Real Time Pedestrian Detection Using Robust Enhanced YOLOv3+

Autonomous pedestrian detection plays a vital role in Computer Vision tasks such as smart video surveillance, smart traffic monitoring system and smart obstacle detections for building smart cities. Real-time performance is much required in self-driving cars particularly while detecting smaller pedestrians without losing any detection accuracy. The proposed paper introduces an anti-residual module in the robust Enhanced YOLOv3+ network to improve feature extraction. The proposed network is optimized by reducing bounding box loss error. This network is trained on Pascal VOC-2007+ 12 dataset, only on the extracted pedestrian images. Experimental results show this network achieves 79.86% detection accuracy while detecting smaller pedestrians and still meets the real-time requirements.