1# Copyright © 2020 Arm Ltd and Contributors. All rights reserved. 2# SPDX-License-Identifier: MIT 3 4""" 5Object detection demo that takes a video stream from a device, runs inference 6on each frame producing bounding boxes and labels around detected objects, 7and displays a window with the latest processed frame. 8""" 9 10import os 11import sys 12script_dir = os.path.dirname(__file__) 13sys.path.insert(1, os.path.join(script_dir, '..', 'common')) 14 15import cv2 16from argparse import ArgumentParser 17 18from ssd import ssd_processing, ssd_resize_factor 19from yolo import yolo_processing, yolo_resize_factor 20from utils import dict_labels 21from cv_utils import init_video_stream_capture, preprocess, draw_bounding_boxes 22from network_executor import ArmnnNetworkExecutor 23 24 25def get_model_processing(model_name: str, video: cv2.VideoCapture, input_binding_info: tuple): 26 """ 27 Gets model-specific information such as model labels and decoding and processing functions. 28 The user can include their own network and functions by adding another statement. 29 30 Args: 31 model_name: Name of type of supported model. 32 video: Video capture object, contains information about data source. 33 input_binding_info: Contains shape of model input layer, used for scaling bounding boxes. 34 35 Returns: 36 Model labels, decoding and processing functions. 37 """ 38 if model_name == 'ssd_mobilenet_v1': 39 return ssd_processing, ssd_resize_factor(video) 40 elif model_name == 'yolo_v3_tiny': 41 return yolo_processing, yolo_resize_factor(video, input_binding_info) 42 else: 43 raise ValueError(f'{model_name} is not a valid model name') 44 45 46def main(args): 47 video = init_video_stream_capture(args.video_source) 48 executor = ArmnnNetworkExecutor(args.model_file_path, args.preferred_backends) 49 50 process_output, resize_factor = get_model_processing(args.model_name, video, executor.input_binding_info) 51 labels = dict_labels(args.label_path, include_rgb=True) 52 53 while True: 54 frame_present, frame = video.read() 55 frame = cv2.flip(frame, 1) # Horizontally flip the frame 56 if not frame_present: 57 raise RuntimeError('Error reading frame from video stream') 58 input_tensors = preprocess(frame, executor.input_binding_info) 59 print("Running inference...") 60 output_result = executor.run(input_tensors) 61 detections = process_output(output_result) 62 draw_bounding_boxes(frame, detections, resize_factor, labels) 63 cv2.imshow('PyArmNN Object Detection Demo', frame) 64 if cv2.waitKey(1) == 27: 65 print('\nExit key activated. Closing video...') 66 break 67 video.release(), cv2.destroyAllWindows() 68 69 70if __name__ == '__main__': 71 parser = ArgumentParser() 72 parser.add_argument('--video_source', type=int, default=0, 73 help='Device index to access video stream. Defaults to primary device camera at index 0') 74 parser.add_argument('--model_file_path', required=True, type=str, 75 help='Path to the Object Detection model to use') 76 parser.add_argument('--model_name', required=True, type=str, 77 help='The name of the model being used. Accepted options: ssd_mobilenet_v1, yolo_v3_tiny') 78 parser.add_argument('--label_path', required=True, type=str, 79 help='Path to the labelset for the provided model file') 80 parser.add_argument('--preferred_backends', type=str, nargs='+', default=['CpuAcc', 'CpuRef'], 81 help='Takes the preferred backends in preference order, separated by whitespace, ' 82 'for example: CpuAcc GpuAcc CpuRef. Accepted options: [CpuAcc, CpuRef, GpuAcc]. ' 83 'Defaults to [CpuAcc, CpuRef]') 84 args = parser.parse_args() 85 main(args) 86