Multi-Color Vehicle Tracking Based on Lightweight Neural Network

Vehicle tracking plays an important role in intelligent traffic management and criminal investigation assistance. At this stage, vehicle target detection technology has reached a relatively mature level of accuracy, but it is difficult to deploy to embedded platforms as the network depth increases. The computing power of the computer also has high requirements. In addition, the continuous increase of vehicle color richness increases the difficulty of target tracking, and it is impossible to identify color types that do not appear in the data set. In response to the above-mentioned problems, this paper proposes an improved YOLOv3 target detection algorithm that can be transplanted to the embedded side. Aiming at the shortcomings of the original YOLOv3 algorithm model that occupies a large amount of memory and is difficult to detect in real time on the embedded side, the lightweight MobileNetv2 depth reduce the network to replace the original YOLOv3 backbone network Darknet-53 for feature extraction, and at the same time make matching changes for anchors to adapt to the characteristics of the dataset for detection. Use the extracted self-built 24-color dataset for training, and then the experimental results show that the parameters and detection speed of the light-weighted YOLOv3 network is significantly better than that YOLOv3, and the recognition accuracy of the our dataset can reach 94.5%.

I. Introduction

According to the data report on cars ownership in 2020, the country’s car ownership reached 281 million, which has increase of 8.1% over 2019. Many cars can be created under the same main color. Different colors, although many colors do not look very different, only gray has 140 different colors in the car market, and blue has 160. It can be seen that the color richness of vehicles on road traffic has increased exponentially, and it also puts forward higher requirements for traffic detection systems, criminal investigation fields, and road detection systems. With the continuous increase of vehicle color [1] categories, the difficulty of real-time decision making cases has increased to a certain extent. At the same time, body color of vehicle as the most significant and stable feature of the vehicle itself, is also one of the main clues widely used in vehicle recognition [1, 2], avoiding the license plate as recognition errors caused by decks, fake number, or man-made damages when recognizing features can improve the accuracy of vehicle recognition to a certain extent [2, 3].