Objective: the objective is accounting the number of people who taking bus in each day. Then we split it into a sub-problem to detect the head and tracking. We are trying to get 30fps(30ms per frame) detection speed, the accuracy should be above 95%, in case we have head occlusion, black hole of head depth image(the depth sensor has functional distance range, within 0.5 meters, the depth sensor cannot calculate the depth) and fast movement of passengers.
The hardware is ARM v7 , 800M HZ, 2 core CPU(the final production is embedded hardware, but we can use android based phone to test it, try to use low-power android as low as ARM v7 In terms of computation capability, it is less powerful than that of Raspberry pi 2),depth sensor to get the grey-scale map. (the 0-255 value indicate the relative distance which can be normalized to 0~1)
software enviroment:ubuntu 14.04. opencv(or use you own algorithm)
Different from traditional method, the bus is crowded and has narrow space, so the traditional RGB camera will encounter lots of problem for human/head detection. So we change to using depth sensor (like kinect without RGB camera) hanging on the ceiling near the front&back door.
we did using state-of-art method deep learning(DL) which has achieved relative good result. but DL itself is hardware-consuming and dedicated chip handling neural network inference is yet to come. Besides that, ARM V7, 800HZ CPU means limited computation capability.
Any further proposal and referred papers , pls see the attachment (pls download the "Tasklist of bus head detection obtained by depth sensor from overlooking perspective.pdf" firstly, it elaborate the tasklist and includes training data-set download link. other 3 pdf docs are reference papers)
1. I know that deep learning is good. I personally have implemented customized Yolo to achieve 200ms per frame, but it is hard to decrease it to 3oms. (30 fps is our goal). So unitl you have binarrized neural networks like xnor.ai (websites to get more), we cannot achieve it realtime as 30fps.
2. I have prepared the training dataset on the word document. you can download it.
I think the first step is trying to make the detection accurate and fast, so offline algorithm is fine without depth sensor.
January 12, 2018
I am willing to pay higher rates for the most experienced freelancers