pedestrian detection for advanced driving assisting system: a transfer learning approach

Pedestrian detection is an important task that must be integrated into an advanced driving assisting system (ADAS). For a pedestrian detection task many rules must be respected like high performance, real-time processing, and lightweight size to fit into the embedded device of the ADAS. In this paper, we propose a pedestrian detection system based on a convolutional neural network (CNN). CNN is a deep learning model generally used for computer vision tasks like classification and detection because of its power in image processing and decision making. The proposed CNN model is named Yolov3 tiny. It was firstly used for general object detection. In this work, we applied the transfer learning technique on the proposed CNN model to make it suitable for pedestrian detection. The pedestrian detection dataset Caltech US was used to train and evaluate the proposed model. The model achieves an average precision of 76.7% and an inference time of 202 FPS.

I. Introduction

An advanced driving assisting system (ADAS) was developed to ensure more safety for drivers and pedestrians. Since its invention in the 1980s its main role was to assist the driver in controlling the vehicle. Generally, an ADAS is composed from a huge number of sensors and cameras where all are connected to a processing unit to process data and generate decisions by applying a number of authorized actions or warnings. Human benign is the most important element in the loop and its safety comes first. A pedestrian detection application is a must for the ADAS. In this case pedestrian detection is a challenging task due to many factors like occlusion, different points of view, deformation and the complex background. All those factors make building a high-performance pedestrian detection system a very difficult task to solve. In an ADAS the pedestrian detection system is based on the information provided by the cameras. So, pedestrian detection is a computer vision task that must be solved. In recent years, the presence of the deep learning technique [1] has boosted state-of-the-art of computer vision applications including image classification and object detection. Deep learning is a set of machine learning algorithms based on deep neural networks. In particular, convolutional neural networks (CNN) are the most used deep learning model for computer vision tasks. CNN is inspired by the biological system for its methodology in image processing and decision making. In addition, CNN models, like all deep learning models, have the ability to learn directly from data without any preprocessing techniques. A CNN model has many hidden layers for feature extraction and selection for decision making. All those features make CNN at the top of algorithms used to solve computer vision tasks. Since the breakthrough of the CNN in 2012 [2], all computer vision applications are now based on CNN models. CNN models were deployed successfully in many applications like traffic sign recognition and detection [3], [4], indoor object recognition [5], image tagging [6], object detection [7]. etc.