An Improved Multi-Pedestrian Tracking System Based on Deep Neural Network 2019-01-1055

The intelligent vehicle driverless technology has become a very hot topic in the past two decades. To solve the road safety problem, which is one of the most important factors inhibiting the development of intelligent vehicle technology, the multi-target tracking system has attracted more and more attention in recent study since it can detect and track multiple objects in traffic scene so as to help the whole driving system plan the safe route. In this work, a novel multi-pedestrian tracking system based on deep neural network is proposed to improve the tracking efficiency while providing high recognition accuracy. The proposed tracking system consists of two parts: 1) pedestrian detector, and 2) pedestrian tracker. For the detector part, we first transform the image convolution operation in spatial-temporal domain to the coefficient product operation in complex frequency domain, and then replace the maximum pooling and mean pooling operation in the traditional SSD detection network by the proposed product operation, and the modified SSD model is used as the pedestrian detector. For the tracker part, we train a corresponding appearance model and calculate the appearance similarity of the detected targets between the continuous frames with respect to the cosine distance. To evaluate the performance of our proposed system, we implement the system on the MOT16 benchmark. The detector can improve tracking accuracy by up to 8.9% and the tracker updates the speed of whole system at a rate of 126Hz, our extensions reducing the number of identity switches by 15%. The experimental results demonstrate that the proposed multi-pedestrian tracking system provides better real-time performance and accuracy than the state-of-the-art methods, which can provide more accurate object information for the following auto-driving system with higher safety.