.4.1
+ && rm -rf /usr/local/src/opencv-4.0.1
 
 # Minimize image size 
 RUN (apt-get autoremove -y; \

README.md

@@ -1,10 +1,12 @@
 # Realtime Object Detection for RTSP Cameras
+This results in a MJPEG stream with objects identified that has a lower latency than directly viewing the RTSP feed with VLC.
 - Prioritizes realtime processing over frames per second. Dropping frames is fine.
 - OpenCV runs in a separate process so it can grab frames as quickly as possible to ensure there aren't old frames in the buffer
 - Object detection with Tensorflow runs in a separate process and ignores frames that are more than 0.5 seconds old
 - Uses shared memory arrays for handing frames between processes
 - Provides a url for viewing the video feed at a hard coded ~5FPS as an mjpeg stream
 - Frames are only encoded into mjpeg stream when it is being viewed
+- A process is created per detection region
 
 ## Getting Started
 Build the container with
@@ -23,13 +25,46 @@ docker run -it --rm \
 -v <path_to_labelmap.pbtext>:/label_map.pbtext:ro \
 -p 5000:5000 \
 -e RTSP_URL='<rtsp_url>' \
+-e REGIONS='<box_size_1>,<x_offset_1>,<y_offset_1>:<box_size_2>,<x_offset_2>,<y_offset_2>' \
 realtime-od:latest
 ```
 
 Access the mjpeg stream at http://localhost:5000
 
+## Tips
+- Lower the framerate of the RTSP feed on the camera to what you want to reduce the CPU usage for capturing the feed
+- Use SSDLite models
+
 ## Future improvements
-- MQTT messages when detected objects change
+- [ ] Look for a subset of object types
-- Dynamic changes to processing speed, ie. only process 1FPS unless motion detected
+- [ ] Try and simplify the tensorflow model to just look for the objects we care about
-- Break incoming frame into multiple smaller images and run detection in parallel for lower latency (rather than input a lower resolution)
+- [ ] MQTT messages when detected objects change
-- Parallel processing to increase FPS
+- [ ] Implement basic motion detection with opencv and only look for objects in the regions with detected motion
+- [ ] Dynamic changes to processing speed, ie. only process 1FPS unless motion detected
+- [x] Parallel processing to increase FPS
+- [ ] Look into GPU accelerated decoding of RTSP stream
+- [ ] Send video over a socket and use JSMPEG
+
+## Building Tensorflow from source for CPU optimizations
+https://www.tensorflow.org/install/source#docker_linux_builds
+used `tensorflow/tensorflow:1.12.0-devel-py3`
+
+## Optimizing the graph (cant say I saw much difference in CPU usage)
+https://github.com/tensorflow/tensorflow/blob/master/tensorflow/tools/graph_transforms/README.md#optimizing-for-deployment
+```
+docker run -it -v ${PWD}:/lab -v ${PWD}/../back_camera_model/models/ssd_mobilenet_v2_coco_2018_03_29/frozen_inference_graph.pb:/frozen_inference_graph.pb:ro tensorflow/tensorflow:1.12.0-devel-py3 bash
+
+bazel build tensorflow/tools/graph_transforms:transform_graph
+
+bazel-bin/tensorflow/tools/graph_transforms/transform_graph \
+--in_graph=/frozen_inference_graph.pb \
+--out_graph=/lab/optimized_inception_graph.pb \
+--inputs='image_tensor' \
+--outputs='num_detections,detection_scores,detection_boxes,detection_classes' \
+--transforms='
+  strip_unused_nodes(type=float, shape="1,300,300,3")
+  remove_nodes(op=Identity, op=CheckNumerics)
+  fold_constants(ignore_errors=true)
+  fold_batch_norms
+  fold_old_batch_norms'
+```

detect_objects.py

@@ -5,6 +5,7 @@ import datetime
 import ctypes
 import logging
 import multiprocessing as mp
+import threading
 from contextlib import closing
 import numpy as np
 import tensorflow as tf
@@ -23,15 +24,20 @@ PATH_TO_LABELS = '/label_map.pbtext'
 # TODO: make dynamic?
 NUM_CLASSES = 90
 
+#REGIONS = "600,0,380:600,600,380:600,1200,380"
+REGIONS = os.getenv('REGIONS')
+
+DETECTED_OBJECTS = []
+
 # Loading label map
 label_map = label_map_util.load_labelmap(PATH_TO_LABELS)
 categories = label_map_util.convert_label_map_to_categories(label_map, max_num_classes=NUM_CLASSES,
                                                             use_display_name=True)
 category_index = label_map_util.create_category_index(categories)
 
-def detect_objects(image_np, sess, detection_graph):
+def detect_objects(cropped_frame, sess, detection_graph, region_size, region_x_offset, region_y_offset):
     # Expand dimensions since the model expects images to have shape: [1, None, None, 3]
-    image_np_expanded = np.expand_dims(image_np, axis=0)
+    image_np_expanded = np.expand_dims(cropped_frame, axis=0)
     image_tensor = detection_graph.get_tensor_by_name('image_tensor:0')
 
     # Each box represents a part of the image where a particular object was detected.
@@ -51,25 +57,55 @@ def detect_objects(image_np, sess, detection_graph):
     # build an array of detected objects
     objects = []
     for index, value in enumerate(classes[0]):
-        object_dict = {}
+        score = scores[0, index]
-        if scores[0, index] > 0.5:
+        if score > 0.1:
-            object_dict[(category_index.get(value)).get('name').encode('utf8')] = \
+            box = boxes[0, index].tolist()
-                                scores[0, index]
+            box[0] = (box[0] * region_size) + region_y_offset
-            objects.append(object_dict)
+            box[1] = (box[1] * region_size) + region_x_offset
+            box[2] = (box[2] * region_size) + region_y_offset
+            box[3] = (box[3] * region_size) + region_x_offset
+            objects += [value, scores[0, index]] + box
+        # only get the first 10 objects
+        if len(objects) == 60:
+            break
 
-    # draw boxes for detected objects on image
+    return objects
-    vis_util.visualize_boxes_and_labels_on_image_array(
-        image_np,
-        np.squeeze(boxes),
-        np.squeeze(classes).astype(np.int32),
-        np.squeeze(scores),
-        category_index,
-        use_normalized_coordinates=True,
-        line_thickness=4)
 
-    return objects, image_np
+class ObjectParser(threading.Thread):
+    def __init__(self, object_arrays):
+        threading.Thread.__init__(self)
+        self._object_arrays = object_arrays
+
+    def run(self):
+        global DETECTED_OBJECTS
+        while True:
+            detected_objects = []
+            for object_array in self._object_arrays:
+                object_index = 0
+                while(object_index < 60 and object_array[object_index] > 0):
+                    object_class = object_array[object_index]
+                    detected_objects.append({
+                        'name': str(category_index.get(object_class).get('name')),
+                        'score': object_array[object_index+1],
+                        'ymin': int(object_array[object_index+2]),
+                        'xmin': int(object_array[object_index+3]),
+                        'ymax': int(object_array[object_index+4]),
+                        'xmax': int(object_array[object_index+5])
+                    })
+                    object_index += 6
+            DETECTED_OBJECTS = detected_objects
+            time.sleep(0.01)
 
 def main():
+    # Parse selected regions
+    regions = []
+    for region_string in REGIONS.split(':'):
+        region_parts = region_string.split(',')
+        regions.append({
+            'size': int(region_parts[0]),
+            'x_offset': int(region_parts[1]),
+            'y_offset': int(region_parts[2])
+        })
     # capture a single frame and check the frame shape so the correct array
     # size can be allocated in memory
     video = cv2.VideoCapture(RTSP_URL)
@@ -81,29 +117,43 @@ def main():
         exit(1)
     video.release()
 
+    shared_memory_objects = []
+    for region in regions:
+        shared_memory_objects.append({
             # create shared value for storing the time the frame was captured
             # note: this must be a double even though the value you are storing
             #       is a float. otherwise it stops updating the value in shared
             #       memory. probably something to do with the size of the memory block
-    shared_frame_time = mp.Value('d', 0.0)
+            'frame_time': mp.Value('d', 0.0),
+            # create shared array for storing 10 detected objects
+            'output_array': mp.Array(ctypes.c_double, 6*10)
+        })
+        
     # compute the flattened array length from the array shape
     flat_array_length = frame_shape[0] * frame_shape[1] * frame_shape[2]
-    # create shared array for passing the image data from capture to detect_objects
+    # create shared array for storing the full frame image data
     shared_arr = mp.Array(ctypes.c_uint16, flat_array_length)
-    # create shared array for passing the image data from detect_objects to flask
+    # shape current frame so it can be treated as an image
-    shared_output_arr = mp.Array(ctypes.c_uint16, flat_array_length)
+    frame_arr = tonumpyarray(shared_arr).reshape(frame_shape)
-    # create a numpy array with the image shape from the shared memory array
-    # this is used by flask to output an mjpeg stream
-    frame_output_arr = tonumpyarray(shared_output_arr).reshape(frame_shape)
 
-    capture_process = mp.Process(target=fetch_frames, args=(shared_arr, shared_frame_time, frame_shape))
+    capture_process = mp.Process(target=fetch_frames, args=(shared_arr, [obj['frame_time'] for obj in shared_memory_objects], frame_shape))
     capture_process.daemon = True
 
-    detection_process = mp.Process(target=process_frames, args=(shared_arr, shared_output_arr, shared_frame_time, frame_shape))
+    detection_processes = []
+    for index, region in enumerate(regions):
+        detection_process = mp.Process(target=process_frames, args=(shared_arr, 
+            shared_memory_objects[index]['output_array'], 
+            shared_memory_objects[index]['frame_time'], frame_shape, 
+            region['size'], region['x_offset'], region['y_offset']))
         detection_process.daemon = True
+        detection_processes.append(detection_process)
+
+    object_parser = ObjectParser([obj['output_array'] for obj in shared_memory_objects])
+    object_parser.start()
 
     capture_process.start()
     print("capture_process pid ", capture_process.pid)
+    for detection_process in detection_processes:
         detection_process.start()
         print("detection_process pid ", detection_process.pid)
 
@@ -115,20 +165,45 @@ def main():
         return Response(imagestream(),
                         mimetype='multipart/x-mixed-replace; boundary=frame')
     def imagestream():
+        global DETECTED_OBJECTS
         while True:
             # max out at 5 FPS
             time.sleep(0.2)
+            # make a copy of the current detected objects
+            detected_objects = DETECTED_OBJECTS.copy()
+            # make a copy of the current frame
+            frame = frame_arr.copy()
+            # convert to RGB for drawing
+            frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            # draw the bounding boxes on the screen
+            for obj in DETECTED_OBJECTS:
+                vis_util.draw_bounding_box_on_image_array(frame,
+                    obj['ymin'],
+                    obj['xmin'],
+                    obj['ymax'],
+                    obj['xmax'],
+                    color='red',
+                    thickness=2,
+                    display_str_list=["{}: {}%".format(obj['name'],int(obj['score']*100))],
+                    use_normalized_coordinates=False)
+
+            for region in regions:
+                cv2.rectangle(frame, (region['x_offset'], region['y_offset']), 
+                    (region['x_offset']+region['size'], region['y_offset']+region['size']), 
+                    (255,255,255), 2)
             # convert back to BGR
-            frame_bgr = cv2.cvtColor(frame_output_arr, cv2.COLOR_RGB2BGR)
+            frame = cv2.cvtColor(frame, cv2.COLOR_RGB2BGR)
             # encode the image into a jpg
-            ret, jpg = cv2.imencode('.jpg', frame_bgr)
+            ret, jpg = cv2.imencode('.jpg', frame)
             yield (b'--frame\r\n'
                 b'Content-Type: image/jpeg\r\n\r\n' + jpg.tobytes() + b'\r\n\r\n')
 
     app.run(host='0.0.0.0', debug=False)
 
     capture_process.join()
+    for detection_process in detection_processes:
         detection_process.join()
+    object_parser.join()
 
 # convert shared memory array into numpy array
 def tonumpyarray(mp_arr):
@@ -136,7 +211,7 @@ def tonumpyarray(mp_arr):
 
 # fetch the frames as fast a possible, only decoding the frames when the
 # detection_process has consumed the current frame
-def fetch_frames(shared_arr, shared_frame_time, frame_shape):
+def fetch_frames(shared_arr, shared_frame_times, frame_shape):
     # convert shared memory array into numpy and shape into image array
     arr = tonumpyarray(shared_arr).reshape(frame_shape)
 
@@ -153,23 +228,24 @@ def fetch_frames(shared_arr, shared_frame_time, frame_shape):
         if ret:
             # if the detection_process is ready for the next frame decode it
             # otherwise skip this frame and move onto the next one
-            if shared_frame_time.value == 0.0:
+            if all(shared_frame_time.value == 0.0 for shared_frame_time in shared_frame_times):
                 # go ahead and decode the current frame
                 ret, frame = video.retrieve()
                 if ret:
-                    # copy the frame into the numpy array
                     arr[:] = frame
-                    # signal to the detection_process by setting the shared_frame_time
+                    # signal to the detection_processes by setting the shared_frame_time
+                    for shared_frame_time in shared_frame_times:
                         shared_frame_time.value = frame_time.timestamp()
+            else:
+                # sleep a little to reduce CPU usage
+                time.sleep(0.01)
     
     video.release()
 
 # do the actual object detection
-def process_frames(shared_arr, shared_output_arr, shared_frame_time, frame_shape):
+def process_frames(shared_arr, shared_output_arr, shared_frame_time, frame_shape, region_size, region_x_offset, region_y_offset):
     # shape shared input array into frame for processing
     arr = tonumpyarray(shared_arr).reshape(frame_shape)
-    # shape shared output array into frame so it can be copied into
-    output_arr = tonumpyarray(shared_output_arr).reshape(frame_shape)
 
     # Load a (frozen) Tensorflow model into memory before the processing loop
     detection_graph = tf.Graph()
@@ -193,6 +269,9 @@ def process_frames(shared_arr, shared_output_arr, shared_frame_time, frame_shape
             if no_frames_available > 0 and (datetime.datetime.now().timestamp() - no_frames_available) > 30:
                 time.sleep(1)
                 print("sleeping because no frames have been available in a while")
+            else:
+                # rest a little bit to avoid maxing out the CPU
+                time.sleep(0.01)
             continue
         
         # we got a valid frame, so reset the timer
@@ -202,22 +281,22 @@ def process_frames(shared_arr, shared_output_arr, shared_frame_time, frame_shape
         if (datetime.datetime.now().timestamp() - shared_frame_time.value) > 0.5:
             # signal that we need a new frame
             shared_frame_time.value = 0.0
+            # rest a little bit to avoid maxing out the CPU
+            time.sleep(0.01)
             continue
         
-        # make a copy of the frame
+        # make a copy of the cropped frame
-        frame = arr.copy()
+        cropped_frame = arr[region_y_offset:region_y_offset+region_size, region_x_offset:region_x_offset+region_size].copy()
         frame_time = shared_frame_time.value
         # signal that the frame has been used so a new one will be ready
         shared_frame_time.value = 0.0
 
         # convert to RGB
-        frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+        cropped_frame_rgb = cv2.cvtColor(cropped_frame, cv2.COLOR_BGR2RGB)
         # do the object detection
-        objects, frame_overlay = detect_objects(frame_rgb, sess, detection_graph)
+        objects = detect_objects(cropped_frame_rgb, sess, detection_graph, region_size, region_x_offset, region_y_offset)
-        # copy the output frame with the bounding boxes to the output array
+        # copy the detected objects to the output array, filling the array when needed
-        output_arr[:] = frame_overlay
+        shared_output_arr[:] = objects + [0.0] * (60-len(objects))
-        if(len(objects) > 0):
-            print(objects)
 
 if __name__ == '__main__':
     mp.freeze_support()