Skip to content

Latest commit

 

History

History
220 lines (135 loc) · 13 KB

README.md

File metadata and controls

220 lines (135 loc) · 13 KB
Raw

Shoplifting-Detection

Project creators

SL -Shoplifting detection Provides real-time alerts for the SMB market retailers, to monitor and report customer behavior when shoplifting occur, by analyzing security camera frames and performing real-time forecasting using cutting edge lightweight deep neural network 3D-CNN architecture

Table of contents

Project highlights

  • A new novel approach in human action recognition (HAR) that combine the advantage of deep neural network 3D-CNN in learning complex features and patterns, and lightweight models architecture suitable for real-time interference and edge devices deployment.
  • SL models are built using a unique dataset designed explicitly for convenience stores and supermarkets containing row videos of theft events performed by actors
  • model designed for use in real-time environments using surveillance cameras and deliver real-time alerts on theft incidents occurring in the store.

Architecture advantages:

  • Extremely low inference time suitable for real-time system deployment
  • Run efficiently on Edge devices memory savings, due to the low number of the network parameters

Introduction

Computer vision and Human action recognition quick overview

  • Artificial intelligence and recent innovations in deep learning (DL) and neural networks took great leaps in recent years. One of the most effective types of AI is computer vision. Computer vision is the field of computer science that focuses on replicating parts of the complexity of the human vision and enabling computers to identify and process objects in images in a similar way has humans accomplish. One of the most challenging tasks in computer vision is Human Action Recognition (HAR) which aims to understand human behavior and assign a label to each action.

  • A common approach in the HAR domain, is to treat video frames as images and apply 2D-CNN (Conventional Neural Network) to recognize the action in every single frame from the video stream.One of the main deficiencies of this approach is that each video frame forms only a tiny part of the video's story, such an approach will be using incomplete information and could, therefore, easily wrongly classified, especially if there are fine-grind distinctions. In the HAR domain a single frame analysis may interoperate as different activity from the realty, when there are different shapes (e.g., person is doing pushups, or he fall down). This is termed interclass similarity, which is a common phenomenon in HAR.

  • the solution for interclass similarity problem is to extend the2D- CNN form 2D to 3D. As such the network has access not only the visual appearance present in each video frame, but also the temporal evaluation across consecutive frames. the training of 3D CNN is very computationally expensive, and the model size also has quadratic growth compared to 2D CNN. The general trend has been to make deeper and more complicated network to achieve higher accuracy. however, these advances to improve accuracy are not necessarily making networks more efficient with respect to size and speed. Despite the effectiveness of 3D-CNN volumetric models, most of the current architectures requires a huge computational power due to the convolution operation and the large number of parameters. Because of the volumetric nature of the convolutional filters to extract the features which increase the computational complexity cubically, which limits the usage of 3D-CNN volumetric in DL models for real time inference tasks.

  • Applications that relay on recognition task to be carried in a timely fashion need to understand the computation limitation of the used platform. This has special importance when we desire to identify an incident that consists of several complex action, in which the time is critical element, and the observed behavior are complex actions that require prolonged and consistent observation to decide whether an incident occurred or not. Accurate and distinctive algorithms need to be designed to solve this problem.

Problem statement

Real-time analysis of each camera has become an exhaustive task due to human limitations. The primary human limitation is the visual focus of attention. The human gaze can only concentrate on one specific point at once. Although there are large screens and high-resolution cameras, a person can only regard a small segment of the image at a time. Thieves are well aware that watching all the video footage is too demanding for "SMBs" such as retailers\ grocery\convenience stores, which makes the technology lose its role as a deterrent.

  • Shoplifting is conservatively estimated to account for 30% – 40% of total retail shrink/losses.
    • (Universities of Florida and Hayes International surveys)
  • The average shoplifting case value in 2021 was $310.11, reflecting an increase of 13.0% from 2020

Project goals:

provides a comprehensive solution for monitoring and detecting unusual events in real-time without the need for human supervision, the system will alert on a variety of scenarios. The following example describes the chain of events in the case of a shoplifting incident, where the customer steals an alcoholic beverage and hides it in a bag.

sl_proecess_1

When one of these actions will detected by our AI model, we will provide the store owner with an immediate alert.

The system monitors basic customer activities in the store, activities that we will define as pre- crime such as:

  • Taking an item off the shelf.
  • Returning an item to the shelf.
  • Examining an item. These activities are routine customer procedures in the store.

And activities that we will define as crime lapse in which our system will monitor and report in real- time on a case of theft in the store. crime lapse activities such as:

  • Concealing an item in clothes.
  • Concealing an item in a bag.

Data

Data collection

In order to train deep learning models, the first step is data collection raw video data collected from security cameras from two supermarkets, the theft was committed by actors in several different theft scenarios inside the store

Scenarios tested:

  • Product taking/returning/ examination of the product
  • Inserting an item in a pocket / coat / bag

All the cases of theft were examined in a variety of shooting angles, and by rotation of actors and clothing.

we collect 4000 video clips after the filtering process. A link to the dataset sample is provided, for the entire Dataset

send email to amitos684@gmail.com

Model architecture

Network name: Gate_Flow_SlowFast

Model description:

Inspired by SlowFast Networks for Video Recognition and the mobileNet-SSD architecture. this Model design combines the Tow gate stream architecture and the SlowFastNetwork architecture. The idea is to simulate the human brain in the aspect of visual information processing and split the data into 2 channels.

  • Slow Simulates a slow information processing process in the brain - the goal of this channel is to teach the network as deep characteristics as possible. Receives as input - 4 Frames

  • Fast Simulates a fast information processing process in the brain - the purpose of this channel is to teach the network local properties temporal feature Receives as input - 64 Frames and within this channel, there are 2 additional sub-channels of RGB, Optical Flow Receives as input - 64 Frames In the Fast-RGB channel, a Lateral connection used to connect properties to the Slow channel Slow_fast

For an in-depth understanding of the topic, I suggest reading the original paper SlowFast Networks for Video Recognition

The model architecture is based on mobileNet SSD. And the highlight of this model is utilizing tow path Slow and Fast, and for each path, there are tow channel one for optical flow and one for RGB channel.

  • Conv3D split into two channels - RGB frame and Optical flows as shown in the figure below.
  • Relu activation is adopted at the end of the RGB channel.
  • Sigmoid activation is adopted at the end of the Optical flow channel.
  • RGB and Optical Flow channels outputs are multiplied together and processed by a temporal max-pooling.
  • Merging Block is composed of basic 3D CNNs, used to process information after self learned temporal pooling.
  • Fully-connected layers generate output.

Common models in the field of HAR

MODEL PLOT

Model training && Evaluation

The model was trained in the AWS-SageMaker environment, on instance of ec2 p3.2xlarge and for machine learning implementation TensorFlow ,Keras Python,OpenCV.

We try the Adam optimization algorithm with the common value for parameters beta 1 beta 2, epsilon as show in the table

T2

Achieved 87% in F1-score Examination of the model on our dataset achieved 85.77% in F1-score Compared to the SlowFast model we got the following results 76% ee

Input-Output

SL_event_record_1__ (1) SL_event_record_1__ SL_event_record_4__ SL_event_record_5__ SL_event_record_6__ (1) SL_event_record_6__ (2) SL_event_record_6__ SL_event_record_7__ (1) SL_event_record_7__ SL_THEFT_3 SL_event_record_4__ (1)

SL_event_record_4__.mp4
SL_THEFT_3.mp4

Reference