Violation Detection on Traffic Light Area Based on Image Classification Using Dimensionality Reduction and Deep Learning

Abstract

The smart city concept is closely related to efficient traffic management, especially using technology to improve safety and smooth traffic flow, including at traffic lights. In this context, traffic light integration into a smart city system uses sensors, surveillance cameras, and intelligent algorithms to adaptively manage traffic based on vehicle volume and real-time traffic conditions. The CCTV already installed in several junction road at Bandung City, Indonesia. Image classification using deep learning is an essential and rapidly growing application in artificial intelligence.  When the number of images in the dataset collected from CCTV gets larger, the total dimension of the data will also increase significantly. The large dimension of image datasets makes the data analysis and processing process more complex and requires extensive computational resources. To reduce computational resource, dimensionality reduction using Principal Component Analysis (PCA) can handle high-dimensional data. PCA is used to process and analyze image datasets efficiently.  We proposed combination of deep learning and PCA to solve classification problem to high dimensional traffic image dataset. Experimental result showed that the non-PCA deep learning model achieved an accuracy of 73.11%, while the deep learning model with PCA achieved an accuracy of 72.73%. In other hand, the combination of deep learning and PCA showed a much shorter training time of only 2.95 seconds compared to the non-PCA deep learning model, which took 80.43 seconds.