Real Time Pedestrian Detection Using Robust Enhanced Tiny-YOLOv3
Balaram Murthy Chintakindi, Mohammad Farukh Hashmi · 2020
One of the key components in developing smart cities in countries like India, autonomous pedestrian detection plays a very crucial role in Computer Vision tasks such as smart video surveillance and intelligent traffic monitoring system. In self-driving cars, real time performance is required without scarifying the detection accuracy, while detecting smaller pedestrians. In the proposed paper, a robust Enhanced Tiny-Yolov3 Network is designed by introducing an anti-residual module which helps to improve network's feature extraction ability. Second, the loss function is improved which reduces bounding box loss error and optimizes the network. Third one prediction scale of size 26 x 26 is removed from the Tiny-Yolov3 network so the computational complexity of the network is reduced. The proposed network is trained on the extracted pedestrian images from Pascal Voc-2007 dataset. Experimental results show this network improves the detection accuracy while detecting smaller pedestrians and it still meets the real-time requirements.