Enhancing Vehicle Count and Classification in Videos via Deep Machine-Learning Models

Vehicle counting and categorization are crucial while analyzing data from videos. This study recommends a technique for counting and classification of vehicles from video data, manipulating a mixture of different computing methods. We examine the efficacy of coupling the bilateral filter and Histogram of Oriented Gradients (HOG) along with the CNN model for uplifting sensitivity, accuracy and specificity in vehicle identification and categorization tasks. In the initial stage, video frames are preprocessed to relieve noise using a bilateral filter, conserving edge details in addition to removal of noise. Furthermore, to enhance the performance of the system, this study proposes a hybrid approach that increases the robustness of conventional computer vision methods with Deep Learning (DL) frameworks. In this hybrid approach, the features acquired from the bilateral filter and HOG along with a CNN are incorporated to enable automated learning of sequential representations from the feed-in data. Assessment of the suggested method is executed on the Car Counting dataset for comparing specificity, sensitivity, accuracy, and mAP with discrete methodologies and habitual Machine Learning (ML) like GBM and ANN. Results acquired show the better performance of the hybrid approach, displacing improved execution in vehicle counting and categorization functions, underlining its efficiency for real-life operations. From the results obtained, the proposed CNN produces accuracy of 92%, sensitivity of 0.90, and specificity of 0.92. The tool used is Jupyter Notebook and the language used is python.