Bharati Vidyapeeth's College of Engineering

About: Bharati Vidyapeeth's College of Engineering is a based out in . It is known for research contribution in the topics: Deep learning & Computer science. The organization has 709 authors who have published 622 publications receiving 3550 citations.

Scene detection and recognition by analysing deep features using PlacesCNN

Abstract: Scene recognition is employed for recognizing images along with some other visual features to collect information from it. As a field, it has turned out to be useful for digital marketers. Digital marketers can identify a consumer's favorite hangout spot like a cafe or bar based on his/her social media posts or uploads. Other applications include using the information from the pictures by the tour guide. CNNs help to identify whether the images belong to a specific class or not like a playground, classroom, dining room depending on the dataset. Different types of CNNs have been used to perform the classification task ranging from PlacesCNN, ImageNetCNN, HybridCNN and much more. PlacesCNN has been implemented using architectures namely AlexNet, GoogleNet and VGG. The objective of the paper is to study and analyze the performance for PlacesCNN based on VGG architecture to classify images into their correct classes along with determining the accuracy for the same. Using the pretrained model for PlacesCNN and the concept of transfer learning, we have been able to perform the task of scene recognition and achieve an accuracy of 98.25% for the same.

ADS Optimization Using Reinforcement Learning

Abstract: Our aim is to optimize the performance of ads for company products which are being deployed at user’s end to increase the revenue generated by getting more clicks and also spending less time and money in Research and Development. This is a specific application that we have chosen, but it can be applied in various fields as per the requirement of the problem. This is done by first analyzing the demands of various products by making use of various factors like product category and the time period of the year. Once it is found that the products have low sales, the ads are pushed to create interest among the users. Multiple ads which are created by the company are deployed and then optimized by analyzing the clicks that they generate over a period of time. Hit and trial way for exploring is one of the characteristic features of reinforced algorithms. For unique cases actions not only affect the present state but also the next state and the succeeding rewards. We have tried to find a way to solve the problem of multi-armed bandit (MAB) problem or N-armed bandit problem. Though several strategies have been suggested over the years, the two most prominent and commonly used are upper confidence limit (UCB) and Thompson sampling (TS). This paper explains why N-arm is preferable over A/B testing in such cases. Comparison of various approaches to solve the N-arm problem has been done. The strategies that we use for gathering of information and exploiting includes two methods, first option being arbitrary selection and the second one is that we are optimistic about uncertain machine initially and we collect the information of getting success in each round from selected machine. These actions having higher arbitrariness are favored because they provide more data advantage. We found out that Thompson sampling slightly outperforms UCB since it does a better job at manipulation.

Poster: DRIZY: Collaborative Driver Assistance Over Wireless Networks

Abstract: Driver assistance systems, that rely on vehicular sensors such as cameras, LIDAR and other on-board diagnostic sensors, have progressed rapidly in recent years to increase road safety. Road conditions in developing countries like India are chaotic where roads are not well maintained and thus vehicular sensors alone do not suffice in detecting impending collisions. In this paper, we investigate a collaborative driver assistance system "DRIZY: DRIve eaSY" for such scenarios where inference is drawn from on-board camera feed to alert drivers of obstacles ahead and the cloud uses GPS sensor data uploaded by all vehicles to alert drivers of vehicles in potential collision trajectory. Thus, we combine computer vision and vehicle-to-cloud communication to create comprehensive situational awareness. We prototype our system to consider two types of collisions: vehicle-to-vehicle collisions based on uploading GPS sensor data of vehicles to cloud and vehicle-to-pedestrian collisions based on detecting pedestrians from vehicle's dashboard camera feed. Sensor data processing in each vehicle occurs on smartphone for GPS values which are then uploaded to cloud and on raspberry pi3 for video feeds to make a cost-effective solution. Experiments over both 4G and wireless networks in India show that collaborative driver assistance is feasible in low traffic density within acceptable driver reaction time of

Analysis of Road Accidents in India and Prediction of Accident Severity

Abstract: Road accidents are a global menace, and no country can curb it. In this paper, an attempt has been made to study the various factors associated with a road accident and its effect on the cause and severity of the accident by analyzing the road accidents occurring in the nation of India from 2000 onwards. The severity of accidents can be measured in terms of human loss as well as economic loss. Further, the data is visualized on the map of India using Folium python library for the convenience of comparison between various states and better visualization. In this paper, decision tree classifier has been implemented for the prediction of the severity of a road accident. For each road accident, different parameters such as lighting conditions, vehicle type, etc., have been taken into consideration. All of the tasks have been deployed on a webpage with the help of the Flask web application framework. The proposed model achieves a testing accuracy of 79.45%.

Effective Background and Foreground Segmentation Using Unsupervised Frequency Domain Clustering

Abstract: Computer vision is the way in which the computer perceives a certain image. Background and foreground detection of an image are based on the concept of computer vision. Traditional approaches in background and foreground detection of an image imply clustering algorithms like K-means clustering, Gaussian mixture model to compute the result in the spatial domain. In spatial domain, we take into account the pixels of an image to classify them as background or foreground. In this paper, we have reviewed the process already been done in spatial domain, then we study some of the state-of-the-art background detection techniques in digital image processing and propose a novel approach to shift the domain of input images to frequency while applying the same algorithms for detection. We have used the concept in which we take into consideration the frequency rather than pixels of an image. The concept of fast Fourier transform (FFT) has been used to determine the frequency. With this solution concept, we aim at reducing the variance in input image by smoothing out the frequency domain image and experimentally demonstrate that the transition into the frequency domain outperforms the majority of techniques employed in spatial domain for background detection.