maritime_safety_security

The worl is fueled by maritime trade and we need to ensure safety and security in sea routes and ports. To do so one of the main means, is the earliest detection of craft or vessel approching our own vessel or entering in the harbord area. Security camera are there for this purpose but, can't be efficiently manned by human. Artificial Intelligence is one of the solution to complement human skill. For this project, we custom trained a YOLOV3 detection model with a six (6) classes vessel dataset available. Here are the project description:

1. find and process a suitable dataset for our purpose: ship detection
   https://www.researchgate.net/publication/327085456_SeaShips_A_Large-Scale_Precisely-Annotated_Dataset_for_Ship_Detection http://www.lmars.whu.edu.cn/prof_web/shaozhenfeng/datasets/SeaShips(7000).zip
   

   two utility python scripts are used to process the dataset in yolov3 format
   seaships.py to produce training files
   extract_xml.py to produce label file
   reference for training and dataset structure: https://towardsdatascience.com/training-yolo-for-object-detection-in-pytorch-with-your-custom-dataset-the-simple-way-1aa6f56cf7d9
   

2. Custom trained YOLOV3 model on our dataset with pytorch For this purpose colab jupyter notebook was used: seaship_custom_train.ipynb you have to provide your own google drive storage with all the files
   reference:https://github.com/cfotache/pytorch_custom_yolo_training.git
   

3. Convert model to Tensorflow for this purpose, colab jupyter notebook was used: convert_tensorflow.ipynb
   reference: https://github.com/mystic123/tensorflow-yolo-v3.git
   

4. Convert model to Intermediate representation (IR) with model optimizer
   python3 mo_tf.py
   --input_model /path/to/yolo_v3.pb
   --tensorflow_use_custom_operations_config $MO_ROOT/extensions/front/tf/yolo_v3.json
   --batch 1
   yolo_v3.json
   [
      {
         "id": "TFYOLOV3",
         "match_kind": "general",
         "custom_attributes": {
          "classes": 33,
          "anchors": [10, 13, 16, 30, 33, 23, 30, 61, 62, 45, 59, 119, 116, 90, 156, 198, 373, 326],
          "coords": 4,
          "num": 9,
          "masks":[[6, 7, 8], [3, 4, 5], [0, 1, 2]],
           "entry_points": ["detector/yolo-v3/Reshape", "detector/yolo-v3/Reshape_4", "detector/yolo-v3/Reshape_8"]
          }
      }
   ]
   reference: https://docs.openvinotoolkit.org/latest/_docs_MO_DG_prepare_model_convert_model_tf_specific_Convert_YOLO_From_Tensorflow.html#yolov3-overview
   link for IR models : https://drive.google.com/open?id=1M1Zb3N6K_sh-lCmEJqj0Hp740xby0jet
   

5. Build an inference engine app for detection
   the inference was build on two files: inference.py, ship_detection.py
   ship_detection.py is build on top of yolov3 openvino demo
   usage: ship_detection.py [-h] -m MODEL -i INPUT
               [-l CPU_EXTENSION] [-d DEVICE]
               [-l CPU_EXTENSION] [-d DEVICE]
               [--labels LABELS] [-t PROB_THRESHOLD]
               [-iout IOU_THRESHOLD] [-ni NUMBER_ITER]
               [-pc] [-r]
         Options:
         -h, --help Show this help message and exit.
         -m MODEL, --model MODEL
          Required. Path to an .xml file with a trained model.
         -i INPUT, --input INPUT
          Required. Path to an image/video file. (Specify 'cam'
          to work with camera)
          -l CPU_EXTENSION, --cpu_extension CPU_EXTENSION
          Optional. Required for CPU custom layers. Absolute
          path to a shared library with the kernels
          implementations.
         -d DEVICE, --device DEVICE
          Optional. Specify the target device to infer on; CPU,
          GPU, FPGA, HDDL or MYRIAD is acceptable. The sample
          will look for a suitable plugin for device specified.
          Default value is CPU
         --labels LABELS Optional. Labels mapping file
          -t PROB_THRESHOLD, --prob_threshold PROB_THRESHOLD
          Optional. Probability threshold for detections
          filtering
         -iout IOU_THRESHOLD, --iou_threshold IOU_THRESHOLD
          Optional. Intersection over union threshold for
         overlapping detections filtering
         -ni NUMBER_ITER, --number_iter NUMBER_ITER
          Optional. Number of inference iterations
         -pc, --perf_counts Optional. Report performance counters
         -r, --raw_output_message
         Optional. Output inference results raw values showing
   

file out.mp4 is created and contains the detections
reference: https://docs.openvinotoolkit.org/latest/_demos_python_demos_object_detection_demo_yolov3_async_README.html
To do :
train a better model with larger dataset. Due to time constraint and computing ressource, this one were train on 20 epoch (2200 iterations). We are targeting 12000 iterations at least implement realtime detection

DEMO VIDEO:
demo_video_0.8_0.4 Confidence_threshold = 0.8 and IOU = 0.4
demo_video_0.9_0.4 Confidence_threshold = 0.9 and IOU = 0.1