E.C.S.T.A.T.I.C.S
E.C.S.T.A.T.I.C.S stands for Enabling Civilian Security Thorugh Adaptive Tracking And Intelligent Computer Systems. Considerable amounts of abandoned luggages are reported every year in crowded areas such as airports and train stations. This has always been a source of safety and security threat to civilians as they can be used to discreetly plant explosives and wreak havoc. A single human resource is typically tasked with monitoring 5-10 screens which increases fatigueness and thus chances of making errors
Following is the brief algorithm used to detect abandoned luggages
1. Run cam at 10fps
2. Capture every 10th frame, convert it to b/w, perform morphological closing and then perform background subtraction
3. Collect 5 such frames
4. Bit-wise and the 5 frames to identify possible stationary objects
5. Collect 10 such bit-wise "and"ed frames and bit-wise and them
6. Perform blob detection to identify blobs resembling luggages
7. Identify the position of the blob in the image and crop it.
8. Pass to a modified version of ai85cdnet which can identify baggage
Currently I have implemented this on Rpi in C++ using opencv and tried the same in python as well. I have attached the results i got on the PETS2006 dataset and the code I used to get those results. It looks very promising. Running the camera at 10fps while doing steps2-7 might be an issue on the MAX78000. But the newer version (MAX78002) which is geared towards video processing should be able to do this without issues! Looking forward to implementing this on MAX78000 and making this a reality! I have also attached a minimalistic version of the MAX7800FTHR board suitable for this project.
I was able to achieve an accuracy of about 57% using just the C++ code to detect luggages. However, adding in the modified version of the ai85cdnet should be able to dramatically improve the accuracy to above 90%.
Future improvements to this code can be
1. Shadow removal from images
2. Improve the working for a wide range of lighting scenarios
3. Bring in most of the pre-processing into the neural net itself so that power consumption will be even lower.
1. Run cam at 10fps
2. Capture every 10th frame, convert it to b/w, perform morphological closing and then perform background subtraction
3. Collect 5 such frames
4. Bit-wise and the 5 frames to identify possible stationary objects
5. Collect 10 such bit-wise "and"ed frames and bit-wise and them
6. Perform blob detection to identify blobs resembling luggages
7. Identify the position of the blob in the image and crop it.
8. Pass to a modified version of ai85cdnet which can identify baggage
Currently I have implemented this on Rpi in C++ using opencv and tried the same in python as well. I have attached the results i got on the PETS2006 dataset and the code I used to get those results. It looks very promising. Running the camera at 10fps while doing steps2-7 might be an issue on the MAX78000. But the newer version (MAX78002) which is geared towards video processing should be able to do this without issues! Looking forward to implementing this on MAX78000 and making this a reality! I have also attached a minimalistic version of the MAX7800FTHR board suitable for this project.
I was able to achieve an accuracy of about 57% using just the C++ code to detect luggages. However, adding in the modified version of the ai85cdnet should be able to dramatically improve the accuracy to above 90%.
Future improvements to this code can be
1. Shadow removal from images
2. Improve the working for a wide range of lighting scenarios
3. Bring in most of the pre-processing into the neural net itself so that power consumption will be even lower.
Discussie (0 opmerking(en))