Indian Roads Dataset (IRD)
Camera based Labelled Dataset Captured on Indian-Roads for Self-driving Cars and Driver Assistance Systems
Sarita Gautam, Anuj Kumar
IRD(Indain Roads Dataset)
Object detection is a mandatory task for self-driving vehicles. As in real-world scenarios, there are multiple objects on the roads that need to be detected and classified accordingly for safer and more efficient transportation. Detecting objects on Indian roads is a very tedious task, as moving objects possess a higher collision risk than the still ones. Also, there are no sufficient datasets available based on Indian roads. In this work, we present a fresh state-of-art dataset captured on Indian roads in two different cities (New Delhi and Chandigarh). The dataset has been captured using a 64-megapixel camera by holding the camera on a camera holder in a Car. Initially, video sequences have been captured from the roads of New Delhi and Chandigarh. Then frames have been extracted using the VLC media player. A step-by-step procedure on how to extract the frames has been shown using pictures.
Videos from Chandigarh road. Please click on the link to watch the video.
In this work, we propose a New Indian roads Dataset captured from New Delhi and Chandigarh. We captured videos from the roads of New Delhi and Chandigarh. Our Dataset contains 3000+ images captured in clear day-light conditions and some images captured in night-light conditions. We have annotated our dataset using the Super annotate tool, it's a freely available annotation tool having various annotation objects such as rectangle, polyline, polygon, bounding box, ellipse, cuboid etc. We can create different classes with different color codes to provide segregation between classes. For example: we kept color code as green for car, red for red_traffic_light, yellow for a person, and orange for a motorbike. Detailed description of annotation tool is given in the subsequent sections.
Frame Extraction from videos:
We extracted frames from the videos using VLC media player. We followed a number of steps as given below.
1. Open VLC player and paused the video.
2. Then under the tools tab we clicked on preferences.
3. After that we landed on the Interface tab, where we need to select “All” preferences under the “show settings” bar.
4. This will lead us to “Advanced Preferences”, here under the video submenu we clicked on “filter” and select “scene filter” from the drop- down menu.
5. A new small window pane will open on the right side of the same dialogue box.
6. Here we can specify the image format, Height, width, filename prefix directory path and the recording ratio.
7. Recording ratio denotes the ration of images to be recorded. Here recording ratio 10 means one image out of ten images are stored. We kept recording ratio as 3.
8. Now after changing all the settings, we played the video. As soon as the video is done playing, go to the specified directory and here you can see images extracted from the video. Steps for frame extraction are given in the following images.
Frames ectraced from video sequences
Super annotate Tool Welcome :
This is the welcome window, This window opens up as soon as you click on the super annotate icon. Here you can create "New Projects" and give your project a name. You have to specify the location of the images in this window for starting annotations on them.
After creating a new project. The new project will open up in a new window where actual annotations takes place.
Annotation Tool Homepage:
On the homepage, After selecting folder on the starting page, your images will get automatically loaded in the task pane. On the left hand side, you can see a toolbar having tools for creating different annotations like rectangle, polyline, polygon, bounding box, ellipse, cuboid, zoom-in, zoom-out, Eyedropper, Bucket etc. Then we created classes like car, traffic light, person, truck, green traffic light, red light, bus etc. Now we selected any of the annotation tools and statrted annotating objects as per the requirement as shown in the figure below.
Effort 3
Classes are selected based on the objects present on the images. Bounding boxes are created around the objects and their specified classes are selected. Differnt objects have different color codes.
Evaluation
We used Transfer learning for the evaluation of our images: Transfer learning is a process of reusing pre-trained knowledge and applying it to process similar kind of new data. During the evaluation process we tried to detect traffic lights. We created a traffic light extraction model to create cropped images of traffic lights and then detected traffic lights and other objects from the cropped images. We used pre-trained transfer learning model “Inception_v3” for training our model trained and detected images using single-shot detector. Here are some of the cropped images as shown in Table-5. We evaluated our dataset on a NVIDIA GEFORCE GTX 1650 ti having 8 gb RAM. Our system achieved 97.23% test accuracy. The line graph is shown in image below.
The link to our dataset is provided here:
https://drive.google.com/drive/folders/1ua96AZ75ETujmG-89U1ezVRngV-2vL0t?usp=sharing
