Road traffic detection and tracking with deep neural network
Loading...
Date
2019-06
Authors
Wong, Kai Kit
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
In Malaysia, road condition is getting more challenging especially in an urban
area like Penang. There are more registered vehicles on the road and more traffic conflict
happened. Traffic managers are required to monitor the traffic condition and analyse it
for future road planning via road closed-circuit television (CCTV) camera. This
monitoring system requires a lot of concentration and long working hours. Thus, human
error is high. Hence, vision-based traffic surveillance system is used to assist traffic
manager to process the input video frame from CCTVs to more meaningful data. This
research’s aim is to build a deep neural network-based traffic surveillance system which
is suitable for Penang traffic road condition. The proposed traffic surveillance system
includes a vehicle detection algorithm by using DNN and vehicle tracking with a
tracking-by-detection algorithm. The first stage is the vehicle detection algorithm. Three
experiments have conducted which are YOLOv2 with pretrain COCO dataset, YOLOv3
with pretrain COCO dataset and YOLOv3 trained with custom dataset. The custom
dataset combined MIO-TCD localization dataset and motorbike of Nepal vehicle
detection dataset. YOLOv2 with pretrain COCO dataset performs poorly with input video
frame which precision of truck and motorbike are 0.00%. YOLOv3 with pretrain COCO
datasets perform well with car and bus class which have 72.97% and 92.00% of precision.
However, the precision of truck and motorbike is just 46.23% and 40.37% respectively.
At the same time, YOLOv3 with pretrain COCO dataset can detect small vehicles that
cannot be detected with YOLOv2 model. For YOLOv3 that istrained with custom dataset,
Motorbike class cannot be detected as the number of motorbike images is insufficient for
training. Overall performance of other classes is good which precision of car, truck, and
bus class are 92.05%, 64.71% and 70.77% respectively. From field test observation, the
trained model does not have the problem that is faced with YOLOv3 with pretrain COCO
dataset. For vehicle tracking algorithm, Euclidean distance is used to identify the same
vehicle in the video frame with two important parameters which are assigned identity
disappear parameter after number of frames and distance parameter to assume centroids
of current frame and previous frame as the same identity.