| Summary: | 碩士 === 國立臺灣海洋大學 === 資訊工程學系 === 105 === The age of automation is coming. Automation technology is already very common in our life, such as automatic cashier in parking lot, automatic ordering system in restaurant, and automatic production line in factory. According to the latest reports, there are 1.2 billion vehicles in the world in 2016; in other words, it is an average of two to three people have a car. Therefore, analysis of vehicle is a significant topic now. The vehicle re-identification system will be an important technique in the future.
Faster-RCNN is a powerful architecture for object detection. A vehicle model is trained for our system to solve the detection of vehicle.
There are two method to implement vehicle re-identification system. First, a HandCrafted method is implemented. The gradient of the vehicle ROI is calculated. The feature points are extracted by a SURF-like algorithm. The distance of feature vector of two vehicle can be calculated to the similarity of two vehicle. Our system output the similarity list which is sorted by bubble sort.
The number one vehicle in the similarity list is regarded as the re-identification vehicle. It is different from the past study that color feature is not used because the input data is all grayscale image.
Deep learning is a popular topic in recent years. The MatchNet network is implemented in our vehicle re-identification system. At beginning, each vehicle ROI is regularly divided to various patches with the same dimension. Input the patches orderly, the Feature network will extract the feature array of each two patches. After input the feature array to the Metric network, the similarity of each two patches can be calculated. The similarity of the Metric network is a score from 0 to 1. Likewise, there will be a similarity list sorted by some simple sorting algorithm. The number one vehicle is the output of our system. This method get a great performance for vehicle re-identification system.
|
|---|
