Overview / Usage

The poor infrastructure and management of traffic at road junctions and curves which leads to increased traffic congestions, road accidents etc. The development proposals of such locations require problems measurements and proper analysis in detail is essential. The process of identification of such locations requires two major things such as traffic flow data collection and analysis to find the solution. The identification can be done easily difficult part is to study what are the problems and to find solutions to them. This proposal addresses them using modern technology and algorithms. To study the affected area a video based data acquisition through surveillance camera is used and the solution is given by the predictive model.
Firstly, Proposal involves the Surveillance camera that records the video at the location under study the data can be acquired through this.A deep learning model which is used to accelerate the video frame rate at edge and distance constrains area coverage to track the vehicular flow precisely. Secondly ,using a edge centric approach where raw video feeds can be pre-processed at the point of capture while integration and deeper analytics is performed in the cloud. Later collected data is funnelled to a predictive model which uses machine learning algorithms and it also involves the supervised learning to analyse the data and predict the best possible solution for the problem.

Methodology / Approach

Data Collection: The surveillance cameras setup at various junctions will consist of Raspberry pi3 ,integrated with Intel Movidius Neural compute stick (NCS) .The low-power vision processing unit (VPU) architecture enables an entirely new segment of AI applications that are not reliant on a connection to the cloud. As the industry’s first always on vision processor , this VPU delivers high performance machine vision and visual awareness in severely power constraint environment .We will be using MobileNets and (SSD) Single Shot MultiBox detector framework for efficient object classification. MobileNetsSSD are based on a streamlined architecture that uses depth wise separable convolutions to build light weight deep neural networks. It uses the Pascal VOC and COCO dataset. In the development phase we use the caffe models to generate the graphs file from prototxt and weights file and run it on the Neural compute stick integrated with the Raspberry pi3. We will be able to achieve 6-8 fps which is much better than other models like tinyYOLO where we can achieve only 3-4 fps. So finally the traffic data is uploaded into the cloud server hence the data is collected in the first phase.

Data Processing: The Collected Data of over 20 months is trained in a neural network . Instead of time series classics like ARMA(Auto regressive moving average) or ARIMA(autoregressive integrated moving average) models or the Kaggle competition classic XGBoost(eXtreme Gradient Boosting). We will use Exploratory Data Analysis using Python here for prototyping and Exploring the data and then train the neural network and hence try to accurately predict the traffic inflow for the next 4 months .The analysis part involves a predictive model which determine the exact problem and help in taking the necessary course of action. The main focus will be on Supervised learning which requires training on the server. The aim is to build a machine learning algorithm that can predict the problem so that smart decisions can be taken in order to ease traffic and benefit the general population.

Technologies Used

● Intel Movidius Neural Compute Stick(NCS)
●Raspberry Pi3
●Camera
● Python
●Intel AIDevCloud
●Python
●Open CV

Collaborators