Identification Of Pedestrian Movement And Classification Using Deep Learning For Advanced Driver Assistance System

Prediction of pedestrian behavior is the most challenging task in today's era. Especially, in India, when a pedestrian is crossing the road it is very important to avoid accidents and slow down the traffic unnecessarily at the same time. The direction of movement of a pedestrian finds its application in autonomous driver assistance systems, security surveillance, medical applications to detect human posture abnormality, pedestrian analysis, biometric systems, robotics, etc. In order to improve the safety of the pedestrians and to enable the cars to drive independently or assist the driver in real life situations, people flow analysis along roadside, at road crossings and junctions is of utmost importance. There have been many proposed algorithms to detect a pedestrian and identify its movement but automation industries still are in search of a robust, efficient, simple and more accurate systems to achieve the desired goal. In recent years, Convolutional Neural Networks (CNN) a technique in deep learning, which is a subset of machine learning, has proved its stance in the field of artificial intelligence. CNN has successfully shown better accuracy and reliability performance than any other techniques used for similar tasks in computer vision. The paper present a system that recognizes a pedestrian moving in a particular direction using Convolutional Neural Networks. By thoroughly studying, understanding and augmenting stage-wise analysis method this paper proposes an innovative thought process that helps achieve better precision as compared to legacy techniques that involve more computational time. In this paper, CNN is used to train and classify binary as well as categorical pedestrian data. Pedestrian walking from left to right and vice versa are the two categories that are considered for binary classification. The accuracy observed for binary classification was 98.10%. This binary class became multiclass by adding two more categories of pedestrian carrying a belonging and walking from left to right and viceversa. The accuracy observed for these four categories was 87.07% for four class data.