AN OPTIMIZED HYBRID FRAMEWORK FOR REAL-TIME EDGE DETECTION AND FEATURE EXTRACTION IN AUTONOMOUS VEHICLE VISION SYSTEMS

Abstract

Autonomous vehicles depend on camera-based perception to recognise lane markings, vehicles, pedestrians, and traffic signs, and the first stages of that perception, edge detection and feature extraction, determine how fast and how reliably everything downstream can work. Classical edge detectors are fast but fragile under glare, shadow, and low light, whereas deep networks are robust but demand more computation than many embedded automotive processors can supply within a strict frame budget. This paper presents a hybrid framework that combines a lightweight, adaptive edge detection stage with a compact orientation-histogram descriptor, a small convolutional network that learns complementary features, and a margin-based support vector classifier that makes the final decision. The edge stage selects its thresholds automatically from local image statistics, so that it adapts to changing illumination without manual tuning. The orientation histogram summarises edge direction in eight bins and provides a stable, low-cost description of scene structure. The convolutional features are fused with the histogram and classified by a support vector machine, which performs well when training data are limited. Real-time behaviour is obtained through region-of-interest cropping, fixed-point arithmetic, operator fusion, and reuse of gradient computations. The framework was evaluated on two public driving image collections and on a network traffic dataset used to test the generality of the feature-and-classifier stage. It achieved a mean edge F-measure of 0.846, an object classification accuracy of 96.2%, and a processing rate of 47 frames per second on an automotive-grade embedded platform, using 38% less energy per frame than a deep-only baseline. A secondary experiment on vehicle network traffic reached a macro F1 score of 0.93 after bias correction. The results indicate that combining classical and learned components can meet automotive timing limits without sacrificing accuracy.