Animal Intrusion Detection Using various Deep Learning Models

Agriculture is the most important sector of Indian economy, but the issue of damage to crops by wild creatures has turned into a noteworthy social issue in current occasions. So far there is no effective solution to this problem and therefore requires immediate solution. Existing methods include building wire fences, electrical fences etc., are not so effective. So, in this paper, we aim at developing various computer vision techniques especially deep learning, for animal intrusion detection and alerting purpose. A total of 3000 images from kaggle were collected and used median filtering to preprocess the data. Later we extract the useful features i.e., count the occurrences of gradient orientation in limited section of an image using hog feature descriptor. The obtained datasets are divided into 80% training and 20% testing for all the algorithms. Furthermore, performance evaluation of various deep learning models such as convolution neural network (CNN), visual geometry group (VGG-16), and residual network-50 (ResNet-50), is carried. From the results, it is observed that detection of intrusion performed well using you only look once-version 5 (Yolov5) with 93% and the tensor flow techniques (reduced total loss accuracy of tensor flow model up to 0.95). Finally, a short message service (SMS) will be sent to land owner through telegram. With the employed frameworks, accuracies of 90%, 95%, and 60% respectively were achieved.