Home - Detections in Road Traffic Monitoring Videos

Detections in Road Traffic Monitoring Videos

posted in 2018

(I) Objective: Detect pedestrians, motorcycles, cars, trucks, and trailers in road traffic supervisory video footage.

(II) Approach: Employed Faster R-CNN with ResNet50 or ResNet101, as well as SSD with a MobileNet, while reducing the number of channels in the intermediate layers to enhance inference speed.

(III) Result: Faster R-CNN achieved higher precision, while SSD provided faster inference speed and achieved a mAP of over 0.857 and an inference speed greater than 10 frames per second (fps) on the Nvidia TX1.

( experiment 1 ) : Faster R-CNN with ResNet101

the accuracy : mAP= 0.93 .
the model size (pb size) : 187Mb.
the memory usage of the model measured by the benchmark tool (avg) : 3.97409e+09(bytes).
the Computation Time of the model measured by the benchmark tool (avg) : 278120(microseconds).
the quantized model size (pb size) : 48Mb.
the memory usage of the quantized model measured by the benchmark tool (avg) : 4.32859e+09(bytes).
the Computation Time of the quantized model measured by the benchmark tool (avg) : 308244(microseconds).

* The memory usage and Computation Time of the quantized model aren't better than the original model. The reason may be GeForce GTX 1080 doesn't have the integer multipliers for the quantized model.

( experiment 2 ) : Faster R-CNN with ResNet50

the accuracy : mAP=0.9 .
the model size (pb size) : 112Mb .
the memory usage of the model measured by the benchmark tool (avg) : 2.20235e+09(bytes) .
the Computation Time of the model measured by the benchmark tool (avg) : 205617(microseconds) .
the quantized model size (pb size) : 29Mb .
the memory usage of the quantized model measured by the benchmark tool (avg) : 2.4e+09(bytes).
the Computation Time of the quantized model measured by the benchmark tool (avg) : 219534(microseconds).

* The memory usage and Computation Time of the quantized model aren't better than the original model. The reason may be GeForce GTX 1080 doesn't have the integer multipliers for the quantized model.

The detection results of the Faster R-CNN with ResNet50 :

( experiment 3 ) : SSD(Single Shot Multibox Detector) with MobileNet

the accuracy : mAP=0.857 .
the model size (pb size) : 23Mb .
the memory usage of the model measured by the benchmark tool (avg) : 1.21478e+08(bytes) .
the Computation Time of the model measured by the benchmark tool (avg) : 38862.8(microseconds) .

The detection results of the SSD with mobileNet :

Google Sites

Report abuse