SlideShare a Scribd company logo

Road Object Detection

•Download as PPTX, PDF•
•
Md. Mahfujur Rahman
Md. Mahfujur Rahman

The document describes a study that used Faster R-CNN, a deep learning approach, to detect road objects like vehicles, pedestrians, and traffic signs in Bangladesh. The researchers collected images to train a neural network to identify 19 object classes. Their model achieved 86.42% accuracy and was able to detect objects in various lighting and traffic conditions, though it struggled when objects were extremely close together. The study aims to help analyze traffic and potentially assist autonomous vehicles in Bangladesh.

Related slideshows

Synops emotion recognize
Synops emotion recognizeSynops emotion recognize
Synops emotion recognize
 
Autonomous Driving
Autonomous DrivingAutonomous Driving
Autonomous Driving
 
Google driverless car
Google driverless carGoogle driverless car
Google driverless car
 
1 of 15
Road Object Detection in Bangladesh using Faster RCNN: A Deep learning Approach Anik Datta, Tamara Islam Meghla, Tania Khatun, Mehedi Hasan Bhuiya, Shakilur Rahman Shuvo and Md. Mahfujur Rahman PAPER ID 282
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Outline • Introduction • Motivation • Methodology • Result and Discussion • Conclusion ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH • 2
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Introduction • Object detection is a combination of computer vision and image processing technology that can detect similar objects of a certain class from images and videos. • The role of object detection is very important for autonomous car, intelligent driving assistance and advanced traffic analysis nowadays. •In the case of traffic analysis and intelligent driving assistance in Bangladesh, it is very important to properly identify all the objects from real time video. •Our main objective is to accurately identify every object on the road such as pedestrians, vehicles, traffic lights and traffic signs in a populous country like Bangladesh and use them to better analyse the traffic situation or use this information in automated driving. 3ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Motivation • Road accident in Bangladesh has become an important issue. •There has been an alarming rise in road accidents, significantly highway accidents, in Bangladesh over the past few years. •Road accidents are mainly occurred for drivers carelessness and violating traffic laws. •To solve such problems we have done this research to save some lives from accidents and to extract important traffic information so that the traffic system can be further improved in the future. 4ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology The reason why did we choose to use Faster R-CNN Faster R-CNN is one of the most popular object detection algorithm. We chose Faster R-CNN for object detection because it is faster and it fixed the problem using Region Proposals Network(RPN).Test time speed of Faster R-CNN for per image 0.2 seconds. Although YOLO is faster than Faster r-cnn. YOLO has some drawbacks in object detection. It is difficult for YOLO in detecting objects that are small and close to each other. Even if the aspect ratio of the image changes, YOLO faces difficulty in object detection. 5ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology Data Accumulation We collect our findable image by clicking photographs and also collect images for our dataset from google. After collecting the images after that we label the desired objects in every image. Faster RCNN We use Faster R-CNN architecture for object detection. It is a modified version of Fast R-CNN. At first Faster R-CNN extracts the input image using convolution neural networks and passes those maps through the region proposal network(RPN). Faster R-CNN uses convolution neural network. It is very effective in image recognition and classification. It can successfully identifying face, object and traffic sign. 6ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology 7 Fig. 1: Working procedure of RPN Fig. 2: Object detection with Faster R-CNN ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology Training and Prediction We have trained our Neural Network through TensorFlow api by Faster R-CNN in 19 object classes. We used inception v2 coco architecture utilize due to high accuracy among convolution neural networks. That inception v2 has the best learning rate for faster RCNN. Our neural network has completed 80,651 global steps during training and has a loss value of 0.0122. We have enlisted the help of Google colab to complete the above process. 8ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion Performance Analysis We took a huge number of instances from the data set to train and observed them with precise needs. For testing purposes, we select 199 instances with same label and 125images as different label. After that, the prediction accuracy jumped up to 86.42%. 9ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion Real-time detection result • We noticed some promising things while checking our results like we were able to detect objects with almost equal efficiency in day and night or in low light. 10 Fig. 3: Detection in day(Bright light) and night(low light) ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • Our neural network has been able to detect objects efficiently with moderate and low traffic on the road. And succeeds in detecting all types of small-to-large-medium objects in all situation. 11 Fig. 4: Detection in medium and low traffic ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • Our purposed neural network can detect and analyze objects well even when the objects are very close to each other and even during high traffic on the road. 12 Fig. 5: Detection in high traffic ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • But it only gets into trouble when the objects are too close together and cling to each other in a crowded state. •When detecting objects on high traffic or in close proximity to the road, objects in the front make it difficult to detect the objects behind because the objects in the front cover the objects behind and therefore our proposed neural network can only accurately detect objects in the front. 13ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Conclusion • In this paper our proposed solution is able to analyse in different situations like night, daylight or high traffic, medium traffic, low traffic fairly efficiently. • And our proposed solution is able to track multiple objects which is possible due to its huge object class. • When we test through this data set, some objects are mistakenly detected and many times some objects are not detected. • The reasons we have investigated are that if it is processed on a low powered machine without a GPU, there is a possibility of incorrect detection and another reason is that the images we used to create the data set have different traffic situations. • It could be further improved in the future to be used more efficiently in real-time traffic analysis. 14ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Thanks 15ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH

More Related Content

Road Object Detection

  • 1. Road Object Detection in Bangladesh using Faster RCNN: A Deep learning Approach Anik Datta, Tamara Islam Meghla, Tania Khatun, Mehedi Hasan Bhuiya, Shakilur Rahman Shuvo and Md. Mahfujur Rahman PAPER ID 282
  • 2. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Outline • Introduction • Motivation • Methodology • Result and Discussion • Conclusion ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH • 2
  • 3. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Introduction • Object detection is a combination of computer vision and image processing technology that can detect similar objects of a certain class from images and videos. • The role of object detection is very important for autonomous car, intelligent driving assistance and advanced traffic analysis nowadays. •In the case of traffic analysis and intelligent driving assistance in Bangladesh, it is very important to properly identify all the objects from real time video. •Our main objective is to accurately identify every object on the road such as pedestrians, vehicles, traffic lights and traffic signs in a populous country like Bangladesh and use them to better analyse the traffic situation or use this information in automated driving. 3ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 4. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Motivation • Road accident in Bangladesh has become an important issue. •There has been an alarming rise in road accidents, significantly highway accidents, in Bangladesh over the past few years. •Road accidents are mainly occurred for drivers carelessness and violating traffic laws. •To solve such problems we have done this research to save some lives from accidents and to extract important traffic information so that the traffic system can be further improved in the future. 4ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 5. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology The reason why did we choose to use Faster R-CNN Faster R-CNN is one of the most popular object detection algorithm. We chose Faster R-CNN for object detection because it is faster and it fixed the problem using Region Proposals Network(RPN).Test time speed of Faster R-CNN for per image 0.2 seconds. Although YOLO is faster than Faster r-cnn. YOLO has some drawbacks in object detection. It is difficult for YOLO in detecting objects that are small and close to each other. Even if the aspect ratio of the image changes, YOLO faces difficulty in object detection. 5ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 6. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology Data Accumulation We collect our findable image by clicking photographs and also collect images for our dataset from google. After collecting the images after that we label the desired objects in every image. Faster RCNN We use Faster R-CNN architecture for object detection. It is a modified version of Fast R-CNN. At first Faster R-CNN extracts the input image using convolution neural networks and passes those maps through the region proposal network(RPN). Faster R-CNN uses convolution neural network. It is very effective in image recognition and classification. It can successfully identifying face, object and traffic sign. 6ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 7. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology 7 Fig. 1: Working procedure of RPN Fig. 2: Object detection with Faster R-CNN ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 8. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Methodology Training and Prediction We have trained our Neural Network through TensorFlow api by Faster R-CNN in 19 object classes. We used inception v2 coco architecture utilize due to high accuracy among convolution neural networks. That inception v2 has the best learning rate for faster RCNN. Our neural network has completed 80,651 global steps during training and has a loss value of 0.0122. We have enlisted the help of Google colab to complete the above process. 8ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 9. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion Performance Analysis We took a huge number of instances from the data set to train and observed them with precise needs. For testing purposes, we select 199 instances with same label and 125images as different label. After that, the prediction accuracy jumped up to 86.42%. 9ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 10. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion Real-time detection result • We noticed some promising things while checking our results like we were able to detect objects with almost equal efficiency in day and night or in low light. 10 Fig. 3: Detection in day(Bright light) and night(low light) ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 11. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • Our neural network has been able to detect objects efficiently with moderate and low traffic on the road. And succeeds in detecting all types of small-to-large-medium objects in all situation. 11 Fig. 4: Detection in medium and low traffic ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 12. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • Our purposed neural network can detect and analyze objects well even when the objects are very close to each other and even during high traffic on the road. 12 Fig. 5: Detection in high traffic ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 13. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Result and Discussion • But it only gets into trouble when the objects are too close together and cling to each other in a crowded state. •When detecting objects on high traffic or in close proximity to the road, objects in the front make it difficult to detect the objects behind because the objects in the front cover the objects behind and therefore our proposed neural network can only accurately detect objects in the front. 13ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 14. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Conclusion • In this paper our proposed solution is able to analyse in different situations like night, daylight or high traffic, medium traffic, low traffic fairly efficiently. • And our proposed solution is able to track multiple objects which is possible due to its huge object class. • When we test through this data set, some objects are mistakenly detected and many times some objects are not detected. • The reasons we have investigated are that if it is processed on a low powered machine without a GPU, there is a possibility of incorrect detection and another reason is that the images we used to create the data set have different traffic situations. • It could be further improved in the future to be used more efficiently in real-time traffic analysis. 14ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH
  • 15. WIECON-ECE 2020 2020 IEEE International Woman in Engineering (WIE) Conference on Electrical & Computer Engineering wiecon-ece.org Thanks 15ANIK DATTA, ROAD OBJECT DETECTION IN BANGLADESH USING FASTER RCNN: A DEEP LEARNING APPROACH