Avoiding Accidents Using Computer Vision 

The increase in the demand and usage of motorized vehicles in all areas has substantially increased the rate of road accidents in recent years. Furthermore, injuries due to road accidents have been increasing every year despite the safety measures introduced by the government.
The recent advancements in technology have opened a new door for the application of computer vision in accident avoidance. Computer vision is an interdisciplinary scientific field that deals with how computers can gain high-level understanding from digital images or videos. Various Object detection algorithms can be used to detect objects in front of a vehicle and avoid accidents.

Background of the Problem

The trend in Road traffic accidents (RTA) injuries and death is becoming alarming in countries like India. The number of fatal and disabling road accidents happening is increasing day by day and is a real public health challenge for all the concerned agencies to prevent. Awareness creation, strict implementation of traffic rules, and scientific engineering measures are the need of the hour to prevent the public. RTA is happening on the roads in India in which countless numbers of people are killed or disabled. Often these are the members whose whole family is wiped out.

The Need of the Solution

As per the Road Accident Report for 2019, a total number of 449,002 accidents took place in the country during the calendar year 2019 leading to 151,113 deaths and 451,361 injuries. Due to road accidents, every year nearly 1 out of 20,000 people lose their lives, and 12 out of 70,000 individuals face serious injuries in India. A lot of research has been done for pedestrian detection, vehicle detection, and animal detection on roads. In this blog, we have discussed various advanced object detection algorithms that can be used for accident avoidance.


Different Methods

1) Histogram of Oriented Gradients (HOG)

The histogram of oriented gradients (HOG) is a feature descriptor used in computer vision and image processing for the purpose of object detection and image classification. . The idea behind HOG is to extract features from an image into a single-dimensional vector, and feed it into a classification algorithm like a Cascade classifier or a Support Vector Machine that will assess whether a pedestrian  (or any other object ) is present in a region or not.


          [Image_Ref:

 In the histogram of oriented feature descriptor, the distribution of directions of gradients i.e oriented gradients are used as features for the given image. Gradients are nothing but the x and y derivatives of the image.  of an image. Gradients are useful because the magnitude of gradients is large around edges and corners. Edges and corners (high-frequency regions) contain a lot more information about object shape than flat regions. which is an important characteristic.

Majors steps to be followed are:

           1.Calculate the Gradient Images

           2. Compute the HOG

           3. Block normalization

           4. Feeding the vector into SVM


2) Region-based Convolutional Neural Network( R-CNN)

Regions with convolutional neural networks or R-CNN is a two-stage object detection algorithm. It combines rectangular region proposals with convolutional neural network(CNN) features. In the first stage, the algorithm identifies regions in the given image that might contain an object, and in the second stage, it classifies the objects in each region.

The first step is to Generate region proposals( regions in the given image that might contain an object) using an algorithm such as Edge Boxes or a selective search algorithm. Selective search algorithms generate 2000 region proposals. The selective Search algorithm works in the following manner. First, it Generates initial sub-segmentation, then it uses a greedy algorithm to recursively combine similar regions, and finally, it uses the generated regions to produce the final candidate region proposals.

            [ image ref: https://www.mathworks.com/help/vision/ug/rcnn.png]
The second step is to crop out and resize the proposal regions out of the image. Then in the third step, we extract features from the cropped and resized regions and classify the regions using a CNN classifier. Which is refined by a support vector machine (SVM) that is trained using CNN features. In addition to predicting whether an object is present or not within the region proposals, the algorithm also predicts offset values to increase the precision of the bounding box. 

3) FastRCNN

To solve some of the drawbacks of R-CNN and to build a faster object detection algorithm the authors of the R-CNN paper proposed a new modified algorithm which is called Faster R-CNN. Unlike the R-CNN detector, which works on the cropped and resized regional proposals,  the Fast R-CNN detector processes the entire image. 

            [ Image ref: https://www.mathworks.com/help/vision/ug/fast.png ]

Thus it generates a convolutional feature map. From the convolutional feature map, The algorithm identifies the region of proposals and warp them into squares. Then it reshapes them into a fixed size by using an RoI pooling layer so that it can be fed into a fully connected layer. And from the RoI feature vector, uses a softmax layer to predict the class of the proposed region. The reason the Fast R-CNN algorithm is faster than the R-CNN algorithm is, In fast R CNN we don’t have to feed 2000 region proposals to CNN every time. Instead, a feature map is generated by performing a single convolution operation per image.


4) SingleShot Multibox Detector (SSD)

It was released in 2016, scoring over 74% mean average precision (mAP) at 59 frames per second on standard datasets such as PascalVOC and COCO. The name of this architecture is being divided into : 

1. Single-shot - In a single forward pass of the network, the object localization and classification is being done
2. Multibox - Used for bounding box regression
3. Detector - For Object Detection. 
Architecture - VGG 16 is used for high-quality image classification tasks and its popularity where transfer learning helps for the results. Enabling to extract features at multiple scales and to decrease the size of the input to each subsequent layer, we would be considering the set of convolutional layers. 
Multibox - There are 2 critical components. First is the confidence loss which measures how confident the network is of the objectness for computing the bounding box and the second is the location loss which measures how far is the bounding box away from the ground truth. The formula for calculating multibox loss is :

multibox_loss = confidence_loss + alpha * location_loss

[image_ref_:https://towardsdatascience.com/understanding-ssd-multibox-real-time-object-detection-in-deep-learning-495ef744fab]

In this for the contribution of the location loss, the alpha term is used. Also, MultiBox starts with the priors as predictions which attempt and help to regress closer to the ground truth bounding boxes.


5) You Only Look Once (YOLO)

          YOLO trains on full images and optimizes directly for detection performance. YOLO is extremely fast. Unlike sliding window and regional-based CNN, YOLO sees the entire image during training and test time. Also, YOLO makes less than half the number of background errors compared to Fast R-CNN. When trained and tested on images, YOLO outperforms top detection methods like DPM and R-CNN by a wide margin. Below diagram shows the detection process using YOLO for preventing accident of car :
Two things that need to be stand out when we use YOLO :
1. Differential weight for predictions from boxes that contain boxes and objects that do not contain objects when used for training.
2. Predict the square root of the bounding box width and height to penalize error in small objects and large objects differently.

Results
The results of avoiding road accidents using Region-based Convolutional Neural Network (R-CNN) 

The results of avoiding road accidents using YOLO

Conclusion

In this paper, We have described and discussed advanced object detection algorithms such as HOG(Histogram of oriented Gradients) Cascade, Regions with convolutional neural networks or R-CNN, Fast R-CNN, Single Shot Multibox Detector (SSD), and You Only Look Once (YOLO).  These algorithms need a good amount of data ( of images ) which is taken in different real/realistic scenarios like time of the day, type of area(for example urban, rural, forest, etc. ), quality of the camera, etc. 

Computer vision is one of the fastest-growing fields. With the increase in the availability of data and deep learning research quality, the object detection algorithms will become more accurate and faster.

References

[1] https://en.wikipedia.org/wiki/Computer_vision
         [2] https://morth.nic.in/road-accident-in-india 
      [3] Girshick, R., J. Donahue, T. Darrell, and J. Malik. "Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation." CVPR '14 Proceedings of the 2014 IEEE Conference on Computer Vision and Pattern Recognition. Pages 580-587. 2014        [4] Zitnick, C. Lawrence, and P. Dollar. "Edge boxes: Locating object proposals from edges." Computer Vision-ECCV. Springer International Publishing. Pages 391-4050. 2014.          [5]https://ieeexplore.ieee.org/stamp/stamp.jsp?arnumber=8832160
         [6]https://towardsdatascience.com/understanding-ssd-multibox-real-time-object-detection-in-deep-learning-495ef744fab
         [7]https://medium.com/egen/region-proposal-network-rpn-backbone-of-faster-r-cnn-4a744a38d7f9
 [8]https://www.hindawi.com/journals/jat/2020/9194028/


Author behind this post are : Nisarg Mehta, Vivek Mankar, Addvija Medhekar, Amol Pawar, Umang Pathrabe

PLEASE COMMENT YOUR THOUGHTS IN THE COMMENT SECTION. HOPE YOU ALL LIKED OUR BLOG


        




               






Comments

Post a Comment