A Practical Animal Detection and Collision Avoidance System Using Deep Learning Model
Harish Kalla, R. Balachandran, Satyasis Mishra, Gemechu Dengia, R Sarankumar · 2022 IEEE 7th International conference for Convergence in Technology (I2CT) · 2022
Nowadays some countries are using advanced traffic light control systems to control the traffic signals in order to ensure smooth travel for road users. Most of the researchers use various methods to identify the density of vehicles using cameras or sensors. Sometimes these roads are occupied by animals to sleep or lie down which disturbs the smooth movement of the vehicles. Due to the occupancy of the animals on the road, traffic also increases, the density of the vehicles may increase and hence there are lots of chances that animals collide with vehicles. Most of the cities are using automatic traffic light control systems. There are no traffic police available to get rid of the animals. Especially in Ethiopia, besides pedestrians, a major number of animals are horses lying down on traffic roads. In this paper, an efficient artificial intelligence NASNetMobile model is proposed to identify the animals on traffic roads and pedestrians using an embedded machine learning system. Further, the system will provide information to the traffic police immediately using short messaging service. Standard Model recognition algorithm based on Convolution Neural Network (CNN) consumes extra computational time in comparison to the training and testing of the model. Transfer Learning is used to specifically adjust the prior models and transfer a standard deep learning model to a specific one with different loads and results. Also, the CNN structure is modified to increase overall functioning, and the traffic environs are trained to the specific scenario. The proposed system has generated test images more than 121,000 images on NASNetMobile architecture consisting even lower number of both true and false pictures and tried out on various images of creatures at junctions. A significant precise value of nearly 86.5% is obtained regarding recognition by utilizing the suggested model.