Traffic Anomaly Detection from Live-Feed through CCTV

Introduction

This project is about the detection of every day anomalies and violations on road traffic using live feed from CCTV surveillance cameras. All the things regarding the models and techniques used for this project will be discussed in detail. The motivation for this project is to decrease the manpower required on the roads, to deal with the traffic violations, as much as possible. As a couple of on-duty wardens are not capable to catch all the violators, keeping in mind the frequency with which the violations happens daily on road, the usage of CCTV cameras will not only help to provide ease to the traffic department but will also be able to create a centralized system for violation detection.

Problem Statement

Accidents and violations are unfortunately happening a lot in the world we live in right now. Some of these events get unnoticed, or are noticed when it is already too late. It is very important for authorities to reach the site of incident as early as possible to minimize the damage caused. Nowadays with CCTV’s placed almost everywhere in cities one might think that nothing escapes the eye of the police. But this is actually not true, with so much heavy traffic on roads in just one place it is really hard for the human eye to concentrate and detect these violations and a lot of them get unnoticed.

To counter this, we need an automated system which can alert the human when a violation or accident is occurring. This takes the load of looking at screens from the human eye. The automated system would take the video as input and analyze it to find any abnormal event that occurs in the video. The need for such a system to be implemented is rising as people want to live in a secure country.

Our project (“Terrific Traffic”) is an automated system which will be capable of detecting different kinds of accidents, crashes, lane violations and identify them as an anomaly. It will currently take live feed as input.

Objective

The objective of this project is to identify the correct anomaly form the video feed and bring to attention, the events, which the society has not been quite worried about lately as it should be. Our objective is to detect anomalies and bring to light these daily traffic errors to the public in order to raise awareness resulting in the minimization of these errors.

Assumptions and Constraints

In this project we assume 3 types (crash, stalled vehicles, lane violation) of events as anomalous and any other type of event happening in the video as a normal event. The 3 type of events will be classified as anomalous and will further be recognized into their respective types.

Our project will take a single video file as input and if any, clip the anomalous part, else, it will classify it as normal show a related message to the user.

Project Scope

We wanted to work on something which would be of help to the Traffic crime petrol departments and safe city projects of the county. Thus, as computer science students and having interest in the field of computer vision and machine learning, we proposed this system, which will be able detect accidents and violations events and alert related authorities accordingly.

Furthermore, this project with some modifications can be implemented in the safe cities project by the Government of Pakistan. The safe cities currently include Lahore, Islamabad and Karachi where they have placed CCTV cameras all around the city to monitor suspicious activity. Using the video coming from those cameras and giving it to our system as input we can automate the job and enhance sense of security and will further enable authorities to penalize the responsible entity in case of incident or violation.

Any company that uses private surveillance can also implement our system to automate the detection of violation and accident activities captured by their surveillance cameras.

Our model uses live stream being captured through already placed CCTV cameras to alert the local authorities. This will save the time of needing someone to make a call and report the incident. Not only this will save lives but will also develop a sense of security and will ensure that people abide by the rule of law. This serves as the scope of our proposed final year project.

Future Work

The present study has brought new dimensions and ideas to understand the research and implementation of deep learning based traffic violation models. Future studies can extend the results of our implementation by improving the accuracy of our model by using better models on the present framework. Another extension of our framework is to implement a single model that can accommodate and process multiple live feeds simultaneously. This will not only make this framework more user friendly but will also reduce the overall cost of the model because a high functioning single node model that can sustain surveillance for multiple inputs keeping good accuracy results is the best case of extension of our framework.