Deep Learning for Pedestrian Detection and Safety in Autonomous Driving

Abstract

The most significant challenge is different lighting conditions, the varies brightness of pedestrian's clothes, or complicated background. Reasonable and suitable feature extractions of different background can get different accurate detection results. Therefore, the most important in practical applications is the size and accuracy, which fits the appropriate feature descriptions, so as to give the highest precision of pedestrian detection. The algorithm should be able to balance detection speed and accuracy. For example, when used to detect in a video stream in real time, the algorithm gets a real-time effect (10fps) when used to detect pedestrians. To balance the accuracy, adding more layers in the YOLO prediction network to increase the receptive field area increases the accuracy of detection. Different ways and different accuracy balance strategies can be selected according to different applications. At the same time, pedestrian detection models should have robustness, so that it can still ensure the stability when there are other factors such as rain, snow, or occlusion that are difficult to identify pedestrians. Moreover, the computational cost of the detection algorithm should not be too high. Because the speed of hardware is limited . Therefore, an excellent detection algorithm needs to balance accuracy, real-time degree, robustness and memory consumption effectively. In this work, a pedestrian detection model that can effectively balance the four factors is proposed. In the specific design process, the YOLO network is chosen as the base detection algorithm model, and a lightweight detection feature network-MobileNetV2 is used as the feature extraction network to balance detection speed and accuracy feasibly [1].