Hey Bro, We thought of the approach mentioned by you to classify the objects as moving towards or away from the vehicle we have a doubt w.r.t to that. Consider the above case in which the same person in two different positions one towards the vehicle and one away from the vehicle, but since a CNN uses strided convolutions we have a doubt if the network can learn the variation that arises due to different spatial positions in the image of similar objects which we thought cannot be learned by the network. There are many cases like this so if we can do something for this then the purpose can be achieved by two classes. It's like classifying two cars moving to the left one on the right part of the image and one on the left of the image as different ones. So, something like cropping the image two equal parts and applying opposite rules on either side may work of this, but we don’t how the network would be like two networks in parallel, and don’t know how to label the image, further this may be a problem for an object in exactly middle. Coming to the next one, we thought of one more approach that might be possible By using the above approach, we want to compensate the lack of images for the two classes and consider the weightage for the change while writing the algorithm. Previously we classified the below person as Front, now we would categorize him as front left. Now coming to the other one we thought of doing something with the location of bounding box that is obtained from the network and using this to further classify the direction so that difference in spatial position can be solved. And we have a doubt if anything could be done like classifying the risk from the images directly, but this may have a problem with alignment of camera, change in type of camera etc. So could you please help us what to do next either with the classes or is there any other new approach that you would suggest?