DeepvisionAI backend

Fully functional code

The backend functions with Tensorflow, Pytorch and yolov5

• Fully tested on the UCSD pedestrians dataset
• Easy execution from anaconda prompt

Features

• Detection of abnormal activities
• Tracking all pedestrians with high accuracy
• Generating foreground masks
• Uses optical flow features
• Uses convolutional saptial auto-encoder

Examples of the output results

1- preprocessing mask

2- Abnormal behaviour detection

3- Full tracking

Files organization

• run .py : the pilot script that runs in order other scripts
• UCSDped1 .py : contains the labeling and reorganization code for the dataset
• utils .py : takes in the number of tests to be performed as an argument and generates a folder containing only the selected samples
• test_script .py : loads the pretrained model results from features/ and processes the test sample
• frames2video .py : a script that compiles a given sequence of images to a video
• track .py : runs yolov5 scripts detect and track different classes of objects MOT (Multi-Object Tracker) the run .py restrains the process to class 0 only 'pedestrians'
• bg .py : generates foreground masks as a preprocessing step for optical_flow .py

Installation & Usage

It's recommanded to use [Anaconda].

Download the UCSD dataset : UCSD_Anomaly_Dataset.v1p2

Extract it to main directroy ./UCSD_Anomaly_Dataset.v1p2

Install the packages

pip install requirements.txt

Run the pilot script :

python run.py -test [number of tests to be conducted default = 36] -tracking [runs tracking too default = true]

Performance results

Below are the performance results compared to other state-of-the-art results.

Method	ROC AUC
Our method	0.91588
self trained deep ordinal regression	0.927
full-BVP	0.836
H-MDT CRF	0.827
STRT unsupervised	0.5945
STRT supervised	0.7118

Roc curve for 36 tests

References for statistics

[Ref 1] : https://www.researchgate.net/publication/293042967_Anomaly_detection_based_on_spatio-temporal_sparse_representation_and_visual_attention_analysis/figures?lo=1

[Ref 2] : https://www.google.com/url?sa=i&url=https%3A%2F%2Fwww.researchgate.net%2Ffigure%2FROC-curves-of-pixel-level-criterion-on-Ped1_fig1_239943156&psig=AOvVaw2NlY3dVHMzMqlnocK6xTd7&ust=1631800725361000&source=images&cd=vfe&ved=0CAoQjhxqFwoTCKCD6MiRgfMCFQAAAAAdAAAAABAD

©️ License

CERIST : Centre de Recherche sur l'Information Scientifique et Technique

ESI : Ecole Nationale Supérieure d'Informatique d'Alger (Ex. INI)

Team :

• Linda Belkessa (Chef d'équipe) https://www.linkedin.com/in/linda-belkessa/
• Sarah Abchiche https://www.linkedin.com/in/sarah-abchiche/
• Salima Mamma https://www.linkedin.com/in/salima-mamma-239002179/
• Sofia Ouanes https://www.linkedin.com/in/sofia-ouanes-18a841182/
• Abdelaziz Takouche https://www.linkedin.com/in/abdelaziz-takouche-9a990b204/
• Massinissa Si Ahmed https://www.linkedin.com/in/massinissa-si-ahmed-0463a316b/

For more informations please refer to one of the members of the team It's totally free for use under license