| 摘要: | 根據統計，大部分的交通事故都是駕駛因為沒有注意而與量發生碰撞，因此先進駕駛輔助系統 (advanced driver assistance systems , ADAS) 已經變成近年熱門的研究議題。在本論文中，我們提出一個可以適應天候的前車碰撞警示系統，此系統可以幫助駕駛者偵測前方車輛，然而車輛辨識的準確度常常受到許多因素影響，最主要是天氣因素影響天候 (例如，白天、夜晚、晨曦、向陽、斜陽、黃昏、雨天、薄霧、濃霧、陰天)，因此我們用卷積神經網路訓練出可適應各種天候環境的前車偵測系統，在駕駛可能發生危險的情況下提醒，避免意外發生。
本論文分為三個部分:第一部分為改進更快速區域卷積神經網路。原本更快速區域卷積神經網路利用VGG16提取特徵，由於VGG16網路較為龐大，需佔用較多的硬體資源，我們以SqueezeNet架構改進原本VGG16網路，以達到減少網路大小及增進速度的效果。第二部分是將更快速區域卷積神經網路中原本的ROI池化層以ROI校準層取代，以改善偵測結果。第三部分是由系統所偵測出來的車輛框，透過相機的焦距與相機安裝於車上的高度和影像中車輛底部的位置透過相似三角形，計算出與前車的距離。
在實驗中，我們以行車記錄器的影片進行測試，影片中包含各種天候，物件偵測系統的mAP可達0.907，在 640×480 解析度的影片測試平均速度為每秒 30 張影像，參數量大小為7.7 M，原本更快速區域卷積神經網路使用ROI池化層的偵測框準確度為79.25%，而我們系統使用了ROI校準層取代了ROI池化層，偵測框準確度達到90.4%，提高了大約10%。
;In recent years, machine learning has flourished, such as face recognition, speech recognition, object detection, etc. However, in terms of object detection, it is closely related to people′s daily life. In recent years, automatic cars have been gradually emphasized, and the safety of automatic cars is also a very important issue. Maintaining an appropriate distance from the front car to avoid collision with the preceding car is a key point of automatic cars’ practice. However, the accuracy of vehicle identification is often affected by many factors, the most influential factors are weather conditions, such as, day, night, morning, sun, rain, mist, fog, cloudy. Therefore, we use a convolutional neural network to implement a front-vehicle detection system that can adapt to various weather conditions and remind the driver to avoid accidents.
There are three parts in this paper, the first part is based on Faster-RCNN and inproving it. Faster-RCNN uses VGG16 for extraction features. VGG16 network is relatively large, it needs to occupy more resources, so we use SqueezeNet architecture to improve VGG16 architecture to reduce network size and speed. The second part is replaced ROI pooling layer by ROI align layer to improve the detection results; the third part is calculating the distance from the front car, We use the focal length of the camera and the height of the camera mounted on the car to calculate the distance from the front car.
In the experiment, we use 697 images as training and 233 images as test data. The film contains various weather conditions. The mAP of the object detection system can reach 0.907, and the average test speed of a 640×480 resolution film is 30 frames per second. The number of parameters size is about 7.7 M. |
